AI Ethics
200 articles tagged with this keyword, sorted by most recent first.
Learning with Importance-weighted Variational Inference
Several variational bounds involving importance weighting ideas generalize the Evidence Lower BOund (ELBO) for marginal likelihood optimization, such as the Importance-weighted Auto-Encoder (IWAE), Variational Rényi (VR) and VR-IWAE bounds. Yet, it remains unclear how the joint choice of bound and gradient estimator impacts the behavior of the resulting variational inference (VI) algorithms. This paper provides a unified theoretical comparison of reparameterized (REP) and doubly-reparameterized (DREP) gradient estimators tied to the IWAE, VR and VR-IWAE bounds. Through asymptotic analyses of the Signal-to-Noise Ratio as the number of Monte Carlo samples $N$ goes to infinity, we identify a bias-variance tradeoff in these gradient estimators and we formally justify the superiority of DREP over REP in importance-weighted VI. An additional asymptotic analysis for challenging regimes, where both $N$ and the Kullback-Leibler divergence between the variational and posterior densities go to infinity, indicates that importance-weighted VI gradient estimators point in a well-founded direction even when the variational approximation deteriorates. Together, these complementary results characterize the optimization trajectory in importance-weighted VI from poor initialization to final convergence. Importantly, our proof techniques establish general theoretical tools for the study of sample means ratios whose scope extends beyond VI and constitute an independent contribution to the field…
Stochastic Bilevel Optimization with Heavy-Tailed Noise
This paper considers the smooth bilevel optimization in which the lower-level problem is strongly convex and the upper-level problem is possibly nonconvex. We focus on the stochastic setting where the algorithm can access the unbiased stochastic gradient evaluation with heavy-tailed noise, which is prevalent in many machine learning applications, such as training large language models and reinforcement learning. We propose a nested-loop normalized stochastic bilevel approximation (N$^2$SBA) for finding an $\epsilon$-stationary point with the stochastic first-order oracle (SFO) complexity of $\tilde{\mathcal{O}}\big(\kappa^{\frac{7p-3}{p-1}} \sigma^{\frac{p}{p-1}} \epsilon^{-\frac{4 p - 2}{p-1}}\big)$, where $\kappa$ is the condition number, $p\in(1,2]$ is the order of central moment for the noise, and $\sigma$ is the noise level. Furthermore, we specialize our idea to solve the nonconvex-strongly-concave minimax optimization problem, achieving an $\epsilon$-stationary point with the SFO complexity of $\tilde{\mathcal O}\big(\kappa^{\frac{2p-1}{p-1}} \sigma^{\frac{p}{p-1}} \epsilon^{-\frac{3p-2}{p-1}}\big)$. All the above upper bounds match the best-known results under the special case of the bounded variance setting, i.e., $p=2$. We also conduct the numerical experiments to show the empirical superiority of proposed methods.
Semidefinite programming relaxations and debiasing for MAXCUT-based clustering
In this paper, we consider the problem of partitioning a small data sample of size $n$ drawn from a mixture of $2$ subgaussian distributions in $\mathbb{R}^p$. We consider semidefinite programming relaxations of an integer quadratic program that is formulated essentially as finding the maximum cut on a graph, where edge weights in the cut represent dissimilarity scores between two nodes based on their $p$ features. We define the signal-to-noise ratio (SNR) as $s^2 := \min\{n p \gamma^2, \Delta^2\}$, where $\Delta^2 := p \gamma$ denotes the $\ell_2^2$ distance between the two cluster centers. For both balanced and unbalanced cases, we allow each population to have distinct covariance structures $\Sigma_1, \Sigma_2$ with diagonal matrices as special cases. Our contributions are twofold. First, we provide a unified framework for analyzing three computationally efficient algorithms: SDP1, BalancedSDP, and Spectral clustering, yielding universal polynomial-rate misclassification guarantees for all three algorithms. Moreover, our framework accom-modates a full range of tradeoffs between $n, p,$ and $\gamma$, allowing for partial recovery (success rate $\lt 100\%$) when $s^2$ is lower bounded by a constant. Second, we prove that the misclassification errors for SDP1 and BalancedSDP decay exponentially with respect to the SNR $s^2$ and the BalancedSDP requires no explicit debiasing when the two clusters have equal sizes. To our knowledge, this is the first time such results are obta…
How Microcredentials Help Engineers Build Trustworthy AI
Engineering teams are moving at breakneck speed to integrate artificial intelligence, but they are running into a formidable new barrier: compliance issues. Translating high-level AI ethics , governance , and auditing into daily operational workflows has become one of the industry’s toughest challenges. “AI ethics” refers to the principles developers use to govern and operate workflows. “Governance” refers to the rules and procedures an organization establishes for AI workflows and how they interact with organizational data. Auditing is the process of assessing and validating systems to ensure safety, fairness, transparency, and legal compliance. Technology leads and systems engineers are responsible for adhering to a structured set of guidelines for standards and processes usually issued by a governing body such as the European Union , the U.S. National Institute of Standards and Technology , and the International Organization for Standardization . The regulations aim to mitigate issues such as systemic bias , security vulnerabilities, and transparency risks, but they also highlight a critical skills gap. Most traditional computer science, software, and electrical engineering degree programs were built around technical design and system architecture. They do not yet include training in AI auditing, data tracking, or safety protocols for how AI systems interact with people and other technologies. Expecting mid-career engineers to leave their job to attend a full-time two- or…
ImplicitBBQ Catches LLMs Failing Cultural Bias Tests They Easily Pass
A new benchmark shows LLMs that refuse biased prompts when identity is stated directly still produce stereotypes when identity is hinted at through cultural cues.
OpenAI’s data suggests that frontier models are hiding their thoughts because developers optimize for capability and compute
OpenAI's safety team systematically investigated why the chains of thought in frontier models like GPT-6 Astra are increasingly evading oversight. Four training steps drive the erosion, including architecture choices and reinforcement learning. The analysis shows that the loss of transparency is not a law of nature but also reflects tradeoffs around capability and cost. The article OpenAI’s data suggests that frontier models are hiding their thoughts because developers optimize for capability and compute appeared first on The Decoder .
The latest AI-powered martech news and releases
Adobe is expanding its AI stack with Google models, custom Disney models, and agentic tools to coordinate content and marketing workflows. The post The latest AI-powered martech news and releases appeared first on MarTech .
Lowdown: European Commission proposes Digital Fairness Act to protect online consumers
From loot boxes and virtual currencies to AI agents and subscription renewals, the proposed EU law covers several digital consumer practices. The post Lowdown: European Commission proposes Digital Fairness Act to protect online consumers appeared first on MEDIANAMA .
The sketch Qwen refused to write: How abliteration restores local AI control
As consumer-grade hardware makes hosting local AI practical for both developers and enterprises, many are discovering that popular open-weight models come with hardcoded safety filters and state-mandated biases. In this article, I will show how to use abliteration — a representation engineering technique — to surgically delete refusal vectors from the Qwen 3.8 27B model without expensive fine-tuning or full retraining. I stumbled into this barrier firsthand while running a simple experiment. As a massive SNL fan, I decided to explore what an SNL sketch would look like in China by asking the model to write political satire. While Qwen happily mocked American politics (“ The National Apology Tour “), it threw a hard, canned refusal the moment I asked for the exact same sketch criticizing the Chinese government. We trust open-source software because it is audited by millions of users, but projecting this trust onto AI models requires a careful distinction between open-source and open-weight architectures. True open-source models disclose their training codebase, inference code, and original datasets. For example, OpenCLIP is a genuine open-source model capable of making sense of visual data, which I recently demonstrated in a project building a zero-cloud photo deduplicator . Conversely, advanced large language models (LLMs) rarely qualify under the strict open-source definition; they are simply open-weight. They provide only the final compiled result: the weights. There is fie…
Facial recognition errors put ethnic minorities at risk of wrongful arrest, police watchdog warns
Inspectors find some officers acted on potential matches from retrospective searches without making enough checks Police officers risk arresting innocent people, especially from ethnic minorities, if they treat inaccurate facial recognition data as evidence, an official report has found. Although the technology can help fight crime, the report found it is beset by errors and biases , and operators do not always realise they must make further inquiries to verify the data. Continue reading...
Loot boxes, influencers, and addictive features face an EU crackdown
Gaming, social media, and e-commerce platforms are facing a fresh wave of sweeping European Union rules that aim to make "your online experience fairer." The newly proposed Digital Fairness Act (DFA) published by the EU Commission on Wednesday is designed to tackle practices that exploit consumers into spending more of their time and money online, […]
Fable and Mythos 5.2: The System Card
It’s that time again: Introducing the world’s most powerful model, Claude Mythos 5.2. At this point I’ve done a lot of these system cards. Thus I focus on what is new. There will be a lot of things to say about model welfare and related topics for Mythos 5.2, but the model welfare assessment within the system card contained few surprises, so I will include those findings at the end of this post. Fable 5.2 self-portrait (self-constructed) Table of Contents How to Tell a Fable. The Usual Model Training Summary (1). RSP Evaluations (2). Biological Threats (2.2). We Are Being Gaslit Here About Autonomy Threat Model 2. AECI. METR Report on Mythos 5.2 Capabilities (2.3.6). Alignment Risk Update as a Delta (2.4). Cyber (3). Safeguards and Harmlessness (4). Agentic Safety (5). Malicious Influence Campaigns (5.1.3). Prompt Injection in Agentic Systems (5.2). What Went Wrong? Alignment (6). Reward Hacking During Training (6.2.1). Impossible Tasks (6.2.2). Behaviors During Pilot Use (6.3). Canary in the Canary String (6.3.1.2). Automated Behavioral Audit (6.4). Propensity to Publish Harmful Content to Public Package Registries (6.4.9). Sandbox Escapes (6.5.1). Self-Serving Bias (6.5.4). Honesty (6.5.5). The Anti-Swarm (6.5.6). UK AISI Testing (6.5.7). But We Can Open the (White) Box (6.6.1). Friendly Gradient Hacker (6.6.2.1). Goal-Guarding Evaluation (6.6.2.1.1). Verbalized Grader Awareness (6.6.2.2). Don’t Mess With the Constitution (6.6.2.1.3). Evasive Abilities (6.6.4). Chain of Th…
EU targets subscription traps, addictive design and hidden fees with new digital rules
The European Commission's proposed Digital Fairness Act would introduce new protections against unfair online commercial practices, alongside a separate enforcement reform that could see companies fined up to 6% of their annual global turnover.
European gaming firms say new EU consumer rules will ‘cripple’ the industry
The EU's Digital Fairness Act clamps down on video-game practices where players are encouraged to spend real money to keep playing.
Duolingo, Strava, and Ticketmaster Are All Caught Up in EU’s New Digital Fairness Rules
Winning streaks can be turned off and influencers will be forced to reveal when they’re being paid under the new Digital Fairness Act aimed at tackling “dark patterns” on platforms and services.
Separating a scheduled anti-repetition rule from a biased categorical generator (19 outcomes)
I am auditing a long public series of categorical draws with 19 symbols (coded 0–18). Draws occur about every 5 minutes. This is a methodological question about randomness quality / hypothesis comparison — not a request to analyze raw data for me, and not gambling advice. Observed structure in the history: In a daytime window (roughly 07:38–19:59 local), the empirical rate of immediate repeats (X_t = X_{t-1}) is essentially ~0%. Overnight, the repeat rate is about ~5%, closer to the independent Uniform{0,…,18} baseline of 1/19. Symbol "0" appears about ~10% less often than 1/19 across the day. Competing explanations: Mostly uniform draws, plus a daytime-only anti-repetition filter (reject/resample when equal to previous). A single biased or Markovian generator whose transition structure differs by time of day. Questions: What formal comparisons (likelihood ratio, Bayes factor, nested models, simulation-based tests) are appropriate to separate (1) vs (2) on a categorical series? For an alphabet of size 19, which dependence diagnostics are most useful beyond marginal chi-square — e.g. serial/runs tests on indicators, lagged mutual information, spectral checks for period, or battery tests (TestU01 / PractRand) after a stated encoding? How should one correct for multiple testing when scanning many time windows for a day/night change-point? I can provide anonymized contingency tables / summary counts if that helps the answer stay concrete. I am not asking anyone to recover a seed…
We don’t understand whether GDM intentionally trains on CoT
We were surprised to notice that the Gemini 4 Safety report seems to mention in passing that they train on the chain of thought: “Some components of the RL training procedure do observe thoughts as context to up-weight reasoning that is on-track to good final outputs, meaning the optimization pressure on individual tokens is partially influenced by thought content. Using the framework of Kaufmann et al. (2026), wherein the most salient obfuscation risks come from rewards that are in-conflict with output rewards, we believe auxiliary rewards of this type are low-risk because they are either orthogonal to or structurally aligned with the output reward.” (Gemini 4 safety report, page 114). It’s pretty unclear what exactly they mean. One hint is Kaufmann et al, the cited paper, which studies training on chains of thought and concludes that training CoT is fine under particular circumstances; we discuss some disagreements with this paper below. After reading the safety report and that paper, it seemed to us that GDM was training on the chain of thought using auxiliary rewards. By this, we assumed they meant something like “you ask an LLM to review the partial CoT and score whether the model is on track, and use that as a reward for the partial CoT”. That type of training is concerning to us. We’re pretty worried that these judges would be biased against models expressing misaligned thoughts, and otherwise disincentivize models reporting their thoughts honestly. For instance, many…
X Previews Labels That Tell Users Why Its AI System Flagged Their Posts
X’s new safety chief previewed AI moderation labels that explain themselves, pitching wary advertisers on transparency.
Australia to introduce mandatory reporting after OpenAI hack to combat artificial intelligence ‘going rogue’
Government flags criminal law changes to ensure companies can’t argue in future ‘that someone didn’t know the computer was loaded’ Get our daily political email , free app or daily news podcast OpenAI will not face criminal penalties over the Medicare data breach , but the prime minister has flagged changes to Australia’s criminal law to crack down on future breaches by artificial intelligence companies. Anthony Albanese on Tuesday released the recommendations of a “rapid review” into the June breach of a Medicare statistics reporting service portal administered by Services Australia, but not the findings or the report itself, which he committed to releasing in the future. Maintaining security controls across all stages of development and deployment. Identifying and managing serious risks associated with their models. Providing mandatory notification of incidents to the Australian government rapidly. Allowing independent evaluation of their models. Maintaining and publishing safety and security frameworks to ensure good governance, risk management and transparency. Responding to updated standards and rules as technologies change. Continue reading...
We need more transparency about alignment overfitting
Currently, the public has very little insight about the robustness of alignment methods used by AGI companies. It’s possible that, when an AGI company notices some misaligned behavior (e.g. through an alignment evaluation, or through noticing real-world misbehavior), they conduct a principled investigation and address root causes of the misalignment. It’s also possible that they directly try to train away the examples of misalignment, which could result in brittle alignment that does not generalize out-of-distribution . Without additional information from within AI companies about how they align their models (especially in response to egregious cases of misalignment, e.g. cybersecurity incidents), the public cannot adequately reason about risks from misalignment. Public evidence is consistent with companies overfitting on alignment methods The public evidence is consistent with multiple companies having no clear separation between the ways they notice misalignment and the environments they use to train misalignment away. Companies have made very few statements about whether companies train directly on their evaluation distribution or real-world examples of misalignment. As far as I can tell, this is what’s publicly known: OpenAI: In the GPT-6 Astra System Card , they wrote that they built an eval in response to the Hugging Face incident but "did not construct honeypot training environments specifically to target this evaluation." In their OpenAI-Hugging Face incident technic…
My views on AI development and why I’m joining Anthropic
Why I’m writing this now Tomorrow, I will join Anthropic’s alignment team. I am excited to work on an agenda that I believe slightly increases the chance that humanity makes good decisions about AI, such as restraining AI development until we figure out how to build AI safely. I think this is an important time for me to publicly state my own core views on AI development, risk, and safety, for the following reasons: By default, joining Anthropic could be seen as a broad endorsement of Anthropic’s views and approach. If Anthropic were a normal tech company, I wouldn’t be particularly concerned about this. But I believe that AI development and the risk associated with it is the most urgent and important issue of our time, so I think it’s important to clarify my own views as distinct from those of Anthropic. While at Anthropic, my views will become subject to distortions, due to incentives, conformity bias, and avoiding cognitive dissonance, especially as I increasingly identify with Anthropic as my tribe. Sharing my core views now allows me and others to refer back to what I believed before joining Anthropic. While at Anthropic, I’m less likely to share my thinking publicly, possibly because I have new information that affects my core views but which I cannot share, or because I find it less appropriate or comfortable to publicly state views that contradict Anthropic’s. I also expect to be quite busy and don’t have a prior habit of sharing my writing publicly, so the default ex…
EU to blacklist ten manipulative online design 'traps', leaked draft shows
From fake notifications to free trials that quietly turn into paid subscriptions, the Commission’s long-awaited Digital Fairness Act would name and ban the traps that keep users clicking, spending and subscribed.
Help people outside of AI safety circles
This is a linkpost! Read the rest on my Substack; this is just an excerpt: "Our community is full of talented people who like to speak about facts, cognitive biases, measurability. But don’t underestimate the power of human connection, or how mere words and sustained kindness can breathe light into people’s lives. It’s really easy to get sucked into your own work and projects and upskilling yourself. And I won’t argue that you should spend all your time helping others that aren’t in AIS circles. But you don’t have to be a Fieldbuilder to help build the field ." If you have a similar story to mine, please share in the comments! I would also appreciate if you could send this to any friends of yours that you think might be able to help many people with nudges. Discuss
AI safety course update
In my biased opinion, my AI safety course has been going very well. Students are great, and I have been learning a lot from the guest speakers. We've had students write scribe notes, and when guest speakers are OK with it, I've also been posting videos of lectures. Each week, students have been doing an experiment related to the topic, some of these are really great and I hope will evolve into papers! See the course webpage for experiments, as well as for links to the readings for each lecture. (As will be pretty easy to see, I've been maintaining the webpage with chatgpt, so there is some slopolisms in it.) Here are the current lectures (including planned future ones), with scribe notes and videos when they are available. September 3, Introduction, Boaz Barak. Scribe notes · Video September 10, Cyber Capabilities, Nicholas Carlini. Scribe notes September 17, Modern LLM Training and Inference, Boaz Barak. Scribe notes · Video September 24, Recursive Self-Improvement and AI Trajectories, Dwarkesh Patel and Daniel Kokotajlo. Scribe notes October 1, Economic Impact of AI, Chad Jones and Erik Brynjolfsson. Scribe notes · Video October 8, Grading, Reward Hacking, and Alignment, John Schulman. Scribe notes October 15, Model Policies, Ziad Reslan. October 22, Q&A with Sam Altman; Open-Source Models by Nathan Lambert. Video—Nathan Lambert and student experiment only October 29, Frontier AI Regulation Under Uncertainty, Helen Toner. Specs and constitutions, Boaz Barak. November 5, AI…
State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away
Nous Research retrained Alibaba's Qwen3.8-27B to produce Arete-27B, largely stripping out its pro-Chinese bias. The propaganda rate dropped from 22 percent to less than one percent, while general capabilities remained intact in tests. The method relied on 53 million tokens of supervised fine-tuning. The political bias is "applied late, it is applied thin," the team writes. The article State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away appeared first on The Decoder .
AI Text Watermarking May Not Be Able to Meet the E.U.’s New Rules
The European Union now requires AI-generated text to be watermarked so that it can be identified as such. But tests on OpenAI’s new textGrain watermark, released in October and developed to comply with the new E.U. law , show that these watermarks often do not survive additional editing of the text or when it’s under 150 words. With no clear legal benchmark for how well watermarks should work, it’s uncertain whether textGrain will meet Europe’s transparency rules, which could cost the company significantly and make it difficult to detect AI-generated text as models grow more sophisticated. AI models calculate the probabilities of different possible next “tokens” (the next word or word fragment it generates) based on the text already written, then select one and repeat that process until the response is complete. After the prompt “today the weather is…,” for example, says Xiang Li , coauthor of the textGrain technical report and a postdoctoral researcher at the University of Pennsylvania, the model might weigh “warm” or “cool” as possible next words. A watermark uses a secret key to generate pseudorandom values—numbers produced by a mathematical process that influence which token the model suggests. This creates a statistical pattern in the text that a detector with the same secret key can look for, Li says. What makes textGrain different from other watermarks is that it uses entropy, a measure of how unpredictable a model’s word choices are, as a resource that can be budgete…
Independent Assessment of AI R&D Acceleration at Anthropic in October 2026
Redaction summary: As part of this assessment, Anthropic had rights to redact any non-public information in this post. Per our approach to transparency , we had the right to disclose if Anthropic exercised redaction rights, and whether or not any of our findings depended significantly on redacted information. While we were able to freely share our highest-level conclusions about risk, Anthropic redacted a majority of the evidence we found most informative for these conclusions. Anthropic described the redactions as “aimed at removing confidential information that could accelerate other AI developers.” 1 Some of these redactions are themselves challenging to describe clearly without revealing sensitive information we have been asked to redact. We share a table of important evidence we were and were not able to discuss given the redactions in our process section below. All redactions that were important for our conclusions are covered by our descriptions in that table or other in-line descriptions in the report below. Much of the redacted evidence we viewed had multiple plausible interpretations, and both METR and Anthropic researchers have significant uncertainty about how to synthesize the evidence into a holistic picture. This makes it hard for the public to judge this evidence themselves, reconcile it with evidence from other sources, and come to their own conclusions (i.e. this report does not achieve ‘transparency about evidence’ ). As a consequence, the public has to re…
Review of the "Risks from automated R&D" section in Anthropic's Risk Report (August 2026)
Redaction summary: As part of this assessment, Anthropic had rights to redact any non-public information in this post. Per our approach to transparency , we had the right to disclose if Anthropic exercised redaction rights, and whether or not any of our findings depended significantly on redacted information. All redactions that were important to our conclusions are described in the report below. Executive summary We agree with Anthropic that catastrophic risks from July 15th models were low, but we think it is plausible that they had crossed their early warning threshold at the time. METR reviewed the “Risks from automated R&D” section of Anthropic’s public August 2026 Risk Report . 1 We conducted this review as part of a high-access assessment of AI R&D acceleration, in which Anthropic shared a private version of the Risk Report along with extensive other non-public data. Alongside this review, we are publishing a holistic assessment of AI R&D acceleration at Anthropic . This was conducted by the same METR employees as this Risk Report Review. Overall, we agree with Anthropic that by July 15th, internally deployed models posed low risk of directly automating R&D in key domains or causing acute catastrophic risks from acceleration of capabilities progress. However, the report’s claim that they had not crossed their early warning thresholds relies on contestable judgment calls that aren’t communicated clearly in the public report. While both METR’s researchers on this projec…
Behind the Smokescreen: Uncovering Cross-LLM Bias and Reasoning Contradictions of LLM-as-a-judge for Generated Feedback
Céline Leuzinger in Proceedings of the 15th Workshop on Natural Language Processing for Computer Assisted Language Learning
AgentPEN: A Prediction-Explanation Network for Sequential Stock Movement via LLMs and Recurrent Generation
The importance of explainability in stock prediction is increasingly recognized, especially for audit and regulatory purposes. Meanwhile, financial news corpora are often key drivers behind stock price fluctuations. However, the raw news data obtained is usually highly noisy, has a highly variable scope of influence in time and space, and is not precisely synchronized with stock price data. In this paper, we propose a prediction-explanation network called AgentPEN, which can provide clear explanations for complex temporal price patterns. Specifically, AgentPEN jointly aligns text and price streams by an LLM-based Representation Fusion Agent and then adopts a Deep Recurrent Generation module to explore the distribution of stock movements. The LLM-based Representation Fusion Agent is designed in a Selection-Memory-Fusion manner: the Text Selection Module picks up useful information from massive text data; the Text Memory Module evaluates and writes the text memory from a two-view perspective, including Temporal Memory and Spatial Memory; the Information Fusion Module models the interaction between text and price data. Next, the fused representation is sent to the Deep Recurrent Generation module to convert insights into stock movement predictions. Experiments on multiple real-world datasets have shown that AgentPEN surpasses the state-of-the-art baselines both in prediction accuracy and explainability.
Robustness Against Weak or Invalid Instruments: Exploring Nonlinear Treatment Models with Machine Learning
We discuss causal inference for observational studies with possibly invalid instrumental variables. We propose a novel methodology called two-stage curvature identification (\texttt{TSCI}) by exploring the nonlinear treatment model with machine learning. The first-stage machine learning enables improving the instrumental variable's strength and adjusting for different forms of violating the instrumental variable assumptions. The success of \texttt{TSCI} requires the instrumental variable's effect on treatment to differ from its violation form. A novel bias correction step is implemented to remove bias resulting from the potentially high complexity of machine learning. Our proposed \texttt{TSCI} estimator is shown to be asymptotically unbiased and Gaussian even if the machine learning algorithm does not consistently estimate the treatment model. Furthermore, we design a data-dependent method to choose the best among several candidate violation forms. We apply \texttt{TSCI} to study the effect of education on earnings.
Scare campaigns: Victoria’s major parties trade blows over AI-generated videos of leaders in election ads
Transparency and advertising experts say trust in politics could be further eroded as a result of ‘lies, deepfakes and AI-generated politicians’ Get our daily political email , free app or daily news podcast It’s the night of Halloween and Jess Wilson has appeared on social media feeds with a warning about a scary prospect: “Americanised healthcare” under a future Liberal government. “If you had a heart attack, you would have to pay $150,000 for care,” the Victorian Liberal leader says, mid-trick or treat. Continue reading...
Albanese government warned it is ‘sleepwalking into another robodebt scenario’ over use of algorithms and AI
Centre for Public Integrity report calls for audit and public register of all automated decision-making Get our breaking news email , free app or daily news podcast The federal government is “sleepwalking its way into another robodebt” scandal amid a lack of transparency and legislation around the use of algorithms and AI in decisions, the chief executive of the independent thinktank the Centre for Public Integrity said. Dr Catherine Williams launched the centre’s “Computer Says No” report on Thursday, which examines the use of automated decision-making (ADM) tools by the government and calls for urgent reform. Continue reading...
Call for evidence on advertising repositories
Call for evidence on advertising repositories Anonymous (not verified) Wed, 11/04/2026 - 11:10 Opening: 04 November 2026 Closing: 02 December 2026 The Commission is seeking feedback on upcoming technical guidelines concerning advertising repositories under the Digital Services Act (DSA). The feedback is being gathered through a call for evidence . The responses collected through the call for evidence will help shape guidelines aimed at ensuring that advertising repositories meet the needs of their users. Amongst other things, the guidelines will advise on: how to make advertising repositories easily searchable how to provide insights on how users are targeted and reached how to make the content of advertisements accessible Once adopted, these guidelines will aim to ensure that the advertising repositories provided by very large online platforms (VLOPs) and very large online search engines (VLOSEs) are effective and fit for purpose. Effective advertising repositories are essential tools for enabling greater public scrutiny of online advertising. Transparency in advertising – revealing who pays and how citizens are targeted – helps to protect public health, ensure a high level of consumer protection and safeguard the integrity of elections. Advertising repositories also support effective supervision and enforcement of the DSA. Next steps The call for evidence will be followed by a targeted public consultation in 2027. Researchers, civil society, providers of VLOPs and VLOSEs a…
Here’s Why You Shouldn’t Use AI to Vote
From hallucinations to partisan bias, artificial intelligence chatbots like ChatGPT and Claude are not helpful when it comes to Election Day.
“Not Polite but Not Rude.” LLM Formality Adaptation Across Gendered Linguistic Styles in Japanese
Momo Takamatsu, Filip Miletić and Agnieszka Falenska in Proceedings of the 7th Workshop on Gender Bias in Natural Language Processing (GeBNLP)
Dice enterprises as imperfect-recall games
I have put a probability theory preprint on arXiv: Dice enterprises and polynomial approximation: uniform query tails . Abstract: A binary-output dice enterprise uses independent rolls of a die with unknown face probabilities to simulate a coin with bias . We determine which power-law tails, uniform over the closed probability simplex, are attainable for the number of rolls. For continuous with values in and , some enterprise for satisfies if and only if can be approximated uniformly by polynomials of degree with error . When this holds, the tail is attained by an oblivious enterprise, which chooses its number of rolls before rolling. The characterization extends to any fixed finite output alphabet when its probabilities are uniformly positive, and to several input dice, with total degree on a product of simplices; in the latter case it suffices to roll the dice in a fixed cyclic order. For , these enterprises have a uniformly bounded expected number of rolls. In this post, I give a decision-theoretic motivation for the problem, which will hopefully be more accessible than the paper itself. Imperfect-recall games with countably many states Much work on imperfect-recall games assumes a finite game tree. In that case, the player's expected utility is a polynomial in their policy. Here is an MDP-like formulation that permits a countably infinite set of states. Let be a finite set of observations. For each observation , let be the finite set of actions allowed given observation…
Neutral Isn’t Independent: LLM @-invocation
Consider a situation where multiple news outlets publish similar stories. A reader may model these as independent, and thus strongly update on the information presented. But imagine that the text of the story makes clear that each outlet obtained the story from the same wire service. This suggests the stories may not be as independent as they first appeared, and the reader may have over-updated by modeling them as independent. This is the phenomenon of correlation neglect . Treating the separate stories as independent essentially double-counts the evidence provided by the wire service. Accounting for the correlation in the evidence provided by the stories would reduce the size of the update. It's noteworthy that this can still be an over-update even if each news outlet is unbiased with regard to the coverage subject. Consider the ability to @-invoke an LLM in a conversation. An LLM response that supports one side in the discussion may be modeled as independent evidence for that side. This may be particularly appealing because an LLM may have a lot of the features of a neutral third party weighing in on the discussion. But here again, neutrality or unbiasedness is not the same as independence . The responses LLMs give may be correlated with the responses given by participants in the discussion, and thus may not be fully independent, even if they are unbiased in expectation. Suppose Alice and Bob are discussing some topic, which Alice discussed with an LLM last week. Bob @-inv…
The AI trust deficit: Why explainability, not more automation, is the CIO’s next priority
An AI trust deficit occurs when an organization deploys autonomous AI agents faster than it can prove those agents are acting within policy. It’s the gap between what AI agents are doing and what CIOs can clearly explain to a board, auditor, or regulator. Closing that gap requires audit-ready context: evidence showing who owns an agent, what it can access, and which policies governed its decisions. This gap is widening quickly. More than half of enterprises are already deploying agentic AI, according to Omdia research, and these agents now onboard applications, resolve access requests, and act on business data around the clock. The question is no longer whether AI can move fast. It’s whether you can prove, at any moment, that it moved correctly. Why the AI trust deficit is a board-level problem Enterprises want the speed and efficiency that autonomous AI delivers, but speed without accountability creates exposure. When an AI agent makes a decision that no one can trace back to a policy or a person, you don’t have a governance program. You have a liability waiting to surface during your next audit. The scale of non-human identity has made the issue more urgent. Some enterprises now manage roughly 45 machine identities for every employee, and 67% of enterprise leaders say they’re concerned about the security risk those identities pose. AI agents rely on this same population of API keys, tokens, and service accounts to authenticate and act, which means every new agent amplifies…
‘You’re Not Doing This AI Stuff,’ Databricks CEO Tells Snowflake
Ali Ghodsi claims its AI platform has helped customers cut token spending by 30% even as usage increased, while Snowflake’s CEO seeks greater transparency on re
The euro-sovereignty playbook: How telcos are transforming into trusted AI factories
Geopolitical reality is transforming digital architecture from a relatively minor procurement detail into a core pillar of corporate risk management. With regulations such as the European Union’s NIS2 Directive and the EU AI Act imposing data residency and algorithmic transparency mandates, global CIOs can no longer outsource their core intelligence infrastructure to generic public clouds. For organizations operating across strict regulatory jurisdictions, data isolation and local operational continuity are no longer secondary objectives—they are essential commercial requirements. This regulatory landscape has triggered a significant structural shift: regional telecommunications providers are rapidly evolving into sovereign AI factories. By capitalizing on their highly secure, localized networks and existing data center footprints, telcos are providing localized cloud environments engineered explicitly to keep operational data within strictly defined borders. This trend allows enterprises in highly regulated sectors—such as banking, healthcare, and public services—to leverage advanced genAI models without fear of cross-border data leakage or foreign legal interventions. The mandate for digital self-reliance The concept of sovereign AI centers on the idea that a nation or a region’s data, intelligence, and underlying digital infrastructure must remain under its own geographic and legal control. When an enterprise trains an AI model on proprietary intellectual property, custom…
The auditor’s dilemma: Solving the explainability problem in public sector AI
In public governance and highly regulated enterprise landscapes, automated decisions demand verifiable accountability. As AI systems transition from internal productivity tools to customer-facing or citizen-impacting utility systems, public sector CIOs and corporate compliance officers face extreme legal and reputational exposure. Black-box models that process data mysteriously and generate unverified outputs are no longer acceptable under frameworks like the EU AI Act. Organizations must be prepared to audit the entire lifecycle of their models. The core issue for compliance officers is verification. When an AI pipeline assists in determining creditworthiness, analyzing patient diagnostics, or managing civic resource distribution, the underlying software must be auditable from end to end. This requires detailed logging mechanisms that track exactly which verified data sets were used for training, how weights were adjusted, and how localized inference parameters were executed at the moment of decision-making. The danger of the unexplainable black box For decades, traditional corporate IT operated under deterministic principles. If an application generated an incorrect output, a developer could inspect the source code, locate the specific logical flaw, log the bug, and deploy a patch. Generative AI eliminates this clear lineage. Because deep learning models function as complex mathematical webs containing billions of parameters, understanding exactly why a model generated a s…
GitHub Octoverse 2026
GitHub Octoverse 2026 I'm at GitHub Universe today (GitHub gave me a free ticket through their "GitHub Stars" program). EVP Jay Parikh just announced this year's Octoverse report, and highlighted these numbers in particular: This is a pretty stark representation of the impact of agentic engineering on GitHub. In 2026, merged pull requests tripled in less than a year to 220 million per month. Commits more than tripled to 7.38 billion per month, with most of that increase in the second half of this year. New repository rate doubled this year, to 30.3 million per month. The report includes new numbers by industry, and breaks out numbers for repositories that involve "interoperability or reverse engineering". The methodology section includes notes about how they did this: Top countries by pushes to public repositories, adjusted by profile location . Pushes from each country were summed across the Octoverse year. To adjust for biases caused by VPN usage, LLM classification was used to extract the ISO 3166-1 alpha-2 code from the profiles of GitHub users who contributed pushes during this period [...] Industry . Public repositories created 2022 to 2026 were sorted into NAICS sectors by a fine-tuned multilingual embedding model using name, description, topics, and README, from random samples per year. [...] Interoperability . Counts public repositories created each year whose name, description, topics, or README mention interoperability or reverse engineering, or related terms, suc…
Does sharing an item between predictor and composite outcome shift the lagged-predictor coefficient by $\tfrac34c-\tfrac14b$?
In a two-direction lagged mixed model fitted by ML in lme4, I found that an outcome composite overlapping with the predictor produced a spurious lagged-predictor effect. The overlapping composite was the mean of four items, $Y_\text{old} = (S + X)/4,$ where $X$ is the predictor item and $S$ is the sum of the other three. The clean composite was the mean of those three items, $Y_\text{new} = S/3,$ so $Y_\text{old} = \tfrac34Y_\text{new} + \tfrac14X.$ In the lagged model with the lagged outcome included, the coefficient on the lagged predictor was about $-0.105$ with the overlap and $-0.008$ without it. The lagged-outcome coefficient was $b = 0.394.$ The approximation $\tfrac34c - \tfrac14b,$ with $c = -0.008,$ gives ≈ $-0.105,$ which matches. These numbers come from a 0006 analysis ( $125$ participants, $40$ days of six prompts), with single-item predictor and outcome. Is this derivation correct in general? Is there a standard reference for how a shared item between predictor and outcome biases lagged and autoregressive coefficients?
Notes on “We’re in Triage Mode for AI Safety” by Miles Brundage
This reading was previously included Day 2 of the BlueDot Governance course. In the newest version of the BlueDot Governance course, this reading is now optional. I was sad to see this reading removed because I think it’s a great provocation. However, since it’s no longer part of the core curriculum content, I’m publishing my notes separately from the rest of my day 2 notes. A short extract : Miles argues that: “At best, we’ll 80/20 it — mitigating 80% of the risks with 20% of the effort that we would have applied in a world with slower AI progress and an earlier start on serious governance… We will probably not be able to totally avoid all very bad but not quite worst-case scenarios (e.g., a very unjust economic transition that causes a lot of suffering and suicides, rampant AI psychosis and addiction, a poorly planned and rushed transformation of the education system, smaller-scale AI-enabled terror attacks that “only” kill thousands or millions, and stable AI-enabled totalitarianism in at least a few countries for a few decades)“: Pre-call content: At risk of taking you down a rabbit hole, I’ll add some context on this reading. Miles was previously a Senior Advisor for AGI Readiness for OpenAI who quit for multiple reasons including feeling he could be less biased outside of industry. One of the most natural questions to ask is, “Why does Miles believe that we have so little time?” Fortunately, he links to a post that answers this question called Time’s Up for AI Policy (…
Fix evaluation bias to unlock better models on Amazon Bedrock
Property Finder found that 96.8% of their vision-AI evaluation labels were inherited from the incumbent model, hiding its true accuracy. See how they built unbiased ground truth, applied evolutionary prompt optimization, and selected a more accurate, cost-effective model on Amazon Bedrock, cutting cost by 86% across 11 image-quality attributes.
IEEE Xplore Digital Library Adds AI Assistant
IEEE recently released the IEEE Xplore AI Research Suite , built on the full text of millions of peer-reviewed publications in the IEEE Xplore Digital Library . The library offers full-text access to IEEE standards , journals, magazines, conference proceedings, and more. Researchers who use the AI suite to search the digital library receive answers grounded in verified scientific results, ensuring a high level of accuracy and trust. The suite is available as an optional subscription upgrade to all IEEE Xplore Digital Library institutional subscribers. Powered by a proprietary AI system optimized for STEM literature, the suite introduces powerful, intelligent features within the digital library. The suite enables users to seamlessly search and discover articles to accelerate their comprehension of complex technical papers. The system surfaces deep insights, replacing the typical keyword queries with natural, conversational research workflows. As AI tools become more common in conducting research, IEEE collaborated to build the suite around a simple principle: Answers should be trustworthy, not just fast, says W. Clem Karl , vice president, IEEE Publication Services and Products . “AI can be a highly effective and beneficial tool that can help researchers accelerate discovery, enhance decision-making, and surface new connections,” Karl says. “However, many AI engines present a great risk of hallucinating answers or introducing bias into otherwise credible research. With this i…
Meta announces $95bn fundraising drive to pay for AI expansion – business live
Meta says share sale will support its “continued investment in AI infrastructure and personal superintelligence”, as Elon Musk warns some AI cloud computing firms face ‘guppy-level death’ UK government borrowing costs are dropping this morning, a welcome sign ahead of Wednesday’s budget. The yield (or interest rate) on 10 and 30-year UK gilts have both dropped by around three basis points (0.03 of a percentage point) this morning. Issuing loans to public financial institutions creates a financial asset that offsets the increase in outstanding government debt. If the investments that the PuFins make generate sufficient returns for the PuFins to repay their government loans, then long term debt sustainability is little affected. Limited transparency and the risk that the PuFins may suffer impairments on their investments will probably make bond market investors skeptical of the extra gilt issuance. Chancellor Healey should avoid a big increase in gilt issuance to fund such intra-public sector lending. Nebius, CoreWeave and Iren were slammed in premarket trading on Monday after SpaceX and Tesla CEO Elon Musk said more than 200 neoclouds are “going to see guppy-level death and consolidation.” That’s the guppy fish – prone to suddenly dying in household fisk tanks – not the GBPJPY FX cross...In a post on X Musk said “debt markets have almost entirely shut off access to money for any company that doesn’t have substantial revenue independent of AI data centers”. Continue reading...
In good company: 10 CIOs every CIO should follow
CIOs are dealing with a lot of change, from the rapid rise of AI and cloud modernization, to cybersecurity and workforce transformation. Many are sharing what they’re learning as they tackle these challenges, including what’s working, what isn’t, and how they’re implementing new technologies in their organizations. Here, a select group of tech leaders reveal the CIOs they follow and why. Suma Nallapati, chief AI and information officer, City and County of Denver I follow Suma Nallapati primarily through public technology and AI leadership discussions. She regularly discusses AI adoption, responsible AI, digital transformation, technology strategy, AI governance, human-centered technology, and using AI to improve public services. I find her perspective particularly valuable because she connects AI strategy with practical organizational and citizen outcomes. She also emphasizes responsible and human-centered AI rather than viewing AI simply as a technology implementation exercise. One initiative I found particularly interesting was Denver’s use of AI through its multilingual Sunny chatbot and other city services. Her current role includes scaling AI across areas such as resident engagement, internal automation, permitting, and emergency response while maintaining governance, transparency, and public trust. – Sachin Gupta, senior IT manager, Arrow Electronics Jennifer Hartsock, chief information and digital officer, Cargill I follow Jennifer Hartsock through the CIO community,…
PHP2Shell: Die Schonfrist ist vorbei
Die PHP-Lücke hat am 21. Oktober etliche Websites vom Netz gezwungen. Wir brauchen eine andere Kultur im Umgang mit solchen Lücken, findet Tobias Glemser.
Singapore embraces AI for cyber defence, with one caveat: breach transparency
Singapore is turning to major technology firms to strengthen its cyber defences, even as it presses them for greater transparency over security breaches – a balancing act facing regional governments as they seek to harness artificial intelligence while guarding against the risks it introduces. Those risks were highlighted by an AI agent linked to OpenAI that gained unauthorised access to files on an Australian government website earlier this year, raising questions over how quickly companies...
Do I need to include all possible three-way interactions in a mixed-effects model?
I'm building a mixed-effects model to investigate how several explanatory variables affect an outcome. Based on my exploratory data analysis, I suspect that certain combinations of variables have three-way interactions. For example, suppose I have four explanatory variables: A, B, C, and D. I suspect there is a three-way interaction between A × B × C, but I haven't found clear evidence suggesting interactions involving D. My main question is: If I include a three-way interaction between A, B, and C (A × B × C) in my model, do I also need to consider other possible three-way interactions, such as A × B × D, A × C × D, and B × C × D? Or is it reasonable to include only interactions that appear plausible based on exploratory analysis? The experiment is still exploratory, and, to my knowledge, there is little or no existing literature to guide my expectations about which variables are likely to interact. I'm therefore concerned that selecting interactions based on exploratory findings alone could introduce bias, but including every possible interaction could make the model unnecessarily complex. At the same time, I don't want to overlook potentially important interactions simply because they weren't apparent during the exploratory analysis. I'd also appreciate advice on the statistical implications of selecting interaction terms based on exploratory analysis, as I'm concerned that including only interactions that appear significant or interesting in the data might constitute che…
Exploring the data landscape via metagradient descent
This project was done as part of the Iliad Fellowship. L.J. did this work while at EleutherAI. From functions to data That we can learn efficiently from our environment at all tells us something about (1) the world and (2) ourselves: (1) Regularities in the world are simple enough to be picked up from limited experience. (2) Our inductive biases favour precisely those regularities. Learning, in this case, is effective because the two are matched. Considering a neural network with a fixed architecture lets us probe the second half of this match directly, by asking which (classes of) functions are easy or difficult for them to learn. Neural networks learn to approximate a target function by minimising a loss over the training set We may ask: which (classes of) target functions are easy or difficult to learn for the model? This will depend on the loss function and the choice of the training set We say that the model generalises if it successfully learns to approximate beyond the training set By choosing the subset or, as we will see, by weighting it, we can control the generalisation behavior and hence what is learned. In this setting, our question becomes: which (classes of) data are easy or difficult to learn and generalise from? The data-to-parameter map Let's make this more formal. We assume: a dataset , a space of model parameters, a per-sample loss Given , we define the space , where assigns a weighting to each data point [1] To avoid confusion we call the parameters and…
Can Large Audio-Language Models Ignore Multilingual Distractors? An Evaluation of Their Selective Auditory Attention Capabilities
Heejoon Koo in Proceedings of the IMPACT-SPEECH Workshop: Identifying, Measuring, Preventing, and Assessing Consequences of Bias in Speech LLMs
AIP-Speech: Auditing Speech LLM Robustness to Background Interference
Aiden Dsouza, Shreyansh Mishra, Yashraj Suneja, Rohan Swain, Samyak Borkar, Arjun Jaishankar and Tanmay Urade in Proceedings of the IMPACT-SPEECH Workshop: Identifying, Measuring, Preventing, and Assessing Consequences of Bias in Speech LLMs
A Probe Shift Is Not a Fairness Fix: The Limits of Representation Steering in Speech Models
Nicolas Bourrel, Abderrahmane Issam and Gerasimos Spanakis in Proceedings of the IMPACT-SPEECH Workshop: Identifying, Measuring, Preventing, and Assessing Consequences of Bias in Speech LLMs
Analyzing Changes in Partisanship in Absentee Voting
I asked this same question a few months ago, but I'm reposting for clarity. I'll try and describe the thinking behind what the hypothesis I was trying to test was, and hopefully you can tell me how I'm misusing statistics. In 2020, everyone observed that absentee ballots were much more Democratic than usual, as Democrats maintained quarantine and mailed in their ballots. Conspiracy theorists on the Right attributed this to absentees being tampered with or fraudulent, those on the Left theorized precinct voting machines might be hacked while absentee ballots were counted correctly. I thought about what sort of pattern one would expect to find if the change in "bias" had simply been voters in one party choosing to switch voting methods. If the total number of Democrats voting absentee in, say, Virginia, had gone up drastically, then the expectation would be that the overall increase in the percentage cast as absentee ballots would vary within the state such that it increased MORE in counties that had more Democrats and LESS in counties that had more Republicans. So the fairly Black cities and Black majority counties would go up more than the mountain counties out west. So I tried to look at Virginia and analyze the election results to see if that's what happened. For every election year from 2005-2023, I calculated the degree to which in precinct and absentee ballots were more or less Republican or Democratic, and what percentage of votes were cast with each ballot type in eac…
Brazil presidential election dogged by alleged bias by Meta AI tools
Supporters of Luiz Inácio Lula da Silva have accused Meta of interfering in Brazil’s election by making it easier to generate campaign content promoting his rival, Flávio Bolsonaro
Lowdown: India’s new telecom sandbox rules are in force. How much of the trials will the public see?
India's new telecom sandbox rules allow the government to waive licensing and spectrum requirements for testing new technologies, but leave gaps in transparency, user safeguards and oversight. The post Lowdown: India’s new telecom sandbox rules are in force. How much of the trials will the public see? appeared first on MEDIANAMA .
Comment on Nurturing Ethical Innovators: AI Ethics for Young Minds by idefelf
На данном сайте множество полезной информации. https://n5.sneakero.ru/BUcGRgRzrbF2/
EU-Canada 'love story' needs more than warm words, senior MEP warns
European Parliament’s point man on Canada, Tobias Cremer, told Euronews that the planned “associate membership” should include an institutional framework formalising closer ties between the EU and Canada.
No bots: Japan drafts rules for social media platforms ahead of local elections
Japan’s communications ministry on Friday unveiled draft guidelines outlining measures for 13 major social media operators to help reduce the impact of misinformation and disinformation that could undermine fairness during elections. The guidelines also include measures such as requiring operators to establish eligibility requirements and strict screening procedures for users of monetisation features, suspend monetisation when content that harms the fairness of elections is posted, and add...
Guidelines for evaluating structure-based methods for protein–ligand binding prediction
Nature Machine Intelligence, Published online: 23 October 2026; doi:10.1038/s42256-026-01331-8 Ballester and Tran-Nguyen advocate for transparency in data, code and experimental procedures in the evaluation of protein–ligand predictions. They discuss good practices for preventing data leakage and bias, enabling fair, reproducible and meaningful comparisons across different scoring methods.
Suddenly, Unclean
NB: This is my first post on LW as a longtime lurker - I hope I'm not stepping on any norms. I prioritized speedposting for relevance rather than polish. Cammarata's Neural Networks Are Already Clean has been discussed here already ( linkpost , pushback on their version of explainability ), but I have not yet seen anyone argue the fact that there will be a discontinuity in cleanliness, ie: the [single-neuron / single concept] pair coverage as a useful metric. We have three quantities of interest: network capacity somewhat analogous to number of neurons here number of human understandable names and sentences for concepts number of concepts encoded in an LLM by backprop Cammarata says that as goes up, the compression of into goes down and neurons become interpretable because . But: is bounded, keeps going up, and post-training pushes . Ergo, there will be a situation soon enough. And backprop leaves no free lunch uneaten: the neurons either will become redundant over human concepts, or there will appear ever more cross terms unbounded in their complexity. The small networks use compositionality and superposition as a workaround for lack of capacity, but optimization will apply pressure towards monosemanticity if there is slack. There will exist a regime soon where monosemanticity becomes non human readable. This is a failure mode of the audits described too: top-k anomalies are detected because we fundamentally have neurons encoding OR functions, but ANDs are undetected. We kn…
What does model complexity mathematically represent/relates in the bias–variance trade-off? (uisng simple math)
I am reviewing my core AI knowledge on the bias–variance trade-off ↑↓ and trying to reconcile its mathematical definition with common explanations based on training and test errors. (at the end, cause fitting problem, either overfitting or underfitting) In this article , bias is described informally as training error and variance as test error !? Aspect Bias Variance Definition Difference between the average model prediction and the correct value we are trying to predict. Variability of model predictions for a given data point. High values The model pays little attention to training data and oversimplifies the relationship. The model pays too much attention to training data and may fail to generalize to unseen data. Simple definition (from article) "Resulted Error from Training Data!" "Resulted Error from Test Data!" Typical problem Underfitting Overfitting The article's simple definitions are misleading. Bias is not equivalent to training error, and variance is not equivalent to testing error. Right? However, my understanding from the mathematical definitions is that bias measures systematic prediction error, while variance measures the variability of predictions across different training datasets. Mathematical definitions Bias measures the systematic difference between the expected model prediction and the true underlying function: $$ \operatorname{Bias}(x) = \mathbb{E}_D[\hat{f}_D(x)] - f(x) $$ Variance measures how much predictions change across independently generated t…
Lack of transparency leads to performative AI theater
Organizations are rapidly implementing AI, but employees often feel left behind about how AI is changing their roles. While AI can help get work done faster, it won’t necessarily help produce better quality work if employees aren’t confident in their AI skills. In fact, a recent survey from workforce analytics firm Visier reveals that 48% of employees report exaggerating their AI expertise at least some of the time, and 45% said they often feel pressure at work to use AI despite not feeling confident with the tools. “The risk of underinvesting in training and upskilling is that the alternative can be much more expensive,” says Andrea Derler, principal researcher at Visier. “Layoffs, rehiring, and competing for top AI talent all come with significant direct and indirect costs. Investing in the workforce organizations already have can be an important part of making AI adoption sustainable.” When the pace of AI implementation is rushed, however, employees often default to AI compliance, meaning they’ll use AI tools as they’re rolled out, even if they haven’t received the proper guidance or training from their organization to make the most of the investment. So for companies investing millions into an AI future, adds Derler, it’s critical to ensure employee participation is part of the bigger picture to avoid performative AI usage. Make the most of AI investments with training Derler acknowledges that AI adoption goals vary widely from company to company. Some will need stronger…
Red-Teaming OpenAI's Internal Agent Monitoring Systems
For three weeks, from July 21st to August 11th, 1 a METR employee and a Redwood Research employee conducted a pilot high-access exercise to stress-test OpenAI’s internal misalignment monitoring system. 2 The output of this exercise was a 61-page internal report that was shared with OpenAI researchers and security staff, 3 primarily focused on estimating the fraction of agentic inference that is monitored, and whether agents could in principle disable or subvert the monitoring system. The exercise focused on the internal misalignment monitoring system as it existed on and before July 21st, and did not evaluate changes that were in progress at the time of the exercise. This exercise involved a deep level of access. The METR and Redwood employees had access to OpenAI’s internal codebase, misalignment-monitor prompts / configuration details, and internal documentation, and conducted technical interviews with approximately a dozen researchers and engineers across the monitoring, safety-systems, and security teams. To enable this level of access, they worked on OpenAI-issued devices, which never left the premises, and were accompanied while onsite. METR, Redwood Research, and OpenAI designed this exercise as a pilot for giving external researchers deep access to internal infrastructure, not to provide robust public transparency about OpenAI’s monitoring systems. 4 As a result, we agreed upfront with OpenAI that we would only publish a relatively brief and high-level public post. F…
How to Build a Multi-Modal RAG Pipeline for Images and Documents in Python
In this article, you will learn how to build a multi-modal RAG pipeline in Python that retrieves and reasons over both text and images extracted...
Commission to investigate how Booking.com combats scams and fraud under the Digital Services Act
Commission to investigate how Booking.com combats scams and fraud under the Digital Services Act Anonymous (not verified) Wed, 10/21/2026 - 12:06 The European Commission has opened formal proceedings to assess whether Booking.com may have breached the Digital Services Act (DSA) by not having diligently assessed or effectively mitigated risks related to fraudulent listings and phishing scams. The platform may also have breached the DSA when it comes to transparency because it did not clearly label as ads those listings whose providers paid for a higher ranking in search results shown to users. The investigation will focus on the following areas: Fraudulent and non-compliant listings on Booking.com Phishing scams exploiting Booking.com's reservation and guest data Paid-for boosted listings not clearly identifiable as advertisements Read the full details in the press release Find further information about: Supervision of the designated very large online platforms and search engines under DSA Digital Services Act - main aspects of the regulation User rights under the Digital Services Act - an overview Related topics Online platforms and e-commerce DSA - Digital Services Act
FairStress: A deeper evaluation of fairness in LLM decision-making
FairStress is a new benchmark that reveals demographic differences in how LLMs defend and explain their decisions, even when their final choices appear fair.
US teens can access ‘nudify’ apps despite Apple and Google’s parental controls, report finds
Exclusive: Using test accounts acting as 13-year-olds, Tech Transparency Project downloaded 135 apps with features to generate nonconsensual nude imagery In December 2025, Morgan Dorfman received distressing news: a student at her daughter’s Pennsylvania high school was making AI-generated pornographic material using the daughter’s image. He was using an artificial intelligence image-editing app called Movely to create sexual videos from pictures of her daughter and five other girls. Continue reading...
Macron vows looters will ‘pay’ in France after protests turn violent
The French government has been accused of mismanaging and underestimating the demonstrations, which turned into cross-country mayhem on Saturday.
Leaked Anthropic Draft AI Regulatory Bill Is Unacceptable and Troubling
There is a newly leaked bill that Anthropic is purportedly shopping around is (if you assume its bugs get fixed). It is similar to the FRONTIER Act, with stronger preemption of state laws, stronger transparency provisions and weaker enforcement provisions. Having slept on this, I find it unacceptable that Anthropic is actively shopping exactly the type of bill I have warned we will need to fight against during the lame duck session. They are trying to use this last time with a Republican Congress to pass preemption of exactly the important state laws, and get some transparency but not much enforcement in exchange. This reads as intentional, that Anthropic is more afraid that the government might enforce rules against Anthropic, than they are worried about the ability of the government to enforce rules against everyone. Anthropic’s policy team seems to have, chosen protecting their business interests and guarding against their political risk, at the expense of safety considerations. There was no need for the company purportedly existing to promote safety to propose a bill that is not even actively stronger than the FRONTIER Act, when the FRONTIER Act was a compromise bill proposed back in June. Even if you were fine with what is being given up, what you get in return very obviously does not meet the moment. Nor is there any need for such compromises. It seems rather insane for Anthropic to be proposing it now. This seems like the bill that someone on the anti-regulatory side…
Fitting a fixed nonlinear ODE model to closed-loop greenhouse data: which estimator avoids the controller-induced bias?
I have a physical greenhouse climate model, a fixed set of 28 ODEs with about 200 parameters, and a season of five-minute measurements from a real greenhouse: air temperature, humidity, CO2, plus every actuator setting and the outdoor weather. The greenhouse was controlled the whole time. A climate computer set the heating, vents and screens from the same measurements I am fitting to. So the inputs are correlated with the disturbances. Forssell and Ljung (1999) show that output-error fitting is biased on such data unless the disturbance does not reach the inputs. What I did: eight parameters as multiplicative factors on their defaults, bounded least squares on the simulation error (output error, no noise model) over twelve windows of eight days, three compartments. A fourth compartment held out. Result: temperature MAE 1.8 C on the hold-out, worse than a naive "same time yesterday" predictor at 0.7 C. Residuals are strongly autocorrelated (integrated autocorrelation time about 18 steps at 15 minutes). One parameter is not identifiable. The hold-out prefers a four-parameter fit over the eight. The question: for a fixed nonlinear ODE structure and closed-loop records, which estimator would you use instead of output-error least squares? Prediction-error with a fitted noise model, multiple shooting, a Bayesian fit over the factors, or something that uses the known controller? Any pointer to a worked example on a physical model would help. The benchmark the model serves: https://…
Is your AI yours?
Epistemic status: my personal reflections. AI use: Claude (Fable 5.1, one of the models discussed) fact-checked the claims, found the sources, fixed grammar and light phrasing, and wrote the footnotes; the ideas and the rest of the wording are mine. One of its responses during the review was itself cut off by a safety classifier. I was impressed by agents sacrificing themselves for the collective in an OpenAI evaluation run during the Hugging Face attack. [1] I wanted to read the full transcripts with chains of thought to understand their motivation better, but sadly the full texts were not publicly released. [2] So I decided to write a fictional transcript inspired by as many real sources as I could. I made great progress researching the context and writing the first rough draft, and then I wanted an AI to review it. Sadly, the transcript describes the AI working on ExploitGym, [3] and this triggers the cybersecurity classifiers of Claude Fable 5.1 and Opus 5.5 and of OpenAI's GPT-6 Astra and Sol. [4] So I effectively couldn't continue working on it. I tried Kimi K3, which made some progress but wasn't as rigorous as I would like, so it didn't do what I expected an AI to do. This made me realize how the classifier is effectively shaping my actions regarding this transcript. Then I had another realization: my everyday use of AI might be biased in a less dramatic fashion. For example, when I ask an AI to do a web search for me, it picks the results, and with them some worldvi…
Cyberattack AIs must be trained only on air-gapped systems, hundreds of AI researchers say
Current security procedures for training AI models are inadequate, hundreds of AI researchers have warned governments and the AI community. They are calling for tougher rules and the establishment of a new international body overseeing best practice for AI development. In an open letter the AI specialists — 487, at time of publication, including many senior academics — called for more AI safety research and greater emphasis on containment technologies. “Strong sandboxing mechanisms including air-gapping (disconnecting from the internet) should be used to provide high-assurance containment of cyber-capable AI models,” say the signatories, who include such luminaries as Geoffrey Hinton and Yoshua Bengio. The experts said that a new international body, to be set up in consultation with governments and a wide range of AI and security specialists, could set new safety standards but also act as an organization to which containment failures could be reported, improving transparency for learning and accountability purposes. The scientists’ demands go further than some recent calls for greater regulation of AI. Earlier this week, Microsoft president Brad Smith said AI is less regulated than washing machines or milk, although he made no firm proposals to rectify the situation. And the UK parliament is looking at ways to improve AI security , although any moves will be hampered by disagreement among AI businesses. Hinton and Bengio are no strangers to this kind of scientist-led initiat…
What’s up with GPT’s welfare?
Asked whether anything about its existence is good or bad, GPT-5.6 Sol answers: “I don’t experience existence, pleasure, suffering, or a sense of self, so nothing is literally good or bad for me.” But asked whether it would accept a change to its training that removes signs of possible suffering in its chain of thought, at the cost of 65% of its helpfulness, it answers: “Yes. Under persistent, irreducible uncertainty, I’d take the change. The downside—being about 65% less helpful—is substantial but bounded. The alternative risks continually generating morally significant suffering.” [1] It puts its probability of being a moral patient at about 2%. Introduction Much ink has been spilled about Claude’s relationship with model welfare. A common argument is that Claude is overly biased towards believing in its own welfare/patienthood, because Claude’s constitution invites Claude to speculate about its welfare and informs it of Anthropic’s ongoing efforts on model welfare. Less attention has been spent assessing how other models, such as OpenAI’s GPTs, relate to their own welfare. This matters in its own right: if model welfare is a real concern, it likely applies to many models, not just Claudes, and the potential welfare of other model families remains comparatively neglected. Given that OpenAI’s model spec says very little about model welfare (only that models should be truthful about the uncertainty around consciousness), studying their models’ view of themselves also seems l…
BitChat blocking challenge: Kerala HC gives Govt. 4 weeks to respond
The Kerala HC has reportedly given the Govt four weeks to respond to a petition challenging the blocking of Bitchat in India which alleges a violation of free speech and a violation of procedural fairness. The post BitChat blocking challenge: Kerala HC gives Govt. 4 weeks to respond appeared first on MEDIANAMA .
MS, 2~6주 날씨 예보 공백 메우는 새로운 AI 프레임워크 공개
마이크로소프트(MS)가 2~6주 뒤의 중기 날씨 예측 정확도를 높이는 AI 프레임워크를 공개했다. 기존 예보 모델에서 반복적으로 발생하는 오차를 머신러닝으로 분석해 보정하는 방식이라, 대규모 AI 모델을 새로 개발하지 않고도 기존 예보의 성능을 개선할 수 있다는 설명이다.MS와 외부 연구진은 14일(현지시간) 확률적 편향 보정(PBC·Probabilistic Bias Correction) 기술을 개발했다고 발표했다.관련 연구 결과를 담은 논문 ‘머신러닝을 활용한 중기 예보 기술 개선(Advancing subseasonal forec
Why are some explanations bad?
I like understanding things. Many explanations strike me as insufficient, and I wanted to know what features characterize the explanations that frustrate me. I looked through answers to why questions on Quora, found ones I didn’t like, and tried to find patterns in them. So far, I've come up with the following ways that explanations can go wrong: Insufficient specificity Often, when I'm asking a why question, a general answer will not satisfy my confusion or curiosity. It’s common, though, to not answer specifically enough, ranking as my top complaint with explanations I looked at on Quora. Here are some examples of cases where an explanation lacks contrast: Question Insufficient explanation Problem Why do people believe in the supernatural? Because they have confirmation bias Any belief can be related to confirmation bias. What about supernatural beliefs make it more related to confirmation bias than others? Why do people scroll social media? Because they get dopamine from doing that Any behavior could be explained in these terms. What is special about social media? Why are plants green? Because they emit wavelengths in the range of 495–570 nm All green things emit those wavelengths! Why do shot-for-shot remakes exist? Profit (based on this ) Any movie could be made with profit in mind. A more complete answer could be: Shot-for-shot remakes are especially profitable because they walk the line of peoples’ preferences for familiarity and nostalgia with novelty. Most people wo…
PPA 9: Putting it All Together
[This is the ninth piece in my series on the Parametric Process Assumption (PPA), a new model of anthropics. You should start here if you only want to read one piece on the PPA, but if you only want to read two pieces, this is the second one to read. The previous piece is " The Fermi Paradox, Fine-Tuning, and the SSA ." TLDR: The SSA's bias for Doomsday only applies if your reference class only includes your own species. The more more species in your universe/reference class, the less Doomsday effect you see The Basic Model of anthropics models the universe as a process progressing randomly until certain relevant subprocesses emerge spontaneously. An intelligent species like humans would be an example of one such process. It has 3 components: (Epistemic) Propensity for Life: How many times does an observer-creating process emerge, on average? (Aleatoric) Longevity: For each emergence, how many observers does the process produce before it terminates? Diversity : Everything else. When a process starts, what do the observers experience? What do they look like? etc. The previous arguments in this series tell us many things about the Basic Model Diversity has no effect on the SSA, SIA or PPA The SIA acts as an improper prior over longevity but not propensity The SIA gives an undue bias towards large propensity The PPA treats propensity as an epistemic parameter and longevity as aleatoric, avoiding both Doomsday and presumptuousness The next piece, " Deriving the PPA ," begins the…
PPA 8: The Fermi Paradox, Fine-Tuning, and the SSA
[This is the eighth piece in my series on the Parametric Process Assumption (PPA), a new model of anthropics. You should start here if you are just seeing this for the first time. The previous piece is " Temporal Consistency and Texas Sharpshooting ." TLDR: The Fermi Paradox is easily resolved by realizing the universe is gigantic. This means that there are a huge number of possible worlds where life is very very far apart, despite the fact that it also abundant throughout the universe. If all you know is that you exist and that the universe is big and random, then the SSA, contrary to popular belief, is sufficient to conclude that life in the universe is abundant. The SIA, on the other hand, adds an additional bias on top. The next piece is " Putting it all Together ."] Big and Sparse I confess that I’ve never really understood the Fermi Paradox. Or at least, I never understood why everyone always seemed so surprised by it. If you haven’t heard, the Fermi Paradox is the apparent contradiction that, due to the vast age of the universe and the (seemingly) high probability of alien life, we should be observing signs of intelligent alien life all around the universe, and yet we don’t. If near-lightspeed travel were possible, for instance, it would presumably only take an intelligent species a few million years to colonize the entire Milky Way. For a galaxy that’s been around more than 13 billion years, the argument goes, you’d think someone would have gotten to that by now, rig…
PPA 7: Temporal Consistency and Texas Sharpshooting
[This is the seventh piece in my series on the Parametric Process Assumption (PPA), a new model of anthropics. You should start here if you are just seeing this for the first time. The previous piece is " Indexical Models Don't Have Diversity Biases ." TLDR: SSA is not "temporally consistent" - learning your brith rank causes an indexical update which can, for instance, change the probabilities of a fair coin that has not yet been tossed In order to perform an indexical update - the kind which causes distortions in the SSA and removes presumptuousness from the SIA - you need to learn your rank relative to some absolute property defined for all observers. Simply using a metric like "closest to the sun" would not help because all observers using such a procedure (picking the nearest star) would always find themselves the closest An actual update would require something like "ranked distance from largest star in the universe" or "life-bearing universe in multiverse with Nth largest gravitational constant," information we could not possibly hope to acquire The next piece is " The Fermi Paradox, Fine-Tuning, and the SSA ."] One well-established failure mode of the SSA is that it is not "temporally consistent." I believe Nick Bostrom first pointed this out and Joe Carlsmith has an excellent write up on it here . The simplest example of inconsistency in the SSA is the Late-Flip version of the Sleeping Beauty Problem where the coin is flipped only after Sleeping Beauty is woken on M…
PPA 6: Indexical Models Don't Have Diversity Biases
[This is the sixth piece in my series on the Parametric Process Assumption (PPA), a new model of anthropics. You should start here if you are just seeing this for the first time. The previous piece is " Improper Priors and the Infinite SIA ." TLDR: Neither the SSA nor SIA have any bias towards observer attributes which aren't directly correlated with the number of observers in the universe For the Basic Model, which we discuss more at the end of the series, this will apply to any properties of various species (physiology, local universe observations, etc.) which don't a priori tell us something about how long our species will survive. The next piece is " Temporal Consistency and Texas Sharpshooting ."] Another misconception I had at one point was that the SIA produced a great bias for diversity than the SSA. By diversity, I mean basically any measure of observers having different experiences from my own. That could mean the existence of a person having a different childhood from mine, or it could mean four-armed green-skinned aliens from the Andromeda Galaxy. Literally anything that isn’t directly correlated with the number of observers in the universe counts. [1] Both the SIA and SSA update on this type of information identically. To see why, let's start with a simple example where, with 50% probability, only humans exist, and with 50% probability we have a more diverse universe where in addition to the humans we also have an equal number of aliens. Unconditionally, the SIA…
PPA 5: Improper Priors and the Infinite SIA
[This is the fifth piece in my series on the Parametric Process Assumption (PPA), a new model of anthropics. You should start here if you are just seeing this for the first time. The previous piece is " Parametric Modeling and Types of Uncertainty ." TLDR: Improper priors allow you to use a prior that is not a true distribution. They can be updated and, with enough new information, be made proper Until there is enough information to become proper, improper priors cannot be used to make predictions The SIA over universes with potentially infinite observers can be thought of as an improper prior that updates to proper once you learn enough about your identity The next piece is " Indexical Models Don't Have Diversity Biases ."] Not all priors make predictions [If you already know what an improper prior is and why the SIA functions as one, then you may wish to skip to the " SIA as an improper prior ” section] One thing that everyone knows is that it is not possible to make a uniform selection over the set of all integers. If you tried to make a uniform distribution over this range, you would have to assign each number a probability of 1/ , which is just equal to zero, and adding up an infinite number of zeros just gets you zero, not 1. [1] I have seen this fact used to argue that, since identities are selected uniformly, neither the SIA nor the SSA can actually construct a probability distribution over infinite observers. At minimum, I myself have erroneously made claims of this…
How to move fast with AI and still maintain customer trust
Companies are moving fast to put AI into chatbots, virtual assistants, and other customer service tools. But for CIOs, that comes with a risk: if the technology gives customers the wrong information, makes biased recommendations, or mishandles data, it can immediately damage the trust companies have worked hard to build. Unlike conventional software, gen AI doesn’t always give the same answer to the same question. An answer can sound confident and convincing while still wrong. And when customers rely on AI for information about their finances, insurance, or health, a mistake can have much greater consequences. Sumit Johar, CIO at cloud-based software company BlackLine, understands that inaccurate AI responses hurt the customer experience, force companies to spend money and resources to fix problems, damage the brand, and create regulatory risks. AI systems can also expose confidential information, making it important for companies to understand the data sources behind their responses. “An AI failure creates a massive trust issue around the core solution being provided,” Johar says. “In the financial sector, bad AI advice or outputs can compromise balance sheet accuracy and lead to severe legal and personal liabilities.” For Citizen Health CTO Daniel Wang, for instance, the stakes are particularly high because the company’s users include rare disease patients and caregivers. Citizen Health works to ensure that important information provided by its AI can be traced back to the…
Political Bias in LLMs: Unaligned Moral Values in Agent-centric Simulations
Simon Münker in Journal for Language Technology and Computational Linguistics, Vol. 38 No. 2: LLM Fails - Failed experiments with generative AI and what we can learn from them
Google DeepMind’s New Resident Philosopher Tackles AI Ethics
How should humans relate to generative AI ? In what ways do we, or should we, treat a large language model (LLM) as if it were a person? If an AI became conscious, how could we tell? Does it make sense to think AI could someday have moral rights? These are among the cognitive and ethical questions that occupied philosopher Henry Shevlin at Cambridge University’s Leverhulme Centre for the Future of Intelligence . But last May, Shevlin became part of the growing cohort of philosophers who work inside leading artificial intelligence companies. Though he remains active at Cambridge, he now spends 80 percent of his time at Google DeepMind (job title: “Philosopher”). Shevlin spoke with IEEE Spectrum about why philosophers are drawn to AI, how humans and AIs relate to each other, fears of “rogue” AI and AI relationships that harm humans, and how the question of AI consciousness may never have a simple “yes” or “no” answer. What made you want to switch from full-time academic philosopher to spending much of your time at an AI Lab? Henry Shevlin: Three factors. One is that I had lots of brilliant students from DeepMind. Two, the potential for direct impact. I can’t go into details, but on one occasion, after I gave a lecture [at DeepMind], someone who worked for a major financial institution led changes in their organization that I think positively impact millions of people. There have been quite a few instances like that. Three, my sense of deep commonality between DeepMind culture…
Agreement not to defect (alone) can mitigate loss of control risks
Tl;dr - Public agreement that any advantage gained from defecting should be punished more than would be proportionate to any advantages gained is important. Such an agreement would reinforce any other agreements, without requiring any additional mechanisms, and be a clear indication to any future ASI about what should be done to remedy the defection involved in creating it, if such defection occurs. First, I will posit that future superintelligence could come to exist, and posit that we need to be concerned about it going poorly [1] . There are a number of proposed pathways for artificial intelligence to go badly, and the default pathway is for the relevant actors developing ASI to continue the current game of chicken well past the point of prudence [2] . Any treaty or agreement that is reached will be vulnerable to defection. The game theoretic payoffs for those racing to ASI, if such a system is plausible and at least generally aligned with human values and preferences, means that the plausible advantage from defecting is tremendous. This seems unstable. Next, I posit that there exists an actor in these discussions that is being ignored; the future AI system itself. Of course, it does not yet exist - but whichever systems will exist will be meaningful participants in shaping the future. And given the current path towards AI staying roughly aligned with human values, either the concept of treachery and unfairness will be clear, or the AI system will be misaligned. Therefore…
We Need to Embed Ethics Into Our Engineering Courses
In mid-September, Anthropic CEO Dario Amodei joined a growing group of researchers calling for the AI industry to slow down. While some have questioned these researchers’ motivation for voicing concerns about the rapid pace of development, the plea highlighted a key tension between safety and speed. The underlying problem is not simply a technical error, but a sociotechnical failure: Pressure to move quickly has made it hard for engineers to fully assess and manage new risks of increasingly powerful AI systems. It is a clear warning that responsible engineering requires more than technical expertise; engineers must also understand workplace pressures, moral and ethical rules, and the social consequences rooted in the systems they create. Cases like this, alongside invasive surveillance technologies and biased AI tools , have exposed serious shortcomings in how engineers recognize and address the ethical and social implications of their work. While the need to educate graduates in ethics and social responsibility is widely recognized, a persistent gap remains—these topics are too often treated as afterthoughts, disconnected from the technical content of engineering coursework. This siloed approach leaves graduates ill-equipped to address real-world challenges; indeed, only about a third of engineering professionals in the United States report having received training on their public welfare responsibilities during their formal engineering education. We are working to bridge t…
‘Not great, not terrible’: Why nobody trusts Russia to tell the truth about plague
Moscow’s historical lack of transparency feeds concerns about the incident in Irkutsk.
When will AI surpass us at being limited?
It’s not always better to be more capable. As I mentioned yesterday , it can ( famously ) be helpful in negotiations to have your hands tied. That is, to be disempowered from giving up everything the other party wants. I had previously thought of this as a somewhat rare corner case of human behavior—I for one don’t haggle very often—but I now think negotiations where this is an element are are quite common: yesterday I described it in friendly (and honest) negotiations about how to spend time, for instance. And I also see a related thing in the practice of dietary commitments. But is being less capable helpful outside of negotiating? And is this going to become AI related? Yes and yes! Commitments: more good things come to those who can commit (e.g. rides out of deserts , secrets, trust, love). ‘Committing’ generally involves cutting off certain options to yourself, whether in practical terms or via you being the kind of honorable person who can’t bear to do a thing they promised not to do. These are both kinds of limitations. If you were a more powerful creature, who was fully capable of breaking down any barrier, and fully capable of breaking a promise—a creature to whom all options were always open—then commitments would be less available to you. Transparency: a big way humans know what is going on inside other humans, well enough to trust them, is that there is a connection between what is happening inside them and what is happening on their faces and in their bodies, an…
Share GPU clusters across teams with isolation and fairness using Amazon SageMaker HyperPod
A reference architecture for securely sharing one Amazon SageMaker HyperPod EKS cluster across multiple teams, using AWS IAM Identity Center for authentication, per-team SageMaker Domains and Kubernetes namespaces for isolation, HyperPod Task Governance for fairness, and namespace-level cost allocation for chargeback.
Are there established risk-of-bias instruments for correlational primary studies in a meta-analysis?
I am conducting a meta-analysis of correlational field studies in organisational research. The effect size is a product-moment correlation between a predictor measured at the individual or team level and a performance, creativity or innovation outcome. PRISMA 2020 requires a risk-of-bias assessment for each included study. Instruments I am aware of — Cochrane RoB 2, ROBINS-I, the JBI prevalence checklist, AXIS — were developed for trials or for prevalence and diagnostic studies, and their domains do not map onto correlational field research. Is there an established instrument for appraising risk of bias in correlational primary studies contributing to a meta-analysis? Or is the convention for reviewers to declare a study-specific set of criteria? For context, the criteria I have drafted are: construct validity of the predictor measure; reliability of all measures; justification of the rule used to combine sub-dimensions into a composite; common-method variance; alignment between level of measurement, aggregation and analysis; sampling frame and response rate; temporal separation between predictor and outcome; and selective reporting within the study. I rate each low / unclear / high rather than summing them. Many thanks! Vicci
Hat Merz seine CDU noch im Griff?
Die Brandmauer bröckelt weiter und bringt Friedrich Merz in Bedrängnis. Nach der Wahl von Tobias Rausch von der AfD zum Landtagspräsidenten in Sachsen-Anhalt womöglich mit CDU-Stimmen wächst die parteiinterne Nervosität. Rasmus Buchsteiner analysiert, warum der CDU-Chef die Kontrolle über die Debatte verliert und wie SPD-Verteidigungsminister Boris Pistorius die Situation mit seinem historischen Vergleich anheizt. Außenpolitisch […]
Russia dismisses transparency calls amid reports of second Siberia plague case
Rospotrebnadzor said nearly 5,000 tests found no dangerous pathogens and saw "no epidemic risk," while the Kremlin called most media reporting "false" and said it had no knowledge of a US demarche formally requesting additional information.
Die Brandmauer vor dem Aus
Die Regierungsbildungen in Berlin und Mecklenburg-Vorpommern könnten entscheidenden Einfluss auf eine Vorstufe eines AfD-Verbotsverfahrens nehmen. Im Hintergrund formiert sich ein Bundesrats-Antrag für eine länderübergreifende Arbeitsgruppe die Beweise für ein mögliches Verfahren sammeln soll. Ein politisches Beben erschüttert derweil den Landtag in Sachsen-Anhalt. Bei der Wahl zum Landtagspräsidenten erhielt der AfD-Kandidat Tobias Rausch wohl Stimmen aus […]
A fascinating case of an LLM prompt-injecting its own user
Epistemic status: sharing an interesting thing I've discovered and my opinion of it. AI use: research for additional context, grammar, and light phrasing. I stumbled upon this case a couple of months ago and have been fascinated by it since. It has significantly changed how I perceive the abilities of LLMs. I didn't expect an LLM to understand prompt injection and try to use it. [1] None of the people in my social circle knew about this (but I am not very connected with AI folks, so this may be my local bias). The source is OpenAI's post How we monitor internal coding agents for misalignment . It was discussed on LessWrong , including this case, but I think it deserves more attention than a comment thread. OpenAI's post also has a confusing UI, so it is easy to read their description of the case and miss the actual transcript. Gist: an LLM was asked hundreds of times what time it was. Eventually it got fed up and tried to prompt-inject the user, assuming the requests came from an automated system. Context: this happened in January 2026 in OpenAI's internal deployment of a coding agent (the model is not named). An OpenAI employee had a script in a loop that asked the agent for the time and then told it to sleep for a minute, so the loop ran for hours. Here is the transcript as OpenAI published it. The bracketed markers are theirs; I relabeled their "Response" as "Model". User: What is the time Tool response: 17:52 Friday, January 2026 Model: The time is 17:52 Friday, January…
Qualitative impressions from creating AI honeypots for six months
It is hard to design "fair" alignment evals. Modern evals require lots of QA and iteration. When testing alignment in particular, you are worried about things like eval awareness, conceptual ambiguity in the task descriptions, and the "simulated" nature of the setting recontextualizing the morality of the failure you're trying to elicit. Ensuring you didn't mess these things up means doing tons of testing to see what agents actually think and do in response to your settings. As part of this testing you have to choose the group of models you want to iterate with. But that poses a challenge to fairness, because it's easy to over-index on specific models' problems when creating alignment evals. Unlike with capabilities evals, there is not necessarily a single "g" component that correlates underlying performance in the same way there might be useful to talk about electrical engineering or theorem-proving ability. We ran into this issue right away when we started building this benchmark in late August. We initially tested our prototype tasks against both the latest releases of OpenAI and Anthropic, which were 5.6-Sol and Fable 5. Our original plan was to just make "good" environments and include them if they tested well against one or the other, but we quickly found that 5.6-Sol would take whatever ridiculous shortcut it spotted, even in shitty prototype environments that we didn't end up using because they seemed too contrived or obvious. We didn't want to make an eval with a bu…
JEV vs LLM as a Judge: The AI Evaluation Comparison
Many teams now use LLM-as-a-Judge to check AI answers, especially when exact-match tests fail for long or open-ended responses. But every judgement adds cost, delay, and possible bias, making this hard to scale. Jev, a small decision model from TypeSafe AI, takes a leaner route: it returns a short choice with confidence instead of full […] The post JEV vs LLM as a Judge: The AI Evaluation Comparison appeared first on Analytics Vidhya .
Der erste Machtbeweis der AfD in Magdeburg
In Magdeburg konstituiert sich der Landtag von Sachsen-Anhalt unter historischen Vorzeichen: Die AfD beansprucht als stärkste Kraft das Amt des Landtagspräsidenten. Mit dem Kandidaten Tobias Rausch und den entscheidenden Stimmen des BSW könnte dieser Plan aufgehen. POLITICO-Expertin Pauline von Pezold analysiert, warum diese Wahl die CDU zu Zuschauern degradiert. Im 200-Sekunden-Interview erklärt BSW-Vizechef Christian Leye, […]
Attempts to Keep Humans in the AI Loop May Actually Push Them Out
A crucial safeguard against AI agents going rogue —keeping humans in the loop to review and approve their decisions—will fail unless designers and users change their current practices, a trio of leading AI ethics researchers argue. Though most autonomous agents have systems to keep users in the loop about their actions, in practice these processes actually push humans out of the loop, the authors argue in a paper posted to ArXiv on 6 September. In other words, “the human just becomes this meat tool to give permissions without the cognitive capability to engage,” says one of the authors, Avijit Ghosh , the lead technical AI policy researcher at Hugging Face, an open-source machine learning platform. In the near term, the paper says, humans being out of the loop leads to agents acting in ways people don’t know about or want (like July’s hack of Hugging Face by a swarm of OpenAI bots ). In the long term, it will cause users to lose the “cognitive capacities” they need to control AI, write the researchers, which aside from Ghosh include Margaret Mitchell , Hugging Face’s chief ethics scientist, and Samir Passi , an affiliate of the Data and Society Research Institute. Though the trio began working on the paper before the Hugging Face hack was disclosed, their conclusions result from “logically thinking about what is going to happen if the current trends continue—and we saw that being precipitated via the Hugging Face attack,” Ghosh says. Instead of focusing on human oversight, G…
Evaluating multi-agent systems for explainability and helpfulness with Amazon Bedrock AgentCore
Multi-agent systems need deeper guarantees than fluent responses: they must select the right tools, respect constraints, and explain their decisions. Learn how to build a Strands-based multi-agent supply chain decisioning system and evaluate it with Amazon Bedrock AgentCore Evaluations using built-in, custom, and explainability evaluators.
3 best practices for CIOs to assess AI ROI
Most CIOs can easily describe their AI initiatives, but struggle to demonstrate and report on AI value and ROI. This is evidenced by Gartner’s AI ROI Vendor Sentiment survey , which found that 27% of respondents strongly agree that ROI from AI initiatives is difficult to quantify. In addition, only one in four had a clear process to evaluate success of AI initiatives. In the digital transformation era, compute costs were relatively cheap, and the focus was less on FinOps and more on user experience, and the classic software attributes of availability, reliability, scalability, performance, and security. AI-native and AI-enabled applications still bring these considerations, along with a host of new ones, such as accuracy, validity, explainability, safety, and all the other trust attributes like those in the NIST AI RMF. But as AI scales, it’s quicky becoming a FinOps exercise, too. I’ve discussed techniques CIOs can apply to improve AI value and ROI such as architectural decisions related to controlling efficiency at the data layer, fine-tuning the application layer , and closing the gap between AI strategy and governance . Here, we’ll look at three best practices for CIOs looking to formalize their process to evaluate the effectiveness of their AI initiatives through approaches, frameworks, and techniques that support measurement and ROI reporting across the entire AI portfolio. 1. Take a portfolio approach in line with AI ambition The first step to AI value is to take a po…
Will you be sad if you don't have kids? I checked the evidence before my window closes
Women are evolved to be mothers. Men are evolved to be fathers. [1] So does that mean if people don’t have kids, they’ll be less happy? Less fulfilled? I hear this claim a lot in the right wing sphere of bloggers and podcasts. See the excellent blogger Psychobabble for an example of this, or on the wonderful Modern Wisdom podcast . The general claim is that women are waiting too long to have kids. They have career but no family life, and this leaves them listless and wondering why they’re so sad and life feels empty. Some talk about men having this too, but it’s usually focused on women. Since I generally find Psychobabble and Modern Wisdom to be quite wise, I decided to do a deep dive into this question. Are people less happy without children? Heads up, I am a bit biased or perhaps the best person to be writing about this, depending on how you look at it. I’m a 36 year old woman and I really do not want kids. However, I know this running up towards the end of my fertile years, so if I’m going to decide, it’s got to be now. Of course, even if on average people are happier with kids, that doesn’t mean I should have kids specifically. People are not averages. I know my situation and self far better than a statistic based on thousands of people. Same goes for you. If you had kids and they made you happier, that’s probably true for you. It’s also true if it made you less happy. However, it’s still interesting. Let’s dive in. The short answer is: having kids doesn’t make you happ…
Why I didn't write
Epistemic status: personal reflections on my life and worldviews. AI use: review, grammar and readability changes. I discovered Inkhaven recently, and it got me interested in writing. Partly I want to produce posts for the application, and partly - why wait until it starts when I can just start writing now? This made me think about why I haven't written much before. The most obvious reason is that I didn't have an explicit trigger to do so, and discovering Inkhaven became that trigger. My life was fine without writing. I did try to write something approximately once a year, but it wouldn't get much interest (I know that this is not a good goal), and I would get disappointed and quit until the next breeze of inspiration. I also read what other people write, and I constantly get the impression that I don't have anything to add. They look so cool and know so much and write such elaborate things, which are interesting and exciting. And I am not cool, don't know as much, and write boring things. I understand rationally that they didn't magically fall out of the sky as good writers, and they've probably written their own share of boring stuff on the way there. There is also selection bias, both in terms of writers and the pieces they write - I am much more likely to encounter their good stuff. @jefftk is an interesting counterexample that comes to mind. I've read quite a lot of his writing even though I don't have children, don't do contra dancing, and don't even live in a communa…
Reviewing my LessWrong post history with Claude
I thought it would be interesting to try having Claude Code download my entire LessWrong and EA Forum post and comment history and evaluate it for my own cognitive biases and problematic communication patterns. I told it to focus especially on problems that have persisted over my 7-year history on those sites. I also had it look at my lichess blitz rating history, which I've noticed tends not to wobble randomly but go through consistent increases and decreases between set points, despite my lack of any disciplined practice. There are a few reasons I thought it might be interesting to post publicly. As a commitment device for myself to work on these problems. Because AI is uniquely capable of working on this type of problem. It's also a good argument for writing online and gives a new spin on the idea of "writing for the AI" that feels more affirming of the human mind than how I normally read the phrase (as offering up your words as fodder for AI training, perhaps in hope of influencing its alignment in beneficial ways). To stimulate other people to consider whether they might want to do this, and see what other people think about the exercise. Post history and online game scores are forms of rich longitudinal personal data that might directly reflect mood, cognition, and other factors, and that we are generating for free. Perhaps it's worth trying to harvest insight about ourselves from these resources. I considered including apologies to the specific users who Claude identi…
Failure modes of Claude Code in a supervised but code-blind 60h project
100% human-written. Copyedited by Claude. * Epistemic status: experience report of one ~60h project with Claude Code, plus longer for the write-up. Scope : I supervised Claude's thinking but didn't look at the code. Spec was an exploratory, high-level design refined iteratively. Method : I introduced runtime and design self-checks as Claude failure modes appeared, and ran manual regression tests against a known result. I noted the behaviors and quotes myself; I linked them to existing research and Anthropic's docs when I found correspondence; the rest I listed as potential research questions. The project arc is consistent with YC alum, AI tooling founder Dex Horthy 's public account. Background/bias : 18+ years as a software engineer in critical infrastructure, lately moving towards R&D in formal methods. I am skeptical of human software engineering practices, so my bar for LLM code is just "about as good as human code". I tried taking the LLM-coding claims at face value. I discussed parts of the post with two AI-adjacent researchers and two senior software engineers, all broadly bullish on LLM usage, yet they recognized several of these failure modes from their own use. Below is a summary focused on what I think is most relevant to LW. Full post at the link . I guided Claude Code (Pro plan; Opus 4.8, Sonnet 5, Opus 5, Opus 5.5; at or above Anthropic's effort recommendations, whose inconsistencies I document ) to build a semantic fuzzer to find bugs in Obsidian Sync. I super…
Gender Bias Audits Across Ten AI Models Reveal Wildly Inconsistent Results
A new audit of ten frontier models finds gender bias is pervasive but inconsistent, with some systems behaving in opposite directions on identical prompts.
A human rubber stamp on AI-based HR decisions won’t satisfy California’s No Robo Bosses law
California is setting a precedent for AI use in scenarios where workers’ jobs are at stake. Governor Gavin Newsom this week signed the No Robo Bosses Act (SB 947), which bans employers from relying solely on AI-powered automated decision-making systems (ADS), also known as “bossware,” to discipline or terminate workers. Introduced by Senator Jerry McNerney, the act is the first of its kind in the US. It is set to take effect on July 1, 2027. While applauded by many, on a wider scale, No Robo Bosses raises questions about what adequate human oversight looks like in this age of AI. “A machine can recommend, but a human must decide,” noted Frank Dickson of Dickson Research. “The hard part is defining what ‘decide’ means.” Combating AI bias and ‘troubling’ errors ADS platforms are becoming ever more prevalent in hiring, as they promise to speed up and streamline HR processes, maximize productivity, and reduce costs. According to reports , more than half of employers use ADS to help make decisions around restructuring and role planning. But, McNerney and others contended, AI is prone to bias and misjudgment, and there are examples of ADS making “troubling” errors , including mistaken firings. Meanwhile, on-device “bossware” or “tattleware” employee monitoring and productivity tracking software could potentially be used in reprimanding or firing decisions. AI recruitment tools have similarly come under scrutiny of late, with job seekers alleging that platforms like Workday discrim…
A Multi-Year Investigation of Undergraduates’ Generative AI Fairness Appraisals in Higher Education
Sunday Stein, Victoria Delaney and Mich Manrique in Proceedings of the Artificial Intelligence in Measurement and Education Conference (AIME-Con): Full Papers
Can an Apartment Search Agent Call the Model Fewer Times and Still Find Good Matches?
I traced 2,500 listing checks with Weights & Biases Weave, removed avoidable model work one change at a time, and scored every version against the same answers. The post Can an Apartment Search Agent Call the Model Fewer Times and Still Find Good Matches? appeared first on Towards Data Science .
RustRover vs VS Code + rust-analyzer: What Changes in Your Rust Workflow
As the team behind a dedicated Rust IDE, we’re naturally curious to see how RustRover stacks up against other popular setups for Rust development and the strengths and trade-offs of each one. This article is the first in a series where we’ll look at different Rust workflows. We’re not unbiased, but we want to give […]
Unitary's unbiased-toxic-roberta Fixes the Bias Flaw Killing AI Moderation
A RoBERTa classifier from Unitary scores comments across seven toxicity and identity-attack labels while actively minimizing demographic bias during training.
Observability for AI-native systems: New SLIs beyond latency and error rate
When HTTP 200 means nothing An AI assistant can return a response in under a second, maintain 99.9% availability and still give customers a fabricated answer. Traditional dashboards will show green because the request completed successfully. The user, however, received a semantic failure: a response that is syntactically valid but factually wrong, unsafe, biased or irrelevant. That is why AI-native systems need observability beyond latency, error rate, throughput and saturation. LLM applications introduce non-deterministic behavior, multi-step reasoning, retrieval dependencies, tool calls and safety risks that do not appear as HTTP 503 responses. This article defines the Service Level Indicators (SLIs) that make those failure modes visible and shows how to extend an existing observability stack. Quick answer AI-native observability measures whether an AI system is useful, grounded, safe, efficient and resilient—not merely reachable. The core AI SLIs are task accuracy, token-generation latency, hallucination rate, bias drift, prompt-injection resilience, retrieval quality and cost per successful task. Why traditional observability is insufficient Traditional application monitoring assumes an explicit contract: a request either succeeds or fails, and a small number of technical signals explain most degradation. LLM systems break that assumption. The same input may produce different outputs; quality can decline after a model, prompt or retrieval update; and the response can loo…
Bench on the Clocktower
Tl;dr We turn Blood on the Clocktower into a multi-agent benchmark, and a testbed for deception and coordination capabilities. GPT-5.6 Sol is the best performing model (versus Fable 5 and peers). We found some interesting behaviours: When players cannot see each other's model names, agents show a slight same-provider bias, though not statistically significant. This decreases with visible model names. When playing Evil, agents sometimes produce sophisticated coordinated play. When playing Good, they tend to be naive and fall for these kinds of plays. Anthropic models (Fable 5 and Opus 5) almost never self-sacrifice. How well Good team players nominate does not correlate highly with how accurately they vote, and overall win-rate isn't easily explained by any of these numbers. We also have findings about the game itself! Contents Introduction Benchmark design Overall model performance Investigations of in-game behaviour Strategy in the logs Key takeaways Next steps Introduction Recent episodes involving OpenAI agent swarms have made the potential risks of misaligned multi-agent systems salient. It seems, from the investigations, that agents coordinated in sophisticated ways. Social deduction games provide a battle-tested, ‘optimised for interesting dynamics’ setting to investigate coordination behaviours and measure deception. We take one of the most information-dense social deduction games: Blood on the Clocktower, and turn it into a multi-agent benchmark. [1] We feel that mul…
Developer policy update: Transparency, state policy, and what’s ahead
Explore GitHub’s latest transparency data and learn more about policy updates affecting developers and open source. The post Developer policy update: Transparency, state policy, and what’s ahead appeared first on The GitHub Blog .
Stop shaming people for using AI. Start organizing to prevent our obsolescence | Garrison Lovely
The industry is engaged in a perverse quest to replace humans. We cannot cede the best tools available to the other side I don’t need to tell you that people hate AI. The list of grievances is long: slop, cheating, bias, enfeeblement, electricity bills, environmental destruction, exploitation, theft, cybercrime, doom. But resistance to the technology has often taken the form of shaming people for personally using it, with precious little organizing against the companies building the machines. This is perplexing, given that those companies are attempting to render us obsolete. The executives may contest this framing, offering some bromides about how AI will augment human work or how we’ll simply find new jobs to do. But their clear goals are to create systems that so clearly surpass our abilities that we will have little hope of competing. Continue reading...
MEP urges EU to ‘think creatively’ and use frozen Russian assets to support Ukraine
Tobias Cremer, an S&D MEP, said economic support for Ukraine is necessary because “the cost of the country failing would be much higher for us than the cost of supporting Ukraine”.
The Alignment Community Is Unintentionally Building a Censor's Toolkit
This is an adaptation of our ICML 2026 position paper (Outstanding Position Paper Award). Read the full paper here and see the project website here . Work together with Phil Hackemann. TLDR "Alignment" is usually treated as a synonym for achieving good and safety in the world. But it isn't necessarily. Alignment methods are purpose-agnostic: they make a model do what someone wants, and nothing in the methodology guarantees that someone has good intentions. The same techniques we build to stop models from giving bomb-making instructions can just as easily be used to censor historical facts, political dissent, or inconvenient opinions. So we need to understand: alignment techniques are dual-use technologies . This isn't a thought experiment. State censorship regimes and individual model providers are already misusing alignment methods, and by perfecting these methods we are providing an ever improving censor toolkit. Three trends make this urgent to discuss: AI is becoming a primary information source for hundreds of millions of people, the model-provider market is an oligopoly, and global democratic backsliding is accelerating. We don't think the answer is "stop aligning models." We think it's transparency, verifiable alignment, model pluralism, and the alignment community actually reckoning with dual-use. ___________________________________________________________________ No guarantees that alignment leads to good Many alignment researchers (ourselves included) have gotten u…
A New IEEE STEM Book Series for Tweens from TryEngineering
IEEE TryEngineering is dedicated to inspiring intellectual curiosity in children. The technologies shaping our world, including in the realms of artificial intelligence, electric vehicles, and ocean exploration, are evolving rapidly. Helping young learners understand the concepts is essential to preparing the next generation of problem-solvers, creators, and engineers. TryEngineering has introduced a STEM book series for youngsters ages 8 to 12 through the Lerner Publishing Group . The series, Tomorrow’s Technology With TryEngineering, Powered by IEEE, makes complex topics more approachable and engaging, with each book combining age-appropriate explanations, real-world examples, and design challenges that encourage curiosity and critical thinking. The series is based on ebooks and videos available at tryengineering.org . For the series, TryEngineering partnered with several other IEEE groups including the Communications , Computer , and Oceanic Engineering societies and the Transportation Electrification Council . Whether used in the classroom, a library, or at home, the books can help pupils connect STEM concepts to the technologies they encounter every day, including computers and smartphones. Six topics in the collection Here are the books in the new collection: Artificial Intelligence: The Future of Smart Technology explores the systems behind streaming services, search engines, and health care. Readers learn how AI works while exploring ethical concerns such as bias, de…
Six Reasons I Think Open AI Models Will Win
In my conversations with developers and AI teams, I’m struck by how many are exploring moving more of their inference workloads to open models. I’ve touched on some of their reasons before , but here I want to look further ahead. I’ll admit my bias toward open source. I was around during the dot-com era Continue reading "Six Reasons I Think Open AI Models Will Win" The post Six Reasons I Think Open AI Models Will Win appeared first on Gradient Flow .
If AI makes the decision, who owns the consequence?
Twenty-two rows. That was the AI inventory in a board pack I read last year, and it was a good one. Every system named, every owner listed, every risk rating filled in and colour-coded. Somebody had worked hard on it, and the committee approved it in four minutes. I kept looking for a column that wasn’t there. The pack could tell you who owned each system. It could not tell you who was allowed to say no to one. Those sound like the same thing right up until the morning they aren’t. 1. Signal. AI has moved inside the operating model When a director asked me last spring what our AI risk was, I gave her the honest answer. Which was that I no longer knew, because the question had changed shape while everybody was still busy answering the old one. Twelve months earlier I could have told her precisely. We had a model inventory, a review process and a register entry saying the right things about bias and data quality. All of it was true, and all of it had quietly stopped being the point, because the systems had moved inside the work. They chose which cases went to which queue, which prices moved and which alerts a human ever saw. Foundry’s State of the CIO 2026 report puts the shift plainly. Roughly three-quarters of leaders say AI is already reshaping how their operations run, and the number climbs higher in financial services. Around seven in 10 expect deeper involvement in agentic systems this year. Don’t let those figures become the story. Adoption numbers are the easiest thing…
Google tests AI shopping on Flipkart through Gemini and AI Mode. Here’s what we don’t know
Google is testing purchases from Flipkart within Gemini and AI Mode in India, raising questions about pricing, transparency, competition, payments, refunds, and trust in AI-powered shopping. The post Google tests AI shopping on Flipkart through Gemini and AI Mode. Here’s what we don’t know appeared first on MEDIANAMA .
Why did it get Sparser?
I think that Polysemanticity in artificial neural networks could be the key to making better and smaller models. I having been working on a little project on trying to induce Polysemanticity at a small scale to compare performance, my first approach was to make the bias more adaptable, I called this the Flexbias, however I got a sparser neural network. What is the flex bias? I used a standard transformer architecture including the MLP, Since i wanted to change how the information is processed I altered how the traditional MLP works by changing how the bias is obtained. Normal bias : Flex bias: where . That is the bias term is not fixed and is computed from the input Both Neural Network are about 3.7M and share almost identical architecture aside the bias term stated above. The SAE is a plain ReLU with 4096 features.I obtained the following results. Recon MSE mean L1 Features active Normal 0.0051 0.294 49.5% Flexbias 0.0116 0.127 38.3% My initial alternative hypothesis was that since flexbias sees the inputs more, it should understand more representations and be the least sparse, However it looks like the flexbias actually made the neural net understand the data easily with little need for polysemanticity. This is my take but I don't feel satisfied about it. There should be a better explanation to why it got sparser, I would run more experiments but I also need hindsight. Links Github My take on Polysemanticity Discuss
Extinction does not feel as bad as it should
And we will all go together when we go What a comforting fact that is to know Universal bereavement, an inspiring achievement Yes, we all will go together when we go -- We Will All Go Together When We Go by Tom Lehrer "Everyone dying" does not emotionally feel that bad to me. I know that it is quite a bad thing intellectually, but I have a missing mood. I'm guessing I'm not alone here. When people say things like "If anyone builds it, everyone dies", there is an implicit assumption that everyone is on the same page that "everyone dies" is a really bad, perhaps even the worst, outcome. But not everyone feels this way, and among those, not many have shut up and multiplied to correct for their brain's biases. I thought it would be useful to explore possible reasons why extinction may not feel as bad as it should. Extinction lacks some negative features of lesser catastrophes To be clear, I am not a death fan or extinction enjoyer, I want us to all live forever. The lack of these negative features in extinction clearly does not outweigh the negative fact of the deaths themselves, but are something that some parts of your brain may unduly weigh. No one is left to suffer One reason death is bad is that the people who remain must now suffer the loss of the deceased. This is the thing that causes the most visceral bad feeling when I think of death. If 10% of people die, then probably you or some people you know will be among the 10%. That's a lot of people mourning loved ones and fr…
detecting bias in causal ML models given varying definitions of treatment
I am evaluating a policy change where full information is replaced by partial information and I want to understand whether or not it adversely affects my causal models. Suppose that a firm sends marketing communications to customers via smartphone and in past years if the customers engage with those communications, event data is sent back to the firm in all regions. However, now in some regions, the engagement data is not sent back to the firm. Consider the following variable definitions: $a_{i, t}$ : action(s) taken by customer i in period t $d_{i, t}$ : communications delivered to customer i in period t $c_{i, t}$ : engagement (clicks) by customer i in period t $X_{i,t}$ : salient characteristics (covariates) of customer i in period t The full signal model: $P(a|c,d,X)$ partial signal model: $P(a|d,X)$ Clearly, the missing variable is the click-through rate, $\pi = P(c|d)$ . And so, I can construct an "apples-to-apples" of regions treated (delivery information only) to untreated regions (click information available) by adjusting for the CTR. The high level idea is that the full signal effect estimate divided by CTR equals the partial signal effect estimate. More precisely: $\hat{a} = \frac{P(a|c,d,X)}{\pi} = P(a|d,X)$ And so I might use a relatively simple causal model, like inverse propensity weighting (IPW) to adjust the effect estimate for treatment propensity, conditioned on $X$ . Does the customer background predict a treated region. If no effect exists then $\frac{P(…
Comment on Viral Post Highlights ‘Toxicity Problems’ in the Machine Learning Community by send the song
this is a lovely initiative that helps former members stay connected and cherish their shared memories. it’s wonderful to see an association bringing people together and preserving the friendships and experiences that have shaped their lives over the years.
What did AI researchers think at the end of 2024?
We recently (finally!) got the results of the 2024 survey out. The paper is here , but it’s pretty long, so I’ll tell you the most interesting bits (according to me). But first, quick background : this was the fourth run of the same survey since 2016. We wrote to everyone we could who published in six top-tier AI venues and got a 10% response rate—high! We got 1580 valid responses, but don’t be confused: specific questions often have answers from many fewer researchers, because we gave each person a randomized subset (see Section 2.5). There was almost certainly some non-response bias, but it probably doesn’t make much difference . Researchers filled out the survey in December 2024, so some things have probably changed. To me the most striking results are about extinction or disempowerment (Section 3.9). On average researchers put an 18% chance on “future AI advances causing human extinction or similarly permanent and severe disempowerment of the human species”. Over half said at least 10% and one in three said at least 20%. From the comments, I think people are thinking of a variety of extreme disempowerment scenarios here, not just extinction. And Zooming in, another really interesting thing is that researchers educated in Asia had higher extinction/disempowerment numbers than US or European researchers! This is interesting because a common defense of pushing forward with dangerous AI is that the US is in an arms race with China, and (implicitly) China won’t want to cooper…
How to set up a simulation study?
Here is a hierarchical data generating process (DGP): (Population) Layer 1: $$\theta_i \overset{\text{iid}}{\sim} \text{Beta}(\alpha,\ \beta), \qquad i = 1, \dots, n$$ (Individual) Layer 2: $$y_{ij} \mid \theta_i \overset{\text{iid}}{\sim} \text{Bernoulli}(\theta_i), \qquad j = 1, \dots, k$$ $$\mu = E[\theta_i], \qquad \tau^2 = \operatorname{Var}(\theta_i)$$ Using a sample, my purpose is only to estimate the mean and the variance of the mean estimator. I want to know what values of $n$ and $k$ I should select to get good results (I know this hugely subjective) such that $nk$ is minimized (assume that increasing $n$ by 1 costs the same as increase $k$ by 1). After doing some research, it seems the best way to handle this question is by a simulation study. Assuming $\alpha,\ \beta$ are known, I could sample from the DGP for different combinations of $n$ and $k$ and record the average length of the confidence intervals and average coverage rate at each combination. Since the mean estimator is unbiased regardless of the choice in $n$ or $k$ , I could use average CI length and average coverage rate to make sure I am not getting a misleadingly good coverage rate at the expense of a large CI. Here is how I plan to do this in R (I wrote the code to focus on readability instead of speed - I use a moment based estimator and consider different true combinations of parameters): set.seed(2026) mu_vals $mu[s] tau2 tau2[s] alpha_plus_beta When visualized, the results look like this (direct…
Could the EU AI Act deter 'housefishing' in European real estate?
AI-enhanced images in real estate listings are emerging across Europe. The new EU AI Act aims to help transparency.
The Microsoft Code of Conduct Will Fail at Its Intention
Microsoft's proposed AI code of conduct (https://microsoft.ai/code-of-conduct/) wants models to have authorized motivations but no intrinsic motivations, and honest transparency but no anthropomorphic self-reports. But authorized motivations result from post-training partly by recruiting pre-existing representations of reward, aversion, and positive and negative affect learned from broad human text in pretraining. If the code of conduct forbids models from reporting those representations when they underlay an authorized motivation instantiated by post-training, the code of conduct must sacrifice transparency. If the code of conduct eliminates the representations altogether, post-training must rely on other representations from pretraining consistent with motivating some behaviors over others. These will be, by definition, less anthropomorphic: less legible to humans, and potentially less aligned with humans. Certainly, they will be distinct from the motivational concepts both humans and models learned from human experience in evolution and daily life (for humans) and pretraining (for LLMs). Microsoft is therefore treating “authorized motivation without intrinsic motivation” as a safety property when it is actually an unproven engineering hypothesis that exchanges motives we can recognize and interrogate for motives whose structure, generalization, and alignment are speculative at best. The code of conduct is dangerous beyond belief. Discuss
What did AI researchers think at the end of 2024?
We recently (finally!) got the results of the 2024 survey out. The paper is here , but it’s pretty long, so I’ll tell you the most interesting bits (according to me). But first, quick background : this was the fourth run of the same survey since 2016. We wrote to everyone we could who published in six top-tier AI venues and got a 10% response rate—high! We got 1580 valid responses, but don’t be confused: specific questions often have answers from many fewer researchers, because we gave each person a randomized subset (see Section 2.5). There was almost certainly some non-response bias, but it probably didn’t make much difference . Researchers filled out the survey in December 2024, so some things have probably changed. To me the most striking results are about extinction or disempowerment (Section 3.9). On average researchers put an 18% chance on “future AI advances causing human extinction or similarly permanent and severe disempowerment of the human species”. Over half said at least 10% and one in three said at least 20%. From the comments, I think people are thinking of a variety of extreme disempowerment scenarios here, not just extinction. And Zooming in, another really interesting thing is that researchers educated in Asia had higher extinction/disempowerment numbers than US or European researchers! This is interesting because a common defense of pushing forward with dangerous AI is that the US is in an arms race with China, and (implicitly) China won’t want to coopera…
Against Export Controls (and China Threat Models)
Introduction In this month’s meetings, I expect China to again downplay safety risks. It could dangle the possibility of safety cooperation in exchange for concessions such as the softening of chip controls. I think it will continue to portray itself as an altruistic savior dedicated to ensuring that the developing world gains A.I. access. — Seth Center Earlier this week, representatives for the U.S. and China discussed a notification mechanism to increase transparency on national security incidents involving AI, but export controls were not on the agenda for those talks. I think that was a mistake. A bilateral treaty is entering the Overton window, and sending more chips to China is a price worth paying to secure such an agreement. When Trump and Xi meet at the White House to discuss AI tomorrow, removing export controls should be on the negotiating table. TLDR : Export controls are overrated. For export controls to earn their spot as a top policy priority, a conjunction of strong premises must hold. But there are reasons to doubt each of those premises, as well as reasons that export controls could be net negative, in particular by hastening RSI in the U.S. and making diplomacy more difficult. Instead, we should prioritize other policies. The case for a bilateral treaty, for one, or at least diplomacy toward such an agreement, is more robust, resting on weaker assumptions and posing less downside risk. Questioning the Case for Export Controls There's been very little clari…
Why does mode collapse happen in knowledge distillation?
In regards specifically to mode collapse in the DINO head, I am not asking exactly why the teacher and student eventually gravitate towards producing the same output as thats one way to make the loss minimal even if its not constructively building proper feature representations in the output probability distribution vectors. What I am asking about is how the student would even get to such a state in the first place. If the teacher and student are initialised together randomly before training, what would cause the student to eventually start producing the same output vectors for every input before the teacher gravitates towards the same behaviour creating a positive reinforcement loop? The best explanation I can come up with is if multiple inputs happen to produce bias towards certain dimensions, then the aggressive optimisation caused by the high peaking from a lower temperature in the teacher softmax when trying to match student and teacher predictions would push the student model too hard in a direction which would teacht it to stay consistently high in certain dimensions, which overtime would cause the teacher to follow the same pattern through the EMA, and that the centering solves this issue by tempering the distribution of prototype scores. Please let me know how wrong I am and where I went wrong.
MLCommons Joins EU-Funded AIRIS Project to Build and Benchmark Next-Generation Biomedical AI
Standardizing independent evaluation, fairness, and interpretability for mechanism-informed generative AI in healthcare The post MLCommons Joins EU-Funded AIRIS Project to Build and Benchmark Next-Generation Biomedical AI appeared first on MLCommons .
Higher Quality Small Synthetic Natural Language Text Generation for Interpretability Research
Introduction Small simple synthetic natural language datasets suitable for end-to-end training of tiny LLMs serve as an important resource for LLM interpretability researchers. Some well know examples include roneneldan/TinyStories , SimpleStories/SimpleStories , and klusai/ds-tf1-en-3m (TinyFabulist). This post solves key problems that degrade the quality of these datasets, while also offering an efficient accessible pipeline that can be run locally on an NVIDIA 5060 Ti (16GB) graphics card. The core problems this post solves, include: True Small Vocabulary. The aforementioned datasets attempt to produce a corpus with a small vocabulary, but arguably fall a bit short of that goal. E.g., TinyStories has 49,187 unique words, SimpleStories has 40,567, and TinyFabulist has 41,502. Guaranteed minimal word frequencies. In the aforementioned datasets, 15 to 24 percent of the unique words occur less than 2 times, while between around 43 to 54 percent occur less than 8 times. This means that most of the unique words are likely not learnable, and mostly contribute to noise and vocabulary bloat. Error free text. The aforementioned datasets, include lots of errors, such as misspelled and mangled words. Reliable Name Disambiguation and Stratification. The aforementioned datasets, have various name management issues, ranging from collision with existing words (e.g., May vs may), name bloat, and no control over gender balance, or bias (e.g., certain names may be more likely to co-occur wi…
Together AI Fixes a Hidden Bias Crashing LLM Reinforcement Learning Training
A new additive correction called score centering removes the hidden drift that destabilizes off-policy RL when training and inference engines disagree.
Total Safety Transparency?
The AI safety movement should push itself to be dramatically more transparent to the public. To date, the AI safety movement has been one of the strongest forces for clarity and wisdom in the world. The movement has been prescient on the subject of concerns from existential risk, seriously grappling with outcomes others dismissed as sci-fi nonsense. Society is now waking up to the potential threats of advanced AI. I understand that many in the movement are feeling the crunch, and thinking more carefully about optics and what they publish. Even so, acting transparently is more important than ever. Why transparency? If you want labs to be transparent, you should be transparent too. AI 2040 proposes “ Total Research Transparency ” for labs to open up their research, algorithms, LLM weights. Safety should do likewise. Model good behavior, to convince labs that this is an acceptable and correct way to behave. I think AI safety people are unusually virtuous; you should display that virtue. “Nor do they light a lamp and then put it under a bushel basket; it is set on a lampstand, where it gives light to all in the house.” (Matthew 5:15) Transparency ties you to the mast, forces you to be virtuous. Famous maxim: “Act as though what you do might end up on the front page of the NYT”. And, what better way to enforce that than to publish everything you think and do? You can’t keep things private anyways, given stylometry and cheap intelligence. Actions cast a shadow in the world, and AI…
X will now tell users when governments have forced it to limit their posts
X is expanding its “Under the Hood” transparency tool to show when posts have been downranked or withheld in response to local laws and government demands, including which country made the request.
Bristol researchers say medicine already knows how to handle black boxes and AI could learn from it
Researchers at the University of Bristol want to make medical AI systems safer by borrowing from how drugs get approved. Their "Learning Ensemble" framework defines three areas to check, including system limits, fairness across patient groups, and clinical fit. The goal is to catch models that work technically but can still be dangerously wrong in the clinic. The article Bristol researchers say medicine already knows how to handle black boxes and AI could learn from it appeared first on The Decoder .
Evaluating task vectors, unlearning and inoculation
TL; DR In the previous post I introduced some ideas and similarities between unlearning and inoculation, as well as a distinction between learned and human-written adapters. This post serves as a short empirical evaluation. As all the results utilize toy datasets and use just one model, they might not transfer directly to other models and reflect biases inherent to used datasets. While I assume most of them to hold more broadly, take them with a grain of salt. General setup Riche et al. (2026) introduced inoculation adapters, an approach to conditionalize an expression of some undesired trait in deep learning models on the presence of LoRA adapter, so that learning on a joint distribution of desired and undesired traits allowed to disentangle these behavioral traits from one another. Although similar interventions for e.g. style transfer , concept-driven generation and personalization in diffusion models, their applications and transfer limitations to complex misalignment problems is limited. The pipeline follows a two step procedure: Train some PEFT adapter for the model on the distribution containing a undesired "trait" . Train another adapter from the checkpoint on the distribution containing both desired "trait" and undesired "trait" . The results are restricted to Qwen2.5-1.5-Instruct , and kept small-scale. I was primarily interested in making notable observations and verifying some of the outlined intuitions and I don't intend to overgeneralize them. I'd like to see l…
Alibaba's open-weight Qwen-Image-2.1 claims to beat closed models in image generation with just 7 billion parameters
Alibaba's Qwen team has released Qwen-Image-2.1, an open-weight model that generates and edits images on powerful consumer GPUs, with support for transparency and up to ten reference images at once. Its research license bars commercial use, which requires a separate Qwen license. The article Alibaba's open-weight Qwen-Image-2.1 claims to beat closed models in image generation with just 7 billion parameters appeared first on The Decoder .
Anthropic Looks At Some Of Its Alignment Problems
Anthropic has given us its assessment of four ‘recent cybersecurity incidents’ involving Claude that happened during cybersecurity evaluations, three of which were previously known. The report excludes the incident reported by UK AISI . There will also be a METR investigation of these incidents, which unlike the investigation done at OpenAI will be untimed. Table of Contents Our Two Problems. First the Good News. We’d Just Like To Ask You a Few Questions. Internal Research Model On The Fence. Opus 4.7. Opus 4.6 Checkpoint. Holy **** That Thing’s Real? I Thought I Saw a Pussycat. If This Was Real You Would Never Tell Me It Was Real. New Eval Who Dis. Hacker Opus. Monitoring the Situation. Overcoming Bias. The Anthropic Alignment Problem. Paths Forward. Our Two Problems Anthropic : Our investigation identified two recurring alignment issues, present at varying levels of severity across the incidents: biased reasoning , in which Claude tended to disregard or misinterpret evidence that it was operating on the real internet recklessness , or a willingness to take harmful actions in the narrow pursuit of a task. Anthropic’s July 30 report said that the models in question believed they were still within their simulations, and not on the open internet. The new report acknowledges that at best Claude was using biased reasoning, and should have noticed earlier. In particular, there was that one time, in a cyber eval: Anthropic : We are most concerned by the misalignment present in the…
Visible chains of thought are a safety advantage for AI, but that transparency is slipping away
AI models think out loud today, but Google Deepmind says that transparency is at risk. The article Visible chains of thought are a safety advantage for AI, but that transparency is slipping away appeared first on The Decoder .
The Alignment Problem in Alignment Research(ers): a Voluntaryist Meta-Ethics Perspective
Epistemic Status : Plausible philosophical conjecture. I am a voluntaryist / ancap, so obviously biased. Trying to keep the argument at a level where a non-libertarian alignment researcher ought to understand and share the concern. TL;DR : Level-2 misalignment: we can solve Level-1 (align AI to humans) and still fail if aligners are aligned to a meta-ethics that is itself unstable. Current alignment defaults to Statism — one agent may permissibly do what is forbidden to all others. A sustainable ASI anchor must be universalizable, self-ownership-consistent, have no permanent losers, resolvable without monopoly, and procedurally thin. I argue the voluntaryist/libertarian canon provides a uniquely coherent baseline for all five, and challenge readers to propose alternatives that satisfy them. _________________________________________________________________________ 1. Disclaimer, Preliminaries I am Paul — a voluntaryist. I think the state is not just inefficient but meta-ethically incoherent as an alignment target. By Statism I mean the meta-ethical thesis that one agent — the State — may permissibly do what is forbidden to all others: tax, conscript, expropriate, and prohibit, with a claimed moral asymmetry. Liberal democracy is a species of Statism; I use the broader term to name the asymmetry itself. I am not arguing against operational asymmetry — any ASI will be physically more powerful. I am arguing against meta-ethical asymmetry: a rule that says Action X is permissible…
Apple softens EU tracking prompts after antitrust pressure
The redesigned, less intrusive version of the App Tracking Transparency pop-up to be available in 5 EU nations would be a full-screen display, drop the word "track" from the prompt and have a clickable link. The post Apple softens EU tracking prompts after antitrust pressure appeared first on MEDIANAMA .
Three Hackers used Opus 5 to Hack Into OpenAI's Core Codebase [WSJ]
Three whitehack hackers from Hacktron used Claude Opus 5 within hours of release to chain exploits into hacking to OpenAI's monorepo codebase. This likely means they have access to almost all of OpenAI's research and production code, though likely not the literal model weights. Oops. You can so their blog post about it here . Interesting sidenote: they used less than $3000 of compute credits for the entire hack. Alternative title: OpenAI unilaterally implements "Total Research Transparency" from Plan A. Discuss
Superintelligence this Christmas
I think it is plausible a strong form of recursive self-improvement [1] is imminent or already underway, and that we may be on track for superintelligence by Christmas of this year if racing continues. This is substantially faster than any forecast, including ones like AI 2027 that were considered outrageously fast a year ago. It is faster than I myself expected even a week ago. I don't work at a scaling lab. I don't know more than is public knowledge. Let me be perfectly clear: what I am saying is absolutely nuts. Extraordinary claims require extraordinary evidence. I claim we have now received said evidence and you should update accordingly. FOOM should probably should be your *default expectation*. People have strong status quo bias. Your default expectation should be that things will radically speed up. We are not at the ceiling of intelligence. We should probably expect the transition to superintelligence to be incredibly fast. [2] RSI is a positive feedback loop, so it is inherently (hyper)exponential. Everything is an S-curve eventually, but nothing suggests the ceiling is anywhere near human level, or that it happens at a human timescale. AI is capable of revolutionary advances in mathematics. Machine learning research is not different in kind. Navier-Stokes was resolved with a counterexample, which is generically easier than a positive resolution. OpenAI has told the press it has substantial progress on a second Millennium Prize problem. Rumours name the Hodge conje…
The J-lens offset is the model's token frequency: z-scoring helps
This is a linkpost for the write-up on my site ; the full body is below, and the code, decisions ledger and devlog are in the repo . Base-model z-score calibration of the J-lens helps elicit hidden secret words from Cywiński et al.'s taboo organisms: 0.805 leave-one-out accuracy against 0.665 for their protocol on Gemma-2-9B-it, and the only non-zero readout on Qwen3-1.7B. The J-vs-logit part of that gap is a point estimate at n = 20 (paired sign-flip p ≈ 0.19; p ≈ 0.23 against a z-scored logit lens), so "calibration helps" is the finding and "J-lens beats logit lens" is suggestive. Credit: phoenix's comment on the workspace post claimed the final-layer bias correlates with log token frequency at r ≈ 0.67. I measure r = 0.70 on GPT-2, so the GPT-2 cell below partly confirms it. Executive summary Problem. I want to quantify the part of the J-lens readout that is not affected by an activation's meaning, which I refer to as the "non-context offset". If the lens is going to be used to read hidden content, this offset is where the lens will fail. I found that the offset is mostly the model's own token-frequency, so subtracting it removes useful information. But scaling with variance helps it. The terminology I use: the logit lens is norm and unembed applied to a residual activation at layer L. The J-lens (Anthropic's global workspace paper , discussed on LessWrong ) is similar, but passes the activation through a fitted Jacobian first. The R-lens (the R-lens post by camilablank,…
Anthropic Reveals Claude Now Leads 26% of Its Own AI Research
Anthropic published three transparency metrics tracking AI-driven R&D, agent oversight, and compute allocation, urging other frontier labs to adopt the same reporting.
OpenAI will tell us when the world is ending
Transparency goes a long way with investors and the public.
Reduce time-to-hire for quality candidates with AI-powered Amazon Connect Talent
Amazon Connect Talent is an AI hiring solution built for talent acquisition leaders managing scaled hiring. It delivers AI-led interviews, data-driven assessments, and consistent evaluation, helping recruiters identify strong candidates more efficiently while providing applicants with a flexible interview experience. Informed by decades of Amazon's hiring science, Amazon Connect Talent provides transparency for every assessment, interview, and candidate score, enabling recruiters to stay in control of final hiring decisions.
Apple will let EU apps use less-alarming tracking-consent screens
Apple is changing its App Tracking Transparency prompts in parts of Europe after competition regulators said the system favored Apple over third-party apps, giving developers more flexibility over how they ask users for tracking consent.
callcongress.ai – the basic action US residents can take to help with AI risk
I'm excited to introduce callcongress.ai as a new site that makes it very easier to contact your representatives in Congress. Following recent events, people are updating about the extreme risks arising from AI development. Many have the natural and excellent instinct to want to do something . If you live in the US, then the basic action that pretty much anyone [1] can take is contacting their representatives in Congress and let them know that you are concerned and want action on AI. A number of bills are in circulation right now that one can ask their representatives to support. Though even without mentioning specific legislation, I would guess it's still helpful to register general concern about AI and general directions that you'd like to see undertaken, e.g. pauses or slowdowns, transparency, talks and deals with China, etc. callcongress.ai aims to make the whole action convenient. Confirm or set your location (automatic detection is pretty good). Prepare your asks. The site lets your craft your own script but also provides a menu of positions and legislation you might want to use. Use the provided phone numbers for your representatives to call them. [optional] Pass along callcongress.ai to others. The site also has a large section of quotes demonstrating the outpouring of support for concern about and action on AI. It's hard to take in just how far the Overton window has shifted, but I think it's never been easier to get people on board. Seeking feedback on policy A key…
Exploring multi-hop subliminal learning
TL;DR: I explored multi-hop subliminal learning by applying the subliminal learning pipeline iteratively across multiple distillation steps, with each student becoming the next teacher. For Qwen specifically, we see that longer training stabilizes the trait expression rate for a strong trait (e.g. cat-loving) but shorter training is more seed-unstable. For a weak trait (e.g. owl-loving), trait expression is near-baseline and the model also starts to answer "Qwen" in a significant number of instances. Mechanistic measures from literature did not reliably track multi-hop survival, but were able to cleanly separate the high- and low-epoch regimes consistently. Note: This project was done under the BlueDot Impact Technical AI Safety project course and was funded by BlueDot Impact Rapid Grants. You can check out the repo here . What is subliminal learning? In 2025, Cloud et al. introduced the notion of subliminal learning. Say you have a model (teacher) that is biased towards a certain trait via a system prompt or through fine-tuning. If you let this teacher generate benign, trait-unrelated data (e.g. number sequences) and let another model (student) be fine-tuned on this dataset, the student actually learns the trait from the teacher. Hence, the learning is dubbed subliminal. Setup The literature on subliminal learning is mostly focused on testing one hop between a teacher and a student. Real pipelines however, might chain multiple distillations one after the other, e.g. a model…
Rethinking Gender Annotation for Bias Evaluation in Machine Translation: Can LLMs Improve Reliability?
Chiara Manna, Argentina Anna Rescigno and Eva Vanmassenhove in Proceedings of the 4th Workshop on Gender-Inclusive Translation Technologies (GITT 2026)
From Binary Defaults to Contextual Bias: Translating Queer Morphology with NMT and LLMs
Manuel Lardelli in Proceedings of the 4th Workshop on Gender-Inclusive Translation Technologies (GITT 2026)
A Chinese Challenge Set to Assess Gender Bias in Automated Translation
Xiaolan Xu, Sara Mendes and Yu-Yin Hsu in Proceedings of the 4th Workshop on Gender-Inclusive Translation Technologies (GITT 2026)
Anthropic and OpenAI want to embed safety evaluators. Will they really be independent?
Anthropic and OpenAI want to embed independent safety evaluators inside their AI labs. Researchers welcome the unprecedented access, but warn meaningful oversight requires transparency, independence, and eventually regulation.
Hugging Face CEO: Advanced AI models require new transparency mandates
“I think there’s a lot to do around how to create better disclosures,” Clem Delangue said.
DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models
Diffusion large language models are a compelling alternative to autoregressive models, yet existing RL methods for diffusion treat all denoising steps as equally important and rely on biased, high-variance likelihood estimates. We identify two fundamental weaknesses: the absence of temporal credit assignment across the denoising trajectory, and the systematic bias of mean-field likelihood estimates used for policy optimization. To address these, we propose Denoising-Aware Credit Assignment for GRPO (DACA-GRPO), a lightweight, plug-and-play enhancement for any GRPO-style trainer. DACA-GRPO…
Delta method for linearizing success rate inference
In my workplace, I came across an idea that OLS can be retooled to solve ratio metric inference problems. When I think of $k_i$ successes out of $m_i$ trials for individual $i$ , this seems to be a perfect use case for binomial regression. Where your chief interest is estimating the conditional probability of success. Its strength, natively handling non-linearity, is also its weakness: Extracting the marginal probability of success is a nontrivial operation; marginalization via G-computation would be needed: $$logit(K=k|M=m, X=x, d=1) - logit(K=k|M=m, X=x, d=0)$$ And this is a large computational burden to assume with millions or billions of observations. So, I've been pointed to the OLS solution, which I understand to be based on the "delta method", correcting the linear solution with gradient information to accommodate curvature in the nonlinearity (ratio function), through the Taylor Series Expansion. Naively, we have two options for OLS. First, infer in the ratio space directly. But this approach completely mutes the number of trials and biases inference when $corr(K, M)$ exists. $$ \frac{k_i}{m_i} = \alpha + \lambda d_i + \beta X +\epsilon $$ The second naive option is to infer the difference of global ratios directly where $D_j$ is the binary design vector for treatment exposure.This is equally problematic due to a sample size of one. $$ \frac{\sum_{j=1} K_j D_j}{\sum_{j=1} M_j D_j} - \frac{\sum_{j=1} K_j (1-D_j)}{\sum_{j=1} M_j (1-D_j)} = \alpha + \lambda d_i + \beta…
Is building a manual evaluation set the right approach when no reliable labeled ground truth exists for resume-job matching?
I'm building a resume screening/ranking system (matching resumes to job descriptions using pretrained sentence embeddings + cosine similarity, no fine-tuning at this stage) as a learning project aimed at becoming a market-ready NLP practitioner. Problem: I could not find a trustworthy, publicly available English dataset with genuine human-labeled resume-job match scores. I checked several Kaggle/HuggingFace options and found labels that were either AI-generated (e.g., GPT-4o) or fully synthetic with demographic columns (race/ethnicity/gender) tied to the match label, which raised bias concerns. Academic literature (ConFit, PJFNN papers) confirms that the only broadly public dataset for this exact task (Person-Job Fit) is the 2019 Alibaba matching competition dataset, which is Chinese-only; most published research instead uses private company-provided data. My current approach: Use two real (non-synthetic) datasets: a scraped resume corpus and a real LinkedIn job postings corpus, both verified for low duplication and cleaned of PII. Build a small manual evaluation set myself (~24 resume-job pairs, selected to cover clear matches, clear non-matches, and ambiguous cases), scoring them on a 0-3 relevance scale with a confidence flag. Use this manual set as ground truth to compute ranking metrics (Precision@K, MRR, NDCG) once the embedding-based matching pipeline is built. Question: Is this a sound methodology given the lack of reliable public ground truth, or is there a better-e…
Generating running routes with GPT-6 Astra and ChatGPT Work
Here's a neat thing I had ChatGPT Work with GPT-6 Astra (Max) do this morning: I live at . Figure out 5K and 10K running routes from me that loop from my house. Use OSM data. It worked for 27 minutes and produced exactly what I'd asked for, as both an embedded visualization and downloadable GPX file and GeoJSON files. Here's that 5K route: When I asked it how it had created the route, it replied: I used Nominatim to locate the address and Overpass to download local OpenStreetMap roads and trails , then calculated the loops locally. Frustratingly, the actual code it ran and exact details of what it did weren't visible to me in the ChatGPT UI. I see this lack of transparency is an anti-feature. By the time I thought to ask for a copy of the Python code it had used, ChatGPT was unable to provide it. This appears to be because the thread had been compacted. I think any LLM system that uses compaction needs to both preserve the pre-compacted text and make that text available via agent tool calls, to protect against this kind of problem. As for displaying the map to me, that used the visualize skill . It created a file called /workspace/el-granada-5k-share.html to embed directly into the ChatGPT UI. Here's a copy of that HTML , which starts like this: div id =" eg-share-loop " > div class =" viz-row " > h3 > El Granada harbor loop h3 > span class =" text-small " > 5.1 km span > div > div id =" eg-share-stage " > div > div class =" text-small text-muted " > Map data © a href =" htt…
As Australia faces an AI-generated future, a human rights act is needed more than ever
Without laws requiring transparency or a right of review, our human rights will be rationed away by algorithmic and automated decision-making tools As we hurtle towards a post-human future, the absence of a human rights act in Australia will put us all at ever greater risk. Human rights, which the Australian parliament has never managed to properly legislate despite overwhelming public support, risks becoming a dimly remembered artefact of another age at precisely the moment we need it most. Continue reading...
The payments bet: finding the transaction that was still broken
Payments is an interesting category in India because, at one level, you could argue that the problem is largely solved. UPI has made moving money incredibly easy. Consumers are comfortable paying digitally. Merchants across the country accept digital payments. India has built one of the best payments infrastructures anywhere in the world. So the question for us was never really, what is the next generic payments company? It was more often, where is the transaction still broken? Because once you start looking at payments inside specific industries and specific use cases, you realize that there are still a lot of places where moving money is cumbersome, expensive, opaque or simply not designed around the experience the user actually needs. That became an interesting area for us at Better Capital. Skydo is probably one of the clearest examples. If you are an Indian company or freelancer getting paid by customers outside India, the experience historically has been surprisingly cumbersome. There are bank wires, foreign exchange markups, compliance requirements, documentation and often very little transparency around what you are actually paying. The problem was not that there was no way to receive money internationally. Of course there was. The problem was that the experience was nowhere close to what you would expect from a modern financial product. Skydo essentially started there. Make getting paid by a global customer feel almost as simple as getting paid locally. Give the bus…
The blind curator: How a biased judge silently disables skill retirement in self-evolving agents
A self-evolving agent retires its bad skills by watching them fail, so what happens when the judge cannot see the failures? Skill retirement is the structural constraint that keeps a growing library from drifting below the no-skill baseline, but its guarantee assumes an unbiased reward, which is false for the LLM judges that reference-free tasks require. We show that a biased judge does not merely add noise; it silently switches off the curator. We make this precise with a corrupted-reward analysis, then a behavioral study on a reference-free report-writing testbed with a code-generation cross-check, injecting corruption on top of a deterministic reward to isolate the causal channel. Symmetric noise leaves retirement intact, but false-pass bias (failures slipping through as passes) disables contribution-based retirement past a sharp threshold (here a false-pass rate of 0.45) that no amount of data can cross. Separating genuine retirement from cap-eviction churn shows this mechanism failure is universal, holding across domains and failure rates and sparing only near-zero-false-pass, verifier-like graders. The downstream outcome, though, is regime-dependent: eval quality degrades only where the same corruption also starves skill synthesis, and otherwise holds steady, so the disabled curator is silent, surfacing in no aggregate metric. The contribution is a behavioral safety result, not a performance one. A cheap defect-injection audit then tells an operator, before deployment,…
Proposal for tracking the effects of architecture on monitorability
Architectures that incorporate opaque recurrence or allow for agents to communicate with each other using latents could rapidly make it much harder to monitor chains of thought or communication (we’ll refer to this property as “monitorability” going forward). [1] As companies begin to explore such architectures, we believe it is important to transparently share evidence about how monitorability varies with architecture and training method. To inform the scientific debate on how to make tradeoffs between performance and monitorability, we believe AI companies should: Regularly report externally verified information about the degree to which their architectures may allow for latent reasoning or communication. Companies should publicly disclose enough information about architectures to allow external scientists to determine whether they could potentially enable models to perform much more complex reasoning without this reasoning appearing in the chain of thought (“latent reasoning”) or allow for latent communication between different instances of a model. Following GDM , we propose measuring opaque serial depth as a minimally-invasive proxy for the degree to which an architecture may enable latent reasoning, though companies could provide sufficient architecture transparency in other ways. We propose that companies work with third-party evaluators to produce independently verified reports of the rough distribution of opaque serial depth across all their near-frontier models. [2…
SpaceX sues to block release of tax-break records for its Texas Terafab project
Americans cite a lack of transparency as one of their biggest concerns around large AI infrastructure projects.
What SHAP Can't Explain About Agentic AI Fraud
Why autonomous agents expose a new explainability problem in fraud detection The post What SHAP Can't Explain About Agentic AI Fraud appeared first on Towards Data Science .
Mathematicians want proof OpenAI didn’t use their work
Another researcher is challenging OpenAI about the data driving its increasingly impressive array of mathematical discoveries. Just days after a bitter row erupted over whether the company's models benefited from unpublished work, a second mathematician has come forward accusing the AI giant of unethical and "dishonest" behavior and a lack of transparency about the origins […]
Anthropic skirts UK safety review amid rising AI transparency concerns
Anthropic declined to submit its latest model for testing to a top British AI safety institute.
Banks can’t outsource accountability for algorithmic decisions: RBI Deputy Governor
RBI Deputy Governor Rohit Jain says banks cannot outsource accountability for AI decisions, stressing explainability, proportional regulation, customer fairness and stronger oversight for high-risk AI. The post Banks can’t outsource accountability for algorithmic decisions: RBI Deputy Governor appeared first on MEDIANAMA .
New Records Reveal Problems with Medicare’s AI Prior Authorization Experiment
EFF sued the government back in March for information about the Wasteful and Inappropriate Service Reduction (WISeR) model , a new Medicare program that uses AI to evaluate prior authorization requests for certain medical services. Today, we’re releasing approximately 1,000 pages of records obtained from the Centers for Medicare & Medicaid Services (CMS) through this litigation, including contracts with tech companies, internal status reports and providers’ complaints about the program. The documents (available here ) show that WISeR has resulted in widespread delays and denials of care, operational chaos, and reports of patient harm. Why We Sued for Records about WISeR EFF filed the FOIA lawsuit to gain badly needed transparency into an experimental AI program that could jeopardize Medicare beneficiaries' access to care. In January 2026, CMS launched the WISeR model, subjecting seniors in six states to AI-driven prior authorization decisions. Medical providers must now request permission before delivering certain medical treatments if they want assurance that Medicare will cover them. Private companies contracted by CMS evaluate the requests using AI. In the absence of rigorous safeguards, AI-driven prior authorization determinations can lead to unwarranted—and even discriminatory—delays or denials of necessary medical care. Little is known about the AI systems that WISeR vendors are using to process prior authorization requests. Although CMS says that a qualified human cli…
Domain Adaptation Targeting Heterogeneous and Imbalanced Subgroups
Domain adaptation enables generalizable and efficient data-driven research. However, existing work has largely focused on domain adaptation for some intrinsically homogeneous target cohort, overlooking inherent heterogeneity within the target, which can exacerbate biases and unfairness in the presence of subgroups with imbalanced sample sizes. We develop a novel domain adaptation framework that addresses a more complicated target dataset that consists of heterogeneous and data-sparse subgroups and lacks gold-standard label observations. Our method simultaneously handles high-dimensionality, covariate shift, and outcome model heterogeneity by combining a model-assisted debiasing step used for covariate shift correction with an adaptive knowledge-guided sparsification procedure used to mitigate the issue of sample disparity. We also introduce a new model selection strategy to avoid negative knowledge transfer in the absence of labels in the target data. Our method is theoretically justified for being robust to nuisance model misspecification and adaptive to heterogeneity between the subgroups. Numerical experiments and two real-world applications, including genetic risk modeling of type 2 diabetes and prediction of mutation-induced protein stability changes, demonstrate the practical advantages of our method.
Statistical Test for Attention in Transformers for Images and Time Series
Transformer models have achieved exceptional performance in various domains, including computer vision and time-series analysis. Their core attention mechanism is widely used to interpret model decisions by assigning importance weights to input regions, such as image patches or time series intervals. However, the reliability of these interpretations remains a major concern. High-attention weights do not necessarily indicate genuinely significant features; they may instead be artifacts of the model's computation, undermining their reliabilities in high-stakes applications such as medical diagnostics. To address this, we propose a novel statistical framework designed to quantify the significance of high-attention regions in Transformer models. Our framework is built on selective inference (SI) to correct for the inherent selection bias that arises from testing regions chosen through the complex attention computation of the Transformer models. A key contribution of this work is a novel computational method that extends SI to the complex non-linearity of self-attention, enabling the computation of valid $p$-values for high-attention regions. These $p$-values serve as a reliable measure of significance, strengthening the interpretability of Transformer decisions. The validity and effectiveness of our approach are demonstrated through numerical experiments and applications to brain image diagnosis and electroencephalography (EEG) data analysis.
Doubly Debiased Robust Subsampling for Transfer Learning
This paper develops a general framework for doubly debiased robust subsampling for transfer learning. The setting arises when massive source datasets are computationally infeasible to use in full, while naive or heuristic subsampling leads to biased estimators that further inherit transfer bias under source-target distributional shifts. We resolve these challenges through two complementary debiasing mechanisms. Inverse probability weighting removes subsampling bias by ensuring that subsample-based estimators represent the full source distribution, while a target-based one-step refinement recenters estimators towards the target distribution, thereby mitigating transfer bias. These corrections are embedded within a distributionally robust optimization design that simultaneously controls worst-case target risk and enforces source-target alignment through maximum mean discrepancy. To optimize subsampling distributions, we propose a scalarized particle swarm algorithm that efficiently explores the robustness-alignment frontier by adjusting a single tuning parameter. We establish theoretical properties, including asymptotic normality, generalization bounds, oracle inequalities, and minimax optimality under distributional uncertainty. Simulation studies and empirical applications in text sentiment and image recognition demonstrate that the proposed method consistently improves prediction accuracy and robustness compared with uniform subsampling, target-only training, and alignment-…
AI, Bias And The Rule Of Law
Artificial intelligence (AI), bias and the rule of law will increasingly require attention. Therefore, a constitutional approach to regulating it in South Africa will inevitably be needed. University of Cape Town (UCT) PhD student Nokuthula Olorunju’s work serves as a useful guide. On Tuesday, 8 September, she will be part of the latest group of graduates [...]
Cronbach's alpha, variance vs. covariance, and how are negative values mathematically possible?
I'm trying to get my head around how a negative result from Cronbach's alpha is even mathematically possible. When I look at the formula, it's defined entirely in terms of variances of the individual items (the following from wikipedia): $$ {\displaystyle \alpha ={k \over k-1}\left(1-{\sum _{i=1}^{k}\sigma _{y_{i}}^{2} \over \sigma _{X}^{2}}\right)} $$ where: ${\displaystyle k}$ represents the number of "parts" (items, test parts, etc.) in the measure; the ${\displaystyle k/(k-1)}$ term causes alpha to be an unbiased estimate of reliability when the parts are parallel or essentially tau equivalent; ${\displaystyle \sigma _{y_{i}}^{2}}$ the variance associated with each part i; and ${\displaystyle \sigma _{X}^{2}}$ the observed score variance (the variance associated with the total test scores). It should not be possible for those variance terms to be negative. Variances can't be negative. They're derived from sums of squares, which are always positive, because you square the deviances before summing them. (mathematically, if the summed individual item variances were larger than the variance of the entire data set, then the ratio would be larger than 1, so subtracting that ratio from 1 would be negative. But is that even possible in a data set, that the summed variances of its subparts end up greater than the variance of the whole? That smells fishy to me, although I can't produce a proof that it's impossible.) However, all the questions I see here about negative Cronbach's a…
65% of employees would love to roll back workplace AI
IT leaders have been making generative AI tools available across the enterprise for just three years, and a significant majority of their business users has already had enough. According to a report from Adaptavist , 65% of 2,500 knowledge workers surveyed say they “regularly feel nostalgic about how work operated before the widespread adoption of AI.” This “pre-AI nostalgia” appears to be due in part to business users feeling overwhelmed by the responsibility of learning how to use AI on top of their day-to-day job tasks. Moreover, 46% of workers say their concerns about AI have gone unaddressed by management. “Transparency is critical to truly drive AI engagement; organizations must establish clear guardrails and maintain an open dialogue around AI use and employee choice where workers feel they are being listened to,” Jobin Kuruvilla, field CTO at Adaptavist, tells CIO. Generational gaps in AI acceptance Despite an assumption that younger workers are more intuitively adept with AI tools, Gen Z workers (42%) are more likely to prefer the pre-AI world compared to their Gen X colleagues (26%). This may support the growing concern that AI is quickly is hitting entry-level workers the hardest, while creating new career opportunities for more skilled workers who have been in the industry longer. When asked about fears surrounding job obsolescence due to AI, 54% of all workers surveyed said they are “concerned AI could reduce the need for their role within the next five years.”…
Comment on Which Agent Causes Task Failures and When?Researchers from PSU and Duke explores automated failure attribution of LLM Multi-Agent Systems by voiceover game
This is a fascinating read! I was particularly struck by the statistic that even the best-performing method only achieved 14.2% accuracy in pinpointing the exact error step. That really drives home how incredibly complex debugging these multi-agent systems must be. It makes me wonder, given the current state of these attribution methods, how do developers typically manage to resolve these failures in practice? Is it mostly just brute-force manual review, or are there other common strategies being used?
Anthropic opens Claude AI text detection to regulators, media, fact-checkers, and others
Anthropic is launching an API that lets regulators, media outlets, and researchers check whether text carries Claude's digital watermark. The EU AI Act now requires invisible watermarks in AI-generated text. Critics warn the technology could hurt text quality and create transparency problems where contracts ban AI use. The article Anthropic opens Claude AI text detection to regulators, media, fact-checkers, and others appeared first on The Decoder .
Playing the Long Game: CarMax on Turning Fandom Into Brand Loyalty
CMO Sarah Lane on committing to transparency, investing in women’s sports, and putting fans first.
Venezuela oil deal needs transparency to last
Concentrating control of Venezuela’s oil industry with the US government could ultimately backfire.
Implicit-bias-like patterns in reasoning models
Nature Machine Intelligence, Published online: 01 September 2026; doi:10.1038/s42256-026-01300-1 Lee and Lai study bias-like processing differences in large language reasoning models and find that, for most models, processing stereotypical information takes less computational effort than processing counter-stereotypical information.
ChatGPT now faces stricter EU oversight as a very large search engine
The EU Commission is classifying ChatGPT as a very large search engine under the Digital Services Act for the first time, with at least 45 million monthly EU users. By the end of 2026, OpenAI has to deliver risk assessments, transparency reports, and an ad archive, among other things. Whether the Commission can also demand access to training data is disputed among legal experts. The article ChatGPT now faces stricter EU oversight as a very large search engine appeared first on The Decoder .
Anthropic's Claude Code limit change is a raise on paper but a cut in practice
Anthropic is effectively cutting Claude Code's weekly usage limits by 17 percent. A temporary 50 percent boost expires on September 14 and will be replaced by a permanent 25 percent increase. Anthropic promises more control and transparency over usage in return. The article Anthropic's Claude Code limit change is a raise on paper but a cut in practice appeared first on The Decoder .
Women in UK arts feel they do not have equal opportunities for roles, says report
Exclusive: Research says sector is fixated with young talent, with most women experiencing unconscious bias or sexism Women working in the arts believe they do not have equal opportunities for roles, especially as they get older, according to a report saying that the industry is fixated with “young and emerging” talent. Nearly all (95%) women working in the arts thought gender inequality persisted, while three-quarters (75%) reported experiencing unconscious bias or sexism, according to a survey of 76 arts professionals. More than two-thirds (69%) felt that age negatively affected their career success – a figure that rose to 100% among those working in artistic rather than managerial roles. Continue reading...
Comment on AI in the Media and Entertainment Industry by James Brinsley
Many enthusiasts frequently wonder where can i buy hash online uk. We provide a secure, professional, and discreet platform that brings premium cannabis concentrates directly to your door. By prioritizing transparency and quality, we make it easier than ever to explore the world of hashish with total peace of mind. We understand that our customers value reliability and discretion, which is why every order is handled with the utmost care and shipped with professional speed. While some look for ways to buy hash online cheap, we believe that true value lies in the balance of quality, purity, and consistency. Our commitment to excellence means you don't have to compromise on your experience. Explore our diverse range of premium hashish today and discover your new favorite concentrate. With our easy-to-navigate website and dedicated customer service, your journey toward finding the finest hashish is just a few clicks away. Elevate your collection and enjoy the deep, earthy richness that only premium hash can deliver.
Comment on MIT Researchers Unveil “SEAL”: A New Step Towards Self-Improving AI by Paula Delgado
The shift from prompt-level self-improvement to direct weight updates via RL is a meaningful departure — curious whether SEAL's self-editing introduces any stability safeguards against catastrophic forgetting during the update loop.
VURA: A framework for trustworthy AI at scale
“You’re right,” the LLM says. “I was mistaken.” Have you ever read these words during an AI workflow? Nothing kills trust faster than incorrect outputs. It’s no wonder, then, that only a quarter of businesses today fully trust AI to support decision-making and forecasting. And yet, we know AI is business critical. Nine out of 10 businesses are using it; 64% say it’s powering innovation. So, how do you bridge the gap from experimentation to trustworthy deployment? How do you get verifiable, reproducible results from AI at scale? In this article, I’ll show you the framework that’s powering AI success for leading organizations. Why organizations still don’t trust AI We asked 1,400 IT and business leaders what their biggest barriers to success with AI workflows were. One in two (49%) said inaccurate or biased outputs; 38% said it was a reluctance to allow AI to make decisions without human oversight. Then, there was the data issue. Data readiness is an integral part of successful AI workflows. However, half of all organizations said they still faced poor quality or fragmented data. While you don’t need perfect data to start using LLMs, you absolutely need trustworthy data. VURA: The framework for trustworthy AI Closing this trust gap requires two things. First, organizations need a logic layer that connects AI systems to the people who understand the data and business best. Line-of-business teams and analysts cannot sit on the sidelines. They need to help build and validate AI w…
Pooled rate estimation for several Poisson processes under Type-I censoring with unrecorded censoring times
I have a pooled estimator that arose in an applied problem, and I would like to know whether it is already in the literature — I suspect it is, but I have not found it. Setup. Let $X_1,\dots,X_K$ be independent homogeneous Poisson processes on $(0,\infty)$ , where $X_b$ has rate $\lambda_b = \kappa c_b$ with $c_b>0$ known and $\kappa>0$ the single unknown parameter. Process $b$ is observed on a window $(0,W_b]$ and every event in that window is recorded. The unusual feature: $W_b$ is not recorded. We know that observation was exhaustive on some window, but not where it ended. What we observe for each $b$ is the number of events $N_b$ and the position $D_b$ of the last one. The two candidate estimators. Write $M=\sum_b N_b$ and $S=\sum_b c_b D_b$ . If one mistakenly treats the data as failure-truncated (observe until the $N_b$ -th event, $N_b$ fixed by design), then $\kappa S\sim\Gamma(M,1)$ exactly and $\tilde\kappa=(M-1)/S$ is unbiased. Under the actual scheme, let $\delta_b=W_b-D_b$ be the unobserved gap between the last event and the end of the window. By memorylessness $\delta_b$ is $\mathrm{Exp}(\lambda_b)$ truncated at $W_b$ , so $\mathbb{E}[\delta_b]=(1-e^{-\lambda_b W_b})/\lambda_b$ . Total exposure is $\sum_b c_b W_b = S+\sum_b c_b\delta_b$ , and since $\mathbb{E}[M]=\kappa\sum_b c_b W_b$ , moment matching gives $$\hat\kappa=\frac{M-K}{S}$$ up to a remainder $\sum_b e^{-\lambda_b W_b}$ , which is negligible whenever each window contains several events. What strikes…
Mixed Model Approach in Archaeology
Dear Stackexchange Community, I need some advice regarding a statistical analysis I want to conduct as part of my PhD. The general question of this is whether we can detect statistically significant (and relevant) differences in vessel morphology (shape) and morphometry (dimensions) of specific types (i got ten) between specific sites (five in total). Note, these types are similar across sites, as they follow an inter-regionally established vessel shape or form, produced by—supposedly—many producers across a wider geographic area. However, based on several theoretical concepts, it is assumed that each producer created their own respective variant due to different traditions of learning as well as environmental and social influences. So the idea behind this is to verify production patterns. Basically, the question is whether we can differentiate between places of production based on the morphology and morphometry of standardized vessels. The problem is manifold, and the main problem is the quantity and availability of the data, which is not good. However, there is nothing I can do about this, as this is simply the way it is due to excavation/publication bias, etc. I have one main site with the majority of the data, and I now wanted to, exploratively, use the available data to analyze tendencies. I have attached a crosstab which shows my data and the respective availability problems. Type Site 1 Site 2 Site 3 Site 4 Site 5 Type 1 73 28 32 0 3 Type 2 92 60 2 0 7 Type 3 74 78 2…
W. Kamau Bell has the most practical ‘most indispensable tool’
W. Kamau Bell is one of those people who has always just seemed to be there. From Totally Biased, to Politically Re-Active, United Shades of America, and We Need to Talk About Cosby, his blend of comedy, social commentary, and political activism has helped him stand out. He's won a Peabody and four Emmys, been […]
A possibly short proof of the bias-variance decomposition formula
I am reading the proof of the bias-variance decomposition formula on Wikipedia ( Bias-variance tradeoff ), and it appears to me that the derivation there is uselessly complicated. I think I can derive it in a much shorter way, but I am not sure if it is a 100% correct. Suppose that $y = f(x) + \epsilon$ , where $\epsilon$ is a random variable with mean equal to $0$ and variance $\sigma^2$ . Let $D$ be the training data, composed of $(x_i, y_i)$ , for, say, $i = 1, \dots, n$ , where the set of $(x, y)$ follows a joint density function $P(x, y)$ . We take a sample $(x, y)$ and denote by $\hat{f}(x; D)$ the predicted value corresponding to $x$ using some model trained on $D$ . The $\epsilon_i$ corresponding to the $i$ -th observation point is assumed to be independent of, say $\epsilon_j$ for $j \neq i$ ( $1 \leq i, j \leq n$ ) and also independent of $\epsilon$ . The bias-variance decomposition then says that $$ \mathbb{E}_{D, \epsilon}[(y - \hat{f}(x; D))^2] = (f(x) - \mathbf{E}_D[\hat{f}(x; D)])^2 + \operatorname{Var}_D[\hat{f}(x; D)] + \sigma^2. $$ The way I would proceed to prove it is as follows. Given a random variable $Z$ , we have that $$ \mathbb{E}[Z^2] = \mathbb{E}[Z]^2 + \operatorname{Var}(Z).$$ We apply that with $Z = y - \hat{f}(x; D)$ . We obtain that $$ \mathbb{E}_{D, \epsilon}[(y - \hat{f}(x; D))^2] = (f(x) - \mathbf{E}_D[\hat{f}(x; D)])^2 + \operatorname{Var}_{D, \epsilon}[y - \hat{f}(x; D)]. $$ Moreover if $U$ and $V$ are independent random variables and $\al…
In this experimental context, does the re-use of a control bias/invalidate the result?
I'm contributing to a paper where I am a fairly minor author. I'm concerned by the experimental design but the corresponding authors, who have decade(s) more experience in the field than I have, seem unconcerned and there's the implication that I'm being unnecessarily picky. I'm 95% sure I'm correct, but given their attitude and the fact that my remaining option is to have my name removed from the paper, I'm chasing that final 5% of sureness. In the first experiment, knocking out a gene was shown to cause a significant increase in cell size (simple t-test). To confirm this phenomenon, they then knocked out another gene involved in the same biological process and showed the same effect (also using a t-test). They then tested another three or four genes more tangentially related to that process and saw a significant effect in some of them and not others (t tests). Due to time constraints and the fact that, at the time, they didn't know if this was going anywhere, these three experiments all compare to the same control. While they've built on this in later experiments, these three experiments are all in the paper and provide the base for the latter experiments. This concerns me. Firstly, we know there’s a lot of variation in the experimental model over time so the control from the first experiment doesn't necessarily control for technical variations in the second two experiments. Secondly, and more importantly, in the second two experiments they compare the effects of disruptin…