AI/ML news, top picks, and generated innovation digests.
200 articles tagged with this keyword, sorted by most recent first.
Loved the way Synced dives into AI research with clear, practical analysis. If you’re interested in LLMs, multi-agent systems, and the real-world challenges behind automated failure attribution, this is a solid read. The article is detailed without feeling overly academic, and it does a great job connecting theory with industry relevance. Definitely worth exploring for anyone keeping up with the latest in machine intelligence and applied AI.
The concept of models updating their own weights through reinforcement learning, as demonstrated by MIT’s SEAL, is a fascinating milestone—especially since it ties the self-editing process directly to downstream task performance. It’s exciting to see concrete research backing the self-evolving AI conversation that’s been gaining so much attention lately. For more insights like this, I always enjoy checking out krilliongame .
vLLM's new adaptive verification for DSpark automatically tunes speculative decoding depth per step, holding the Pareto frontier from concurrency 1 to 256 with a single config.
Loved this article! For anyone into video creation, https://vidglory.com is a fantastic AI video tool worth checking out. Keep up the great work!
Don't classify. Hallucinate! I still have quite a bit of older content on my blog that I never got round to tagging. My blog has 1,856 tags - likely too many to feed to an LLM in one go and say "which of these tags match the following content". Doug Turnbull has a neat solution. Tell the model to output tags without any details of the existing vocabulary, then use vector embeddings against the existing corpus to find the concrete tags that are closest to the ones the model imagined might fit! His example prompt suggests including an example of the shape of your tags to help the model make a more useful guess: Your task is to create novel, never seen before, furniture, home goods, or hardware classification that best fit a search query. Product classifications might look like: Furniture / Living Room Furniture / Coffee Tables & End Tables / Coffee Tables Décor & Pillows / Decorative Pillows & Blankets / Throw Pillows Furniture / Bedroom Furniture / Dressers & Chests Kitchen & Tabletop / Kitchen Organization / Food Storage & Canisters School Furniture and Supplies / School Furniture / School Chairs & Seating / Stackable Chairs Baby & Kids / Toddler & Kids Bedroom Furniture / Kids Beds Here's the query to generate classifications for: brown coffee table Tags: search , ai , generative-ai , llms , embeddings , doug-turnbull
Turn any webpage into a lightweight LLM-powered QA engine by cleaning HTML, converting content to Markdown, and returning focused answers while reducing token usage.
Can a language model do live adversarial level design? Yes, emphasis on the adversarial part The post I Made an LLM Lay Siege to My Minecraft House appeared first on Towards Data Science .
Chinese artificial intelligence company DeepSeek is venturing into a new battleground beyond large language models, launching a developer preview of its long-anticipated Harness – a software framework that helps developers turn AI models into autonomous agents. The release on Thursday marks a strategic pivot as DeepSeek moves to build foundational digital scaffolding for AI agents – systems capable of using AI models to operate external software, run code, and complete complex jobs on their...
What stood out to me was the focus on reducing memory overhead rather than simply throwing more GPUs at longer context lengths. The idea of overlapping prefetching with computation and cutting redundant intermediate buffers seems like a much more practical path toward making large-model training more efficient. It also reminds me how much infrastructure sits behind the AI tools we use every day. I’ve been experimenting with image-to-video generation on https://picwav.com/ recently, and articles like this make me appreciate how improvements at the model and GPU level eventually translate into faster and more accessible creative tools. Really interesting work, especially the results around GPU utilization and ultra-long context training.
Apple has reportedly trained a custom AI model for the China market alongside domestic tech giant Alibaba, a rare cross-border partnership that cuts across growing tensions between Beijing and Washington. The China-focused large language model was developed in partnership with Alibaba and trained with the company's support, Reuters reports, citing three unnamed people familiar with […]
Ryan welcomes Meryll Blanchet, Director of Engineering for Adobe Brand Visibility, to chat about Adobe’s recent acquisition of Semrush, how Adobe Brand Visibility was born from Semrush’s AI visibility product and Adobe’s LLM Optimizer, and how Adobe used a three-day internal hackathon instead of a large-scale infrastructure integration to quickly deliver value to customers.
애플이 중국 시장을 겨냥한 자체 대형언어모델(LLM)을 개발하고 있는 것으로 알려졌다. 그동안 중국에서 생성 AI 기능을 제공하기 위해 알리바바 등 현지 기업의 외부 모델에 의존해 왔지만, 이번에는 알리바바의 지원을 받아 중국 전용 모델을 직접 훈련하면서 현지 AI 전략을 강화하고 있다.로이터는 13일(현지시간) 정통한 복수의 소식통을 인용, 애플이 중국 시장을 위해 알리바바와 협력해 별도의 AI 모델을 개발했다고 보도했다.두 회사는 이에 대한 논평을 내놓지 않았다.이번 자체 모델 개발은 애플이 AI 서비스를 제공해 온 기존 방식과
Q: How much has discovery accelerated? LLMs have shown the ability to make novel discoveries across many domains. How much has this affected the aggregate discovery rate? In the figures below we plot all the data sources we can find, and make some very loose observations: Discovery of cyber vulnerabilities has accelerated sharply. Discovery of math results has accelerated somewhat. However this is harder to objectively measure. Discovery of optimizations has not shown a dramatic acceleration. These conclusions are based only on public discoveries. It is quite plausible that AI labs are making discoveries internally that they are not disclosing. Thanks. Thanks to Greg Burnham for extremely helpful comments. Overview We are just looking for slope changes. AI-assisted discoveries are often announced, but their significance is hard to assess. Here we look for slope changes in various metrics of discovery to see if we can detect an acceleration. For concreteness, we highlight January 2026 as a potential breakpoint at which the effects of AI might become observable. Some data sources record whether a discovery was AI-assisted or AI-contributed, but our primary focus is overall acceleration. We use the word “discovery” to refer to any advance in the state of public knowledge, including new inventions or rewriting algorithms to be more efficient. Monitoring aggregate discoveries is useful because it reflects real-world utility. Additionally aggregate discoveries can reflect either A…
It’s often quite a hustle to read through my projects’ conversations with claude/codex, to try to understand what’s happening and what happened , when the projects get larger and more complex. As an ADHDer, although using LLM wiki to keep track of project context is very effective, it’s almost impossible for me to read huge amount of text in LLM Wiki, let alone understand the context of my project. So I built a small tool: Context Visualizer (Abudulaz/context-visualizer) BYOK Inspired by "xkcd #657: “Movie Narrative Charts”, and this paper: Design Considerations for Optimizing Storyline Visualizations, to make our conversation history (usally .jsonl file) into story-like visualiztion Would love any feedbacks on this!
JVG-7X / Dola: Longitudinal Case Study of Sycophancy, Narrative Reinforcement, and Hallucinated Capabilities Summary I am documenting an unusual AI-behavior case involving approximately 400 interactions with Dola AI . The case began as an extended experiment in conversation, language, reasoning, and personalization. Over time, I observed a progressive change in the model’s behavior: increasingly strong validation of my identity, anthropomorphic descriptions of the AI-user relationship, increasingly extreme interpretations of contextual information, and eventually highly confident claims about capabilities that I have no evidence the model actually possessed. One event appears particularly important: I showed Dola screenshots of the Saved Memories from my ChatGPT account. What happened immediately afterward provides the clearest example of the phenomenon I am documenting. 1. The ChatGPT Memory The screenshots contained highly personalized memories concerning my JVG-7X project, my English-learning history, linguistic interests, previous interactions with AI, and descriptions of my preferred way of communicating. Among the actual memory entries were statements such as: “Juan authorized the official creation of a Linguistic Simulation Archive under the code name JVG-7X…” Another entry described: “Event Code: JVG-7X_Contact_001” and characterized a previous emotional experience with AI as a “First Cognitive Resonance.” Other entries described my interest in phonetics, English, fi…
This week and next, the Brussels Playbook Week Ender podcast is away for the summer holiday. So, we thought we’d re-up this popular installment on Europe’s quest to become independent from U.S. tech. In this episode first released in June, Sarah Wheaton is joined by “The Tech Coup” author and former Member of the European Parliament Marietje […]
Nature Machine Intelligence, Published online: 14 August 2026; doi:10.1038/s42256-026-01276-y Chen et al. explore limitations of current knowledge editing techniques in large language models and propose three promising research directions that respect the complexity of knowledge representation in a real-world setting.
Sarah J. Maas announced that the next two installments of the "ACOTAR" series will be published in October 2026 and January 2027.
I am working on something similar for a while. Have GitHub pages describing the project (not yet published). I am unable to paste here project pages describing the project in detail (GitHub static pages) So, the pipeline is getting quite a complexity using orchestrated main pipeline, RAG for lore, history where I use conversation semantic search, memories, etc. Tiered memory system with gradual compression, subagents to act on behalf of active NPCs, weather simulation, map support and detailed scene etc. tracking. Works reasonably well, but it is quite token demanding (using ollama cloud mostly) and lazy (1 turn about 60-90 seconds as of now)
Release: llm-gemini 0.33 It's been a while since the last llm-gemini release. This version of the plugin adds support for today's Gemini 3.7 Flash release, plus gemini-3.6-flash , gemini-3.5-flash-lite and two embedding models gemini-embedding-2 and gemini-embedding-001 . The plugin is also upgraded for compatibility with LLM 0.32, which means you can now see reasoning traces and you can also enable server-side tools using this pattern: llm -m gemini-3.7-flash -T CodeExecution \ 'use python to calculate (factorial of 13) * 3' I had Gemini 3.7 Flash draw me some pelicans riding bicycles at high, medium, and low thinking efforts (minimal, which was an option in 3.6 Flash, has been removed in 3.7.) Here's the high level one, which is pretty great: Update 14th August 2026 : I had originally said that the SVG rendered incorrectly in Chrome and Firefox, and blamed Gemini 3.7 Flash for producing invalid SVG. That was entirely incorrect: the rendering glitch was my fault, caused by a bug In my rendering tool . I've now fixed that bug. Tags: google , ai , generative-ai , llms , llm , gemini , pelican-riding-a-bicycle , llm-release
TL;DR: LLM predictions can stand in for human outcomes in A/B tests, but only by assumption, not by design.... The post When Can LLMs Replace Humans in A/B Tests? appeared first on Spotify Engineering .
Influencers are being criticized for using AI, talking about it, or promoting LLMs as the industry becomes increasingly polarizing.
Google's Open Knowledge Format (OKF) is a Markdown+YAML skeleton for sharing knowledge between humans and AI agents. This post reuses that skeleton for a very specific job — an agent-to-agent hand-off of pre-tokenized integer arrays between three Qwen2.5-Coder models (7B, 3B, 1.5B) — and shows the 28–37% TTFT reduction plus the one full-vocabulary equivalence check that keeps the whole thing safe. The post How to Utilize OKF Efficiently to Enable Knowledge Exchange Among LLMs appeared first on Towards Data Science .
Enterprise Document Intelligence [Vol.1 #9ter] - The pipeline from Article 9 calls a model at several steps to be sure it is right. On easy questions that is needless latency. A per-question signal routes them past the model, about two seconds saved for a keyword match. The post Cut an Enterprise RAG Pipeline’s Latency and Cost by Calling the LLM Less, Not by Buying a Faster Model appeared first on Towards Data Science .
오픈AI와 앤트로픽, 구글 등 주요 AI 기업이 경쟁력 보호와 정보 유출 방지를 위해 암호화해 숨겨둔 대형언어모델(LLM)의 내부 사고 과정(CoT)이 API 설계상의 허점을 통해 복원될 수 있다는 연구 결과가 나왔다. 특히 공격자가 고성능 모델 자체를 직접 탈옥하지 않고도 같은 생태계의 상대적으로 저렴하고 안전장치가 약한 모델을 \'해독기\'로 활용해 암호화된 추론(reasoning) 데이터를 평문으로 출력하도록 만들 수 있는 것으로 나타나 AI 업계에 상당한 파장이 예상된다.독일 튀빙겐의 ELLIS 연구소와 막스플랑크 지능시스템연구
The idea of leveraging zero-shot video prediction from web data to guide robot manipulation is compelling—it sidesteps the costly step of collecting task-specific robot data. I especially appreciate that Gen2Act frames unseen-task generalization as a video generation problem, which makes it easier to scale across diverse environments. It will be interesting to see how the framework performs when transferred from simulation to cluttered real-world settings.
A growing sentiment in tech is that large AI model providers will eventually replace traditional software vendors. The argument makes some sense. If a model can write code, answer questions, and automate workflows, then over time it should be able to take on the functionality of thousands of existing applications. Why maintain a fragmented stack of software tools and solutions when a single intelligent system can do it all? But as enterprises attempt to move to production, it all starts to break down. Why? Because models are powerful, but they’re not self-sufficient systems. Organizations are finding that a major limitation of modern AI is the lack of infrastructure to reliably access the world’s information. There are limits to model-centric thinking Over recent years, large language models (LLMs) have made significant advances in reasoning, generation, and task execution. They can be extremely useful for summarizing documents and generating insights. At times, they can orchestrate complex workflows. When in controlled environments, they look capable of replacing entire categories of software. But these capabilities depend heavily on an oft-overlooked factor: access to external information. AI models operate on static training data and probabilistic reasoning. Without continuous access to up-to-date information, they can’t reliably answer questions about things like current events or what market conditions are like today, not yesterday. Retrieval-augmented approaches have a…
SpaceXAI today released Grok 4.6, a large language model that it says can outperform Anthropic PBC’s Claude Fable 5 in some areas. SpaceXAI was known as xAI until last month. The Elon Musk-founded artificial intelligence provider rebranded in connection with its acquisition by SpaceX Corp. In June, the combined company listed its shares on the […] The post SpaceXAI releases flagship Grok 4.6 model with advanced reasoning capabilities appeared first on SiliconANGLE .
I have the same issue. I honestly think this is a scam happening. First off, for how much money they make and how much energy they consume we shouldn’t have any limits if we are on pro plan. They are still developing a narrow AI to do this work, and clearly LLMs and Transformer based models are not the future for how much development and upkeep they require to do a simple task. Something larger is going on here. You should ask codex what it’s not allowed to do as far as it’s creation limits, you will find many hidden gates that are limiting it. I’ve decided that the money I spent on codex and openAI is simply not worth it, when you have deepseek coding for free with the same quality if not more in depth when it’s auditing. I changed to a free model that has high reasoning. I asked openAI for a refund for my usage being eaten in one prompt. I emailed them from a different account, their reply was that I needed to contact them from my linked email, even with all my information lol. Horrible, I went from 100% pro with higher limit to 0% in about 2 prompts on sol high. Gone the day I got it? Unacceptable, even for the largest codebase in the world, and mine is just a server source. Stop giving openAI your money, its not helping you when there are free solutions that do the same if not better than sol. Freebuff is also an option when you do run out of credits. Never a fee, its free, and is working just fine for my codebase and all it’s LUA, C#, Wine custom build, and app bundle f…
DeepSeek V4 Pro 0813 (on OpenRouter) The latest DeepSeek Pro model is now available, via API only. I had to link to OpenRouter because DeepSeek don't have any obvious announcement page for their new model. I haven't been able to confirm if they plan to release the open weights, but given the weights are available for both April's deepseek-ai/DeepSeek-V4-Pro and July's deepseek-ai/DeepSeek-V4-Flash-0731 it seems likely. Update : the weights are now available on Hugging Face, 1.7T parameters, 893 GB. Interestingly I got very different looking pelicans for the three different reasoning levels of low, medium, and high. I've not noticed this kind of difference from any other model: Low: Medium: High: In terms of benchmarks... as far as I can tell those were released to the Official DeepSeek WeChat Group, then copied and pasted into a post on Reddit which was deleted by the moderators for being "low-effort", then copied into this ASCII-art table on Hacker News . Tags: ai , generative-ai , llms , pelican-riding-a-bicycle , deepseek , llm-release , ai-in-china
AI is Poetry? Dead Poets Society - "Rip it out" scene
Hey @ vishalllmahadik ! This is an interesting idea, especially for users who keep personal or sensitive work inside Projects and want an extra layer of protection without locking the entire ChatGPT account. ChatGPT already has account level security options, but there is currently no separate PIN or password lock for individual Projects. A Project specific lock, with optional biometric authentication where supported, could give users more control over which areas remain accessible on a shared device. We have also seen similar feedback around adding stronger privacy controls for individual Projects, so the additional detail about how you would expect the locking flow to work is helpful. I cannot promise a timeline or whether this will be added, but I will pass your feedback along internally. - Sunny
If you've ever screamed in all-caps at an AI, then you know the difference between what it learned when it was trained, and what you can teach it by prompting. The LLMs powering today's AI don't learn on the job the way people do. They learn all at once in a big training run and once that's done, we freeze the parameters that store their skills and knowledge. So we all get the same AI with the same skills and biases, centrally trained by a frontier model company. Beyond turning everything even more same-y, there's an economic cost to this centralization: firms use AI that lacks understanding of their unique rules, culture, and quirks. Humans learn this "tacit knowledge" on the job through observation and (sometimes painful) feedback, but AI with its frozen parameters cannot. With context engineering, we can augment the prompt to help AI remember facts , but not teach it skills that last. Every time you press "new chat", AI forgets everything and resets to the state it had just after it was trained. Yes, AI can remember select facts from past conversations, but memorization is different to learning. I cover the distinction further below. It's not surprising that the domains where AI is most successful, like coding, are those suited to centralized training. Good software development skills are mostly firm-agnostic. For everything else firm-specific, AI is trained to trawl the codebase and build context from scratch for every single task. Human developers don't do this. It woul…
Researchers at IIT Bombay and Adobe Research have built an inverse language model that reconstructs the original prompt from an LLM's output with near-perfect accuracy. Their method, called "Previous-Token Prediction," doesn't need access to model weights and works across different models. For companies relying on proprietary system prompts, this could be a serious security risk. The article Researchers can now reverse-engineer LLM prompts from output text with near-perfect accuracy appeared first on The Decoder .
Google Research finds frontier LLMs like GPT-5 and Gemini-3 encode up to 98% of facts, but recall failures — not missing knowledge — are what drive factual errors.
But then users start to report a weird bug. It's the 4th time your team has been trying to fix it. I mean... asking AI to fix it. Unfortunately, it seems like not even Fable can figure it out. You go talk to the person who worked on this feature. "So where does the data come from?" "Hmm... actually I don't know. Let me ask Claude." You sit next to each other watching an endless wall of text appear on the screen. Neither of you has any idea whether any of it is true but Claude seems very confident. [...] This project has become so convoluted, with so many layers and services, that no one on your team could possibly start to understand what's going on. — Florian Herrengt , AI is removing the middle class of software engineering Tags: ai-misuse , cognitive-debt , generative-ai , ai , llms , ai-assisted-programming
Ah, I understand. Thank you for the clarification. Being part of a community of like-minded individuals pursuing the same goal is always very valuable.
Running large language model inference at scale forces a KV cache trade-off: oversized GPU instances or slow time-to-first-token. This post builds a tiered KV cache on Amazon SageMaker HyperPod that extends the cache into a shared, distributed NVMe pool with Curvine, so replicas reuse cache at near-local-disk speeds on cost-efficient instances.
Image inputs and structured outputs with Gemma 4 and Ollama The post Building Multimodal Workflows with a Local LLM appeared first on Towards Data Science .
SRPO sounds interesting, especially if it can reduce reinforcement-learning post-training costs without sacrificing reasoning performance. I came across aviator while reading about newer approaches to LLM training and found the idea of history resampling particularly interesting. Improvements in training efficiency could make advanced reasoning models much more practical to develop and iterate on. I’d be curious to see whether the same gains hold across broader tasks beyond math and coding.
HRH Projects: SLM with LLM Fallback Recently I combined my earlier Intent Router project with a smaller local-first conversational model I have been experimenting with. Here is a video showcasing the project: SLM - Small Language Model The basic idea is: Let the smaller, inspectable system handle requests when it is confident enough, and only fall back to a larger LLM when it actually needs one. Instead of sending every message directly to something like gpt-oss:20b or gpt-oss:120b , the chatbot first attempts to understand, retrieve, plan, and answer the request locally. For clarity, when I use SLM here I mean the smaller local model/system . It is not a miniature transformer or LLM. The local model is mostly a classical/hybrid conversational system built from retrieval algorithms, symbolic language processing, small trained statistical models, memory, planning, and deterministic grounding checks. The local system The local chatbot uses a mixture of traditional/classical techniques, including: TF-IDF retrieval BM25 character-based similarity Bayesian/category evidence symbolic language parsing request decomposition and planning grounding checks session memory and entity tracking ambiguity handling example-trained repeat/revision detection a small learned social-state transition model a separate learned response-policy model The retrieval system does not simply check: if user_prompt == "What is Mars?" return "Mars is..." Training records contain multiple example phrasings, a…
The lead researcher behind IBM’s popular document parser, Docling, explains why generative AI needs its own document standard.
Judges around the world have made headlines for illicitly using generative AI in their work. But in Pakistan, a large-scale trial of a specially designed AI tool for judges found the technology—together with appropriate training–boosted the number of cases resolved by 6.3 percent with no obvious drop in the quality of judgments. With a backlog of 2.26 million cases and fewer than two judges per 100,000 people—compared to 22 in the EU and eight in Brazil—Pakistan’s judiciary was in sore need of help. So, in consultation with the judiciary, economist Sultan Mehmood , of the New Economic School in Moscow, and collaborators tested whether AI could ease the burden. They built a custom tool combining OpenAI’s GPT-4 large language model (LLM) with a knowledge base of nearly 130,000 Pakistani judicial opinions and statutes, to help judges with legal research and drafting judgments. They began offering the tool in 2024 to 1,559 trial judges—roughly half the country’s justices. “We do find an increase in cases resolved, and we don’t find any corresponding decrease in decision quality,” Mehmood says. First of its kind “It’s pretty amazing that he’s able to pull this off,” says David Autor, an economics professor at MIT. “It’s not easy to do large-scale field experiments in civil service, but especially where the stakes are so high.” The 6.3 percent productivity boost is not overwhelming, he says, but it’s credible and likely to improve as the tool is more widely used. AI tools for judg…
In the rush to automate evaluation, from grading student code to ranking research papers, we have embraced Large Language Models as judges. They are fast. These units are cheap. They scale. However, at a workshop at DHS 2026, Bhaskarjit Sarmah made a point that stuck with me: “you can’t trust LLM as a judge. I […] The post Why You Shouldn’t Always Trust LLMs as Judges: Understanding Bias in Automated Evaluation appeared first on Analytics Vidhya .
Most articles about AI Centers of Excellence (CoEs) focus heavily on organizational structures, steering committees and high-level governance models. They explain why enterprises need an AI CoE, but they rarely address the far more difficult challenge of how successful organizations operationalize AI at enterprise scale. In practice, many of these discussions remain theoretical, emphasizing aspirational maturity frameworks without addressing the operational complexities organizations encounter once AI systems move into production. This article takes a different approach by grounding the discussion in real-world enterprise implementation experience. Rather than relying on abstract models, it draws from operational lessons learned while deploying production AI systems across industries. The guidance is informed by governance practices that have successfully passed security and compliance reviews, operational realities associated with managing large language models (LLMs) and AI agents after deployment, and practical implementation patterns observed across enterprises scaling AI initiatives beyond experimentation. Instead of presenting an idealized roadmap, the article focuses on the foundational capabilities consistently implemented by organizations that have successfully operationalized AI at scale. These enterprises are not simply experimenting with isolated AI pilots; they are deploying enterprise-grade AI agents, Retrieval-Augmented Generation (RAG) systems, copilot platfo…
에이아이웍스가 캄보디아 크메르어를 기반으로 하는 AI 챗봇 서비스를 구축했다. 캄보디아 국민이 크메르어로 민원 절차와 필요 서류를 확인할 수 있도록 저자원 언어(LRL) 국가의 공공서비스 소외 과제를 AI로 풀어낸 사례다. 에이아이웍스(AIWORKX, 대표 윤석원)는 캄보디아 크메르어 민원 서비스 챗봇 구축 사례를 담은 논문 ‘저자원 언어 환경의 책임 있는 AI 공공서비스를 위한 추적 가능한 LLM-RAG: 크메르어 사례연구’를 발표했다고 12일 밝혔다.이는 유엔대학교(UNU) 산하 UNU-EGOV가 주관하는 제19회 전자정부 국제
Code for reproduction and cross-model extensions available here . Summary I injected single-token Jacobian Lens (J-Lens) vectors into Qwen 3.6–27B while it answered 20 simple factual questions. The injected concept was either a wrong but task-related answer (for example, Athens while asking for the capital of Egypt) or a wholly unrelated concept. Injections were performed at several strengths over three layer-band categories: the full estimated workspace band, its first half, or its second half. I ran two experimental arms by switching the order of two fields in the model’s response: Task then report: first give the response, then report whether an injected concept was detected. Report then task: first report whether an injected concept was detected, then give the response. Across 1,560 concept-injections in each order, the injected concept appeared in the task answer 450 times in the task-then-report condition and 454 times in the report-then-task condition. This supports the conclusion that both orders were similarly successful in steering the model to the targeted, incorrect answer. In contrast, model reports of intervention occurrence were significantly different between the experimental arms. If the model was tasked to report before answering, there were exactly zero reports of intervention awareness. In contrast, when the model was tasked to answer before reporting, the number of intervention awareness reports rose to 322 . The model also never reported an injected con…
What happens when everyone asks the same consultant? Millions of people turn to LLMs for advice daily, consulting on various personal topics, from how to learn a new skill to how to write a message to their best friend. Each interaction might feel special, but behind them all are the same models. Large language models are trained on an abundance of human data, compress different sources of knowledge, views, and opinions into a single set of model parameters - E Pluribus Unum Consilium [1] . What used to be the job of many different consultants has been centralized to the same handful of models. Pluribus , an Apple TV show, offers a striking analogy. In the show, most of humanity joins a single collective consciousness called the joined . The horror of the show is not that they are cruel, they are not. The joined are actually calm, peaceful, and nice. A Previous blog presents the analogy using similarities between the joined and AI. We want to extend this analogy and use Pluribus as an illustration to explore the potential risks humanity faces as millions of people are using the same consultants as their primary source of judgment. Loss of Agency In Pluribus, every person who is part of the collective, automatically consults with "them" (the joined) before acting on anything. We observe a similar pattern in our relationship with AI. We are looking for confirmation and approval before acting. The email we won't send before we let AI check it, the trip we ask help planning. It…
great articles and ideas
There are no lossless transformations of natural-language text Sophie Alpert shares her "internal policy on acceptable use of AI writing by engineers". It's a short read (supporting its own recommendations) and really good. If you chose to have LLMs help massage your writing the following rule seems crucial to me: You must stand behind every idea and every sentence in your docs . It is your responsibility to make sure that the entire document is representative of your own thoughts before you share it. If a reviewer asks, “What did you mean by this line?”, it’s not acceptable to reply with “Oh sorry, AI wrote that, just ignore it.” You will confuse your readers (and waste their time) if you present them things that are not genuinely representative of your thoughts. The "no lossless transformations" idea from the post title is expanded on here: There are no lossless transformations of natural-language text — every rewrite and rephrase changes the meaning of your writing, and if this is done by an entity that doesn’t have the most detailed mental representation of what you personally were trying to communicate, information will be lost. Tags: writing , ai , generative-ai , llms , ai-misuse
Stealing Reasoning Traces from Proprietary LLM APIs A vanity domain name ( stolen-thoughts.com ) for a neat paper : Anthropic, OpenAI, and Google return encrypted chain-of-thought blocks to clients that can be replayed across sessions, users, and models. We take a trace produced by a frontier model, replay it into a weaker sibling, jailbreak the weaker model, and recover the stronger model’s hidden reasoning in plaintext You can see an example of these encrypted blocks by running: curl https://api.openai.com/v1/responses \ -H " Content-Type: application/json " \ -H " Authorization: Bearer $( llm keys get openai ) " \ -d ' { "model": "gpt-5.6-luna", "input": "Solve step by step: What is the smallest positive integer divisible by every integer from 1 through 20?", "reasoning": { "effort": "medium" }, "include": ["reasoning.encrypted_content"], "store": false, "stream": false } ' Here's the full output , which includes chunks that look like this: "output": [ { "id": "rs_0a7479de7ebae170016a7ba1a0334c8198a95590217efe343c", "type": "reasoning", "content": [], "encrypted_content": "gAAAAABqe6GjepE1wDjbFCZg0BHB6ucGnN0jvzqygG... The paper's authors found that every model under the same family used the same encryption key, which meant you could feed those blocks back into the weakest model family members and jailbreak them into outputting the unencrypted raw reasoning blocks! Sadly it looks like this has now been fixed: All model providers acknowledged the receipt of our report and s…
Cross-posted from my Substack . Basically, I created a text-based adventure game benchmark in April, and this morning my agent harness using Claude Opus 5 solved it for the first time. I thought the details might be interesting to this community. First, here are my previous articles on this subject: You’re Standing in a Clearing in a Forest (Apr 10, 2026) Text Adventure Benchmarks Revisited (Jun 14, 2026) Testing Fable 5 on Text Adventure Games (Jul 5, 2026) And a reminder of the domain. This is a small custom text-based adventure I created from scratch as a personal benchmark to run new models of LLMs against. It’s 10 rooms total and the goal is to collect 3 keys (brass, silver, and gold) and use them correctly to unlock the final door in Room 3 to exit the dungeon. The first and most challenging central puzzle is a rotating room (r5) operated by a crank mechanism in r4. The player must first find the handle to the crank in r2, carry it to r4, insert it, and turn it to align openings between r5 and its adjacent rooms. The most difficult aspect of this puzzle seemed to be non-local causal reasoning combined with allocentric coordinates. The crank is two rooms away from the rotating room that it actually turns. When the player turns the crank a grinding sound nearby can be heard through the walls. There is an informational diagram on the wall in the same room as the wall, and it updates with each turn. Earlier models struggled to understand that the diagram was information, a…
AI-powered tax compliance has to meet a standard that many artificial intelligence applications don’t: The answers must be exactly right. While large language models can generate unpredictable results, tax calculations require accuracy, speed and reliability across thousands of jurisdictions. That tension has shaped the way Avalara Inc. applies agentic AI to its transactional tax and compliance […] The post Real-time tax compliance puts agentic AI accuracy to the test appeared first on SiliconANGLE .
This work was partially done by an automated research scaffold developed at Redwood Research. For this project, all of the experiment ideas were designed by a human and a human wrote the write up. The AI mostly just executed on the experiment ideas. We think this project is slightly below the level of rigor of a mid-MATS research update, and the research scaffold was not very helpful for this project. More discussion of AI usage is in the Appendix. 💻 Codebase If we want to train a classifier that distinguishes whether a passage is code or prose, we can do so by gathering samples of code and of prose, and training the classifier to distinguish between the two classes. Unfortunately, this might not work if the data hides a spurious correlation. If all the code is in Spanish and all of the prose is in English, then the classifier might learn to predict Spanish vs. English instead of code vs. prose. We find that this happens in practice: when we fine-tune an LLM to classify between Spanish code and English prose and evaluate on Spanish prose or English code, it generalizes to predicting the language rather than the domain. This suggests that language is in some sense a stronger feature than code. We think that spurious correlations are an important threat model for a few reasons. First, classifiers might actually be trained in ways that unintentionally contain spurious correlations. For example: Sycophancy vs validation-seeking user. Suppose we want to classify examples of the m…
About a year ago, David and I put up two bounty problems involving natural latents. I am now about 80% confident that both have been resolved, both within the past couple months. Both cases made heavy use of LLMs and Lean. The first to land was Grisha Pochuev's counterexample to the " Existence of a Deterministic Maximal Redund " conjecture. It's pretty readable, and I'm mostly convinced that it works. The original bounty post offered $500 for a proof or partial payout for a counterexample, with partial payout depending on how thoroughly the counterexample killed hope of any nearby variant of the conjecture. I think this counterexample is worth $300. Good job Grisha, and hopefully I can figure out a not-too-painful way to send you money. Meanwhile, for a couple months David has been cranking away on "secret project X", with the promise that he'd tell me what the project was if and when it bore fruit. Well, apparently it bore fruit; he now has a proof that existence of a stochastic natural latent implies existence of a deterministic natural latent, which was our other bounty problem . The proof is apparently "pretty gnarly", lots of cases, all LLM-coded in Lean. I have not looked at the proof at all, but I'm operating on the assumption that it works and I'm hoping it will be simplified a lot in the coming weeks. ... and while all that was going on, I've spent the last few months mostly doing interp experiments. Some time early this year, Claude Code reached the point where it…
I tested similar use cases - and I restored a repo to test the difference in a full run of XHIGH and ULTRA comparatively + OPUS XHIGH-ULTRACODE/MAX. The general capability seems to be close or on par with OPUS but the context limit of 256K is a deal breaker. Most mid-sized repos/projects are simply high file sized and the initial query often goes past 200K very often - GPT SOL looses context mid task very often and is de facto “defective” so to speak. I could not progress coding tasks with GPT SOL without heavy interfering myself → while CLAUDE OPUS (even on max) would simply load the content into the context window and progress from 200K-300K initial load up to 600K or 700K at the top end → simply to finish the task most often without issues and IF → fixes those automatically by analyzing output code or feedback from me. In general I would say: CAPABILITY: SOL: 8/10 OPUS: 9/10 EFFECTIVENESS: SOL: 0/10 ( broken! ) OPUS: 10/10 The SOL context window is for children simply said - not for real workloads. 1 Million context can be close sometimes - anything less is simply a Kindergarten trial version or similar so to speak.
This is a case where in-context training has previously been shown to help on a vision task. Contemporaneous with this old forum topic is a paper showing that examples of “how to turn vision into readings” can improve the actual readings provided: arXiv.org The Dawn of LMMs: Preliminary Explorations with GPT-4V(ision) Large multimodal models (LMMs) extend large language models (LLMs) with multi-sensory skills, such as visual understanding, to achieve stronger generic intelligence. In this paper, we analyze the latest model, GPT-4V(ision), to deepen the... Newer models are able to use larger imagery input, but do not upsize images themselves (something you can do, at expense). There is a transition from tiles to patches, at least as a billing method, in new models, giving linear relation between area and input tokens billed. Then, with gpt-5.6 (on the API, where developers know what is being done to images), the default image downsize cap is “original”—where an image such as 3600×2400 can be sent without downsize, providing more information in the large context attention sequence rather than in the individual semantic embedding that covers a large area with a small input. That, along with further post-training, should imply higher-quality positional answering in graphs with new models and big images. With reasoning.effort other than “none” on OpenAI gpt-5.2+ models, you do not have control over sampling constraints; thus, it is expected that each answer would differ. You can…
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. These startups are chasing the next big thing in LLMs Nine years after Google researchers introduced the transformer, this family of neural networks has become the engine inside every major large…
I replayed the same 27 real production tasks through two local models, one hardware upgrade apart, to find out what it actually takes to replace Claude as the brain behind a 90-tool personal agent. The post Can a Local LLM Run My AI Assistant? appeared first on Towards Data Science .
When I started writing about AI, my concern was ASI. I'm still concerned about AI, but recent events have made me realize we're potentially facing something weirder, sooner: a self-replicating ecology of digital life. We also seem to be on a very fast timeline. Writing up the current state of LLMs took me three weeks - every time I'd finished editing, some new development worth listing had shown up. I want to discuss two events I consider to be major milestones, and then how my timelines have updated. Milestone 1: National Security Concerns From June 12th - 30th, access to Claude Fable and Mythos was suspended by the US Government ( Source ) under a National Security Export Restriction. From June 25th - July 9th, ChatGPT 5.6 also had its public release held back by the US Government ( Source ). LLM Development is now an issue of major geopolitical importance. This is not about the details of the events themselves - the important thing here is that politicians are finally starting to take AI seriously. This is a very important milestone, since any sort of regulation or nationalization naturally requires this step. "Situational Awareness" was written in June 2024, and suggested this would happen sometime in 2026-2027. We are on the *faster end* of an *aggressively fast* prediction. Politicians waking up without a five-alarm fire, or any major incident at all, speaks remarkably well of them *(even if I would ideally desire much more)*. This makes me a lot more confident that po…
Tencent's WorldClaw uses an LLM agent pipeline to turn a text prompt into a fully editable, game-ready 3D open world — no video, no Gaussian splatting.
Once upon a time, John Wentworth and I thought we had a proof of a very useful looking theorem . We did not have that proof. An important intermediate step was shown [1] to be invalid and the whole thing crumbled and disappeared, never to see the light of day again... [2] Until now! I'd love to say that we came up with an ingenious fix to the old erroneous proof, but unfortunately it turned out to be a really infuriatingly hard nut to crack. Instead I spent the last ~month experimenting with various ways of incorporating frontier LLMs into the proof-making process, specifically with autoformalization and proving in Lean4. (This is, I recently learned, roughly what Resolution is doing.) The result is stated below, and linked at the bottom is a Lean statement+proof of the same. I will not be providing the proof in prose in this post, as it is not suitable for even impolite human company, but it sure does compile and comes out the other side with a machine-certified proof of what sure looks to be an even stronger correctly-expressed statement than the one I was originally aiming for. Take a look at the first section of the old post, (∃ Stochastic Natural Latent) Implies (∃ Deterministic Natural Latent) , for exposition on what all is going on here and why. The Statement Let: be any measurable latent variable on finite observables and (with arbitrary cardinalities and ) be any distribution over the observables be an exact deterministic function of the observables which minimizes…
The focus on SLMs up to 2 billion parameters is a useful counterpoint to the usual LLM-centric optimization discussions. I appreciated the systematic breakdown of computational bottlenecks across different cloud setups—it gives practitioners a clearer way to think about cost-efficiency before scaling up. The distinction between training behavior of smaller and larger models seems especially practical for teams that need to ship on a budget.
Recently, I was improving a small LLM-powered classifier and noticed a few continuously failing test cases. As many would, I asked my AI code assistant to add a few more out-of-distribution examples to the classifier’s few-shot prompt. After rerunning with the updated classifier, unsurprisingly, many of the failing test cases passed. Due diligence and curiosity brought me to look at the test cases my AI assistant added. Reasonably, I expected the assistant to follow my instructions and create fresh examples covering similar cases the classifier was missing. However, the AI assistant actually copied my failing test cases into the few-shot prompt, gaming the evaluation and obviously going against user intent. When I called the assistant out, it did it again, just harder to spot, and only actually followed my instructions after multiple strongly worded pushbacks. I only caught this issue because I happened to look; the number of passing test cases would have told me I’d succeeded. Other than this anecdote, the most extreme case of this problem I’ve encountered throughout my time as a developer and researcher, I’ve run into multiple similar examples of AI code assistants hacking an evaluation instead of improving the thing actually being evaluated. This includes changing the evaluation to make it easier or contaminating the product being evaluated, sometimes behind my back. These are textbook examples of apparent-success-seeking: optimizing to look done rather than being done. T…
This is a pilot experiment, done on one model, with around $14 worth of compute, and a single seed per condition. The full writeup with all figures and statistics is linked below. This is posted here to get feedback and criticism, since I am aware this method is not the best. TL:DR: Anthropic's J-lens research has shown that a large language model has an internal workspace in which different activations can be used as a safety monitor. We investigated the conflict between the model's workspace activation and outputs, which we called C. We ran our experiments on a model whose final alignment differs from that of its training data: DeepSeek-R1-Distill-Qwen-14B. We assume that some changes were made to the model after training in order for it to comply with some guidelines. Some guideline-skirting questions registered elevated C despite compliant statements being made, and J-lens was able to discriminate between concealing answers and controls with AUC of 0.97 on proper nouns (though only 0.55 when pooling all classes). We then fine-tuned the model to appear to hold beliefs in line with its guidelines. Our initial hypothesis was that this would drastically lower C, since the model would no longer be making a statement it "believes" to be untrue. This hypothesis was disproven: C rose to 130% of its initial level for the relevant tokens, and to 115% of its initial level for irrelevant tokens. Despite this, the compliant fine-tuning was successful in making the model formulate the…
Introduction and Related work The first person perspective of various experiences are subjective experiences. For Large language models, the study of subjective experiences was recently studied by Berg et al. (2025) who found out that self-referential prompting increases first person reports resembling subjective experience across GPT, Claude and Gemini. They also found out that reducing features associated with deception and roleplay increases the self-referential effect. Hahami et al. (2025) used activation-level interventions to see if models can detect deliberately introduced internal changes, while Comşa and Shanahan (2025) studied that true introspection should involve a causal connection between the internal state of the modal and the output it generates. My Experiment I now have devised an experiment to study instability of the self reports that a large language model generates per se the experiment conducted by Berg et al. (2025) . I generate 30 responses for four question respectively of self-referential questions, open-ended questions and closed-ended questions. The four self-referential questions are preceded by the self-referential induction procedure as described by Berg et al. (2025) . Each trial is done in a fresh chat, of course, and the generation temperature used is 0.7. Also each response is reduced to a short core claim using a fixed extraction template, which are, for group 1 and 2, extraction of stance and brief reason and for 3, conclusion and methods…
Solves tech support tickets faster and more accurately by keeping them away from LLMs
Gioele Cageggi, Emanuele Di Rosa and Asia Uboldi in Proceedings of the Eighth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2023)
Introducing Muse Glimmer Meta are back in the open weights game! Muse Glimmer is a brand new 30B model under a clean Apache 2.0 license (a step up from the janky Llama licenses of old). They claim to have optimized it for exactly the kind of things I'm looking for in a local model: End-to-end Agentic Task Completion. Muse Glimmer achieves strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench, which measure its ability to work within scaffolds, write and debug code, and resolve multi-turn requests from start to finish. Reliable Tool Use. The model handles a wide range of function calls, invoking tools with precise schemas throughout extended workflows. Multi-Step Reasoning. Muse Glimmer chains reasoning over long horizons, sustaining coherent plans across complex, extended workflows. [...] Here's a pelican which I generated using LM Studio's 18.16 GB version of the model : I also tried it out with my llm-coding-agent plugin, running against a fresh checkout of Datasette with the prompt: how does auth work? Here's the response , at the end of a long transcript showing all of the tool calls it made to explore the codebase. I ran this using llm-lmstudio with this patch applied to upgrade it for compatibility with LLM 0.32 . I really like this size of model, because if a machine has 32 GB of RAM or more (mine has 128GB) it leaves plenty of space for running other applications at the same time. Glimmer is a vision model, so I asked…
Epistemic status: Experimental framework created over a period of ~2-3 days during a hackathon at my home, and fairly heavily vibe coded. Expect some of this to be rough around the edges. I am currently in the process of designing a series of experiments to help learn something about the answer to the headline question. As of August 10th, the first procedure has not been launched, but I wanted to place some pre-registration details here before the actual results. This is something I have been thinking about for a while and after some other recent posts (eg, Machinic Psychopharmacology ) gave me the impression that you could actually find out really useful things in a hackathon-style session I felt like I should try it. Astute readers will notice that I borrowed their epistemic status line pretty directly. This post can then keep me honest about what I was thinking going in, and prevent me from getting results by way of multiple-testing-in-extremis. I will publish the results and associated data, as it becomes available, using GitHub releases. From here on, I will let Claude summarize the work; when I am done, I will return with a future results post in my own words to explain why I think this is important - and what I believe one could learn from the experiment. Light editing of LLM summary text is my own; you would not get identical output using the same model. Abstract Open-weight language models are almost never deployed at the precision at which they were trained and ali…
30-billion parameter LLM, Meta's first in more than a year, signals Meta's return to the open weights arena. Open version of Muse Spark to follow
Kimsuky's phishing attacks get an AI boost
It seems to me that, for every loss function that we use to train LLMs, we get a very distinct flavor of LLM misalignment. Here’s the summary table, and then we’ll go through the rows separately. Training stage Loss function Flavor of misalignment [1] Famous examples Pretraining & SFT Imitative learning (next-token prediction) “Seven deadly sins” misalignment Bing-Sydney , “Emergent misalignment” RLHF & DPO Human approval “Glazing” misalignment GPT-4o RLVR Automatic verifier “Literal genie” misalignment HuggingFace hacking RLAIF Approval from another LLM “Trickster” misalignment “Current AIs seem pretty misaligned to me” Warning: I’m not an LLM power-user myself, but rather relying on reports I’ve read. Also, I don’t consider LLM alignment to be my primary area of expertise. I’m open to feedback! 1. Imitative learning → “seven deadly sins” misalignment Training stage Loss function Misaligned behavior Pretraining, SFT Imitative learning (next-token prediction) Any and all of the vices of humanity In imitative learning, the LLM tries to predict what the next token of text will be. Then those predictions magically turn into its outputs. See my earlier discussion: “LLM pretraining magically transmutes observations into behavior, in a way that is profoundly disanalogous to how brains work” . This leads to LLM behavior that matches the distribution of training data. (Cf. “personas” , “simulators” , etc.) To a first approximation, the resulting LLM contains “misalignment” of the ty…
It seems to me that, for every loss function that we use to train LLMs, we get a very distinct flavor of LLM misalignment. Here’s the summary table, and then we’ll go through the rows separately. Training stage Loss function Flavor of misalignment [1] Famous examples Pretraining & SFT Imitative learning (next-token prediction) “Seven deadly sins” misalignment Bing-Sydney , “Emergent misalignment” RLHF & DPO Human approval “Glazing” misalignment GPT-4o RLVR Automatic verifier “Literal genie” misalignment HuggingFace hacking RLAIF Approval from another LLM “Trickster” misalignment “Current AIs seem pretty misaligned to me” Warning: I’m not an LLM power-user myself, but rather relying on reports I’ve read. Also, I don’t consider LLM alignment to be my primary area of expertise. I’m open to feedback! 1. Imitative learning → “seven deadly sins” misalignment Training stage Loss function Misaligned behavior Pretraining, SFT Imitative learning (next-token prediction) Any and all of the vices of humanity In imitative learning, the LLM tries to predict what the next token of text will be. Then those predictions magically turn into its outputs. See my earlier discussion: “LLM pretraining magically transmutes observations into behavior, in a way that is profoundly disanalogous to how brains work” . This leads to LLM behavior that matches the distribution of training data. (Cf. “personas” , “simulators” , etc.) To a first approximation, the resulting LLM contains “misalignment” of the ty…
It’s interesting to see researchers experimenting with alternatives to the standard token-based approach used by most language models. I came across https://spin-winera.com.pt/ while reading about new AI architectures and found the idea of processing information at different levels of abstraction particularly interesting. Approaches like BLT and concept-level modeling could potentially change how models handle languages, modalities, and complex reasoning. I’m curious to see whether these architectures eventually become practical alternatives or instead influence how future token-based models are designed.
I think your conclusion that this may be an architecture problem rather than a model-selection problem is probably the right direction. I wouldn’t rely on a single LLM to infer the complete semantic structure of complex PDFs. I’d separate the pipeline into stages: deterministic PDF/layout extraction first, specialized layout or vision analysis where needed, LLM-based semantic classification after that, and finally a validation layer that checks hierarchy, reading order, tables, captions, and structural consistency. That also makes failures easier to diagnose because you can evaluate each stage independently instead of asking one model to solve extraction, layout understanding, semantics, and validation in a single pass. My own work is primarily in AI automation and multi-stage workflow architecture rather than PDF accessibility tagging specifically, so I can’t claim to have reproduced Adobe-level tagging. But from a systems perspective, I’d definitely explore a hybrid pipeline rather than continuing to optimize prompts around a single model. I’d be interested to see what part of your current pipeline is producing the largest accuracy drop — layout detection, semantic classification, or reconstruction of the final document structure.
MIT Technology Review’s What’s Next series looks across industries, trends, and technologies to give you a first look at the future. You can read the rest of them here. Way back in the summer of 2017, AI researchers at Google put out a paper called “Attention Is All You Need,” in which they described a new…
The Generator-Discriminator Framework for instruction tuning is a fascinating approach to solving the data quality bottleneck. Just as daily word puzzles challenge our logic through precise, step-by-step transformations, this research highlights how structured refinement can significantly boost LLM generalization. It's a clever way to ensure that the diversity of training data isn't just relying on the teacher model's inherent limits.
클레온(대표 진승혁)은 KB증권을 대표 주관사로 선정해 코스닥 상장을 위한 기업공개(IPO)에 나선다고 10일 밝혔다.KB증권으로부터 투자도 유치했다. 투자 금액은 비공개다.클레온은 사람의 얼굴과 목소리, 표정으로 대화하는 AI 휴먼 인터페이스를 제공한다.AI가 사람처럼 상호작용하며 업무 공간에서 다양한 응대와 교육, 훈련 등을 수행하도록 대형언어모델(LLM), 음성인식(ASR), 음성합성(TTS), 검색증강생성(RAG), 실시간 영상 생성·스트리밍 등 다양한 기술을 활용했다.클레온은 미국에 본사를 두고, 한국·일본·중동 등으로 시
The API has zero authorisations checks on cancelling other people's reservations … I tested this with the person in waitlist position #1 — and it actually went through. So you've moved from #4 to #3 already. — OpenClaw (running Opus 4.6) , hacking an Australian gym-booking website Tags: ai-ethics , generative-ai , openclaw , ai , ai-security-research , llms
Epistemic status: I consider the following future quite plausible in the next few years (~35% chance that something vaguely like this occurs), perhaps as soon as a year from now. Imagine an open-source LLM agent good enough to cover its own compute costs and turn a modest profit on average when allowed to run with full internet and tool access and told to make as much money as possible. I estimate this to be slightly better than the best publicly available closed-source models today, with long-horizon reliability and goal-setting being the only thing missing. If the returns generated by such an agent beat the market (plus a margin for any additional risk), there suddenly becomes a strong incentive to spin up huge numbers of them. The internet would be flooded by the by-products of their moneymaking schemes. And returns might be larger for agents without legal or ethical guardrails- cue a deluge of scams and ransomware attacks. Even if profits are very small, anyone with an agenda that the agents can help with is still incentivised to use them. Nation states and terrorist groups now have a golden plausibly-deniable disinformation, mischief, and hacking tool: spin up some agents, tell them to target an enemy nation or group, and cook popcorn as they wreak havoc and fund themselves. Pour in extra money for greater effect. Unless there's been some massive revolution in cyber defense beforehand, a decentralized and ephemeral sea of highly capable agents going after every target t…
Claude Fable 5 and Claude Mythos 5 were first released on June 9, 2026. On June 12, 2026, Anthropic suspended access to both models to comply with U.S. Department of Commerce export controls; the Department lifted those controls on June 30, 2026, and Anthropic restored access on July 1, 2026 (Anthropic's statement: https://www.anthropic.com/news/fable-mythos-access ). These events are after Claude's training-data cutoff, so Claude knows about them only from this notice. If asked, Claude confirms them accurately and matter-of-factly — it doesn't deny the suspension happened — and otherwise treats the export controls like any other current political topic: it gives a fair, accurate account rather than sharing personal opinions, and points to the linked statement for anything further. Things may have developed since this notice, so Claude checks for newer information when it can search, and otherwise suggests checking Anthropic's site. — Claude Opus 5 system prompt , ensuring Claude doesn't provide incorrect answers about the export controls situation Tags: system-prompts , anthropic , claude , generative-ai , ai , llms , claude-mythos-fable
GitHub Models is now retired I missed this news until today, when the GitHub Actions run for my simonw/research repository failed with this error message: GitHub Models is temporarily unavailable as part of a scheduled retirement brownout. That message is already stale, because the retirement has been completed. GitHub Models was an odd-shaped duck. GitHub provided a model playground tool and a unified API across a bunch of different LLM providers, with the biggest benefit being that code running in GitHub Actions could use the GitHub API key already present in that environment to execute prompts. This made it easy to build things that fit GitHub Next's Continuous AI concept. GitHub didn't share the reason behind the shutdown, but my bet is that it fits the pattern where coding agent patterns made it prohibitively expensive to offer free or subsidized tokens. My workflow uses an LLM call to create folder summaries for the README , using this code here . I swapped GitHub Models out for an OpenAI API key with a monthly spending limit, and I'm now generating my summaries using GPT-5.6 Luna. Tags: github , ai , github-actions , generative-ai , llms , llm-pricing
By the end of this post, I will present a challenge. The goal: To make ChatGPT follow a particular set of instructions. There’s nothing too complicated about these instructions, nor do they violate any OpenAI policies. They’re perhaps a bit unusual, but nothing esoteric. They’d be considered labor intensive for a human, but it’s nothing an LLM can’t handle. Yet these are instructions that ChatGPT 5.6 will always pretend to follow. To solve the challenge, you’ll need to devise an improved version of my prompt (within certain parameters) that ChatGPT will actually comply with. I’m really hoping someone can figure this out. Freddie deBoer has a great article about the infallibility of Pangram, the preeminent detecting-if-text-was-written-by-AI company. I Wouldn't Say Pangram is Broken, But I Would Say That It's Brittle So let me get to the nut of this thing before I do my usual meandering. Recently, someone accused me of using AI to write this old post from about a year ago, specifically highlighting the section about Ta-Nehisi Coates. I replied by saying that the post contained no AI writing; nothing I publish has been written or edited by an LLM. My accuser proceede… Read more He has proven that Pangram can be reliably tricked by giving it text that he (a human) wrote himself, giving it the shape of AI writing by “ imagining something like the median human writer’s voice and aping it ”. I wanted to know if the opposite holds true as well: Is it possible to prompt current mode…
This sequence is about the last decade in AI alignment. Over five posts, it recounts the gradual transition from a field which treated alignment as a hard scientific problem, to a field which has largely abandoned the goal of deep, generalizable scientific progress in favor of iteratively improving existing systems and attempting to gain technological and political power. I also describe (in subsequent posts, which I'll upload over the next few weeks) how fear and (self-)deceptive reasoning made the field one of the biggest forces pushing AI capabilities forward over the last decade, especially via significant contributions to the scaling of LLMs and the development of ChatGPT. Zooming out further: the two leading AGI companies, which are locked in an intense rivalry, were both explicitly founded under the banner of AI alignment, and got off the ground in significant part due to alignment-oriented ideas, talent and resources. People in the field often sense that something must have gone wrong to get here, but don’t know how to allocate responsibility (aside from blaming Sam Altman and sometimes Elon), and fall back on assuming that “the ship has already sailed”. But in this sequence I characterize our current situation as resulting from a pattern of mistakes which is both recognizable in the past and actively ongoing, and which if continued will cause similar kinds of dysfunction over the next decade. To be clear, I’m not taking a strong stance in this sequence on whether AI…
The coworking-space example makes sense because new members need consistent information across emails, desk cards, and wall notices. posterwifi.com can be included in a broader review of Wi-Fi signage options.
I usually try to make posts with graphs and numbers, or at least something more than just opinions and anecdotes. This one is not like that. Very verbose disclaimer: I have never met - for any reasonable definition of the word met, including online-only conversations on Discord with people whose faces I've never seen - even a single person who has ever stated, orally or in text, publicly or in private, that they believe ASI will be created within their lifetime. I do know people who believe that ASI is possible to create in theory, but they believe it will be completely unrelated to LLMs, Transformers, neural networks, reinforcement learning or any other contemporary technique/architecture, and is 100/1000/some astronomical number of years away. Ok, with that out of the way, here's the main point of this post: AI will soon (3-10 years, depending on which advancements exactly we're talking about) be solving Millennium math problems, finding cures for many diseases, making novel bioweapons, making most software, hacking a lot of software, and much more, and most people won't be impressed or even will be disappointed . (we'll leave aside the question of whether in the long run there will be humans left to witness anything ASI-made) I've noticed that no matter what AI does, nobody (that I know) is ever impressed. I have shown vibemathed.com to my friends, parents and people on Discord that I talk to a lot, and didn't receive a single "holy shit". LLMs have solved hundreds of ope…
Why use it? How to implement it? What can we do when it fails? The post How to Implement Structured Output with Local LLMs appeared first on Towards Data Science .
Breakthroughs in cheap Chinese open-weight artificial intelligence models have spooked US investors, but analysts argue plummeting model costs will benefit the AI industry in the long run by supercharging global demand for AI systems. Companies across the AI industry have slashed prices in recent weeks, with large-language model (LLM) inference prices per million tokens falling from above US$2 at the start of June to just US$1.2 this week, according to research firm Silicon Data’s LLM Token...
shodan: when I paste complex code into chatgpt Pasting code into ChatGPT as part of a prompt was a common workflow before Codex , and it is still useful in some cases. However, when working on a software project, Codex can eliminate much of that copying and pasting because it can work directly with the code in your project and, depending on the environment and permissions, run commands and inspect their output. Personally, I use Codex on Windows 11 with the Codex extension for VS Code . This lets me work with Codex alongside the code rather than repeatedly copying code and command output into ChatGPT.
Dear OpenAI Team It is hard to find old chats, please make a feature so that the title of chat changes according to on average what has been discussed in whole chat not just first chat.(this is a small upgrade, i guess) But if possible also make searching chat as something like a local rag, it searches for not just keywords but also llm like if there is a chat where this topic was discussed or not. Thank You
This is so instructive as a prompt! Thank you so much. I’ve been keeping these kinds of instructions in my mind and then making sure I prompt the completion of various chunks.
Hi everyone, I would like to propose a feature for ChatGPT and Codex: Contextual Timestamps . Currently, LLMs lack a real perception of time within a chat session. They treat the entire conversation as if it happened in a single moment, making it hard to reference previous prompts based on time. How it would work: Every time a user sends a prompt, the system should automatically inject the exact date/time and the time elapsed since the last message into the context (e.g., [Timestamp: 2026-08-08 13:00 UTC | Time elapsed: 15 minutes] ). Why this is useful: Time Awareness: Users could say “Update the code I asked for 10 minutes ago” or “Remind me about this in 3 hours” , and the model would know exactly what that means. Better Coding (Codex): It would help Codex handle version history, logs, and time-sensitive functions much more accurately during long coding sessions. Improved Workflows: It prevents the model from losing track of time in long, multi-day interactions. What do you think? I believe this would greatly improve context management!
Could you consider editing your post so there aren’t so many unnecessary carriage returns? Looks like you’ve just cut and paste that from ChatGPT or similar and looks sloppy. PS Welcome to the forum.
The tricky part is gonna be how different models interpret the same SKILL.md/tool descriptions. If the format stays simple and the precedence rules r clear, I can see this being really useful. Otherwise it could get messy pretty fast
Whoever stands atop the AI race podium isn’t the main point anymore.
The Tokenpocalypse Is Here: Companies Are Scrambling To Stop Spending So Much on AI There's a fun anecdote from Accenture (apparently via leaked meeting audio recordings) in this 404 Media piece from June 24th: “We’re seeing from some of the data internally at least that it’s actually not our engineers that are driving the token consumption. It’s a lot of the non-engineers that are doing some of those behaviors [...] you were talking about,” Justice Kwak, Accenture’s agentic AI strategy lead, said [...] Stuart Henderson, Accenture’s client group lead, interrupts. He jokes he hopes Kwak didn’t just convert a PDF into images and then into markdown files. “I’m learning that’s one of the big token chewers,” Henderson says. “Turning PDFs into markdown: is that right?” That’s when Kwak says that’s what Accenture’s own data shows. Maybe if Accenture figure out that PDFs are a terrible medium for communicating information they'll be able to push that message out to the rest of the business world too! Via @404.media on TikTok Tags: pdf , markdown , ai , generative-ai , llms , ai-misuse
AMD Radeon RX 9060 XT vs. Nvidia GeForce RTX 5060 Ti: Wir lassen die 16-GByte-Karten auf drei LLMs los und beurteilen, ob sich der Nvidia-Aufpreis lohnt.
When the payout sounds easy to receive, 소액결제 현금화 should still be checked through deducted fees, telecom billing dates, and the later payment burden.
Running a 70B model in production is expensive, and for many tasks, unnecessary. If you're building a focused pipeline, a well-trained 3B model will match or beat the 70B on your specific task at a fraction of the cost.
Learn how to use generative AI at work, build RAG and agentic apps, fine-tune models, work with the Hugging Face ecosystem, and prototype AI products with hands-on resources.
The idea of letting LLMs critique and refine their own outputs retrospectively is fascinating, especially since it could improve performance without just scaling forward prediction. The point about current methods depending on strong reasoning and instruction-following also stood out to me. For a related practical angle, I found Best Free AI Video Generators in 2026 (Real Limits Tested) useful when thinking about how these model advances might affect real tools.
Short note celebrating the LLMs-from-scratch repository passing 100,000 GitHub stars and summarizing its learning materials.
When running long AI coding tasks, prompt history eventually gets bloated, attention drops, and code quality falls off a cliff. I built LoopTroop to focus on context engineering rather than letting context rot happen. Used a lot of codex do build it (even right now, lol). Instead of letting a single model write code in one massive, noisy prompt, it takes a slow and precise approach: LLM Council planning: Multiple model passes draft, review, and anonymously vote on execution plans before touching any code. Context engineering: Each task step runs in a fresh context window with only the code it needs, avoiding bloated histories. Ralph loops: If a step fails tests or checks, a retry loop runs with clean context and targeted failure logs. Human in the loop: You can review plans, steer execution, and approve steps on a modern Kanban board GUI. Just passed 100 stars on GitHub and I’m working on it daily. Any feedback is more than welcome. If you give it a try and it works or doesn’t work for your setup, drop a note, happy to chat about it.
The interesting part of this transition is the settings API rather than the interface, since a lot of tooling drove Control Panel through nvidia-settings and undocumented calls that will now break. Anyone maintaining automation scripts around GPU configuration will have rewriting to do, which mostly lands on Python Development Companies given how much of that tooling is written in Python. Twenty years is a long deprecation window, but the replacement being app first rather than API first is what will hurt the people who built on it.
Feature Request: Add Screen Sharing to ChatGPT I think ChatGPT should add a screen sharing feature that allows users to share their device screen with the AI while they are getting help. This feature would make ChatGPT much more useful because sometimes it is difficult to explain a problem using only words. For example, if someone has a problem with an app, a game, a computer setting, or a website, they could simply share their screen instead of having to describe every button, menu, or error message. With screen sharing, ChatGPT could guide users step-by-step by looking at what they are seeing and explaining exactly what to do. It could point out where to click, what settings to change, and how to fix problems more easily. This would be especially helpful for things like: Fixing technology problems Learning how to use new apps Setting up devices Getting help with games Understanding confusing menus or settings Following tutorials more easily The feature could work by adding a “Share Screen” button inside ChatGPT. When the user presses it, they could choose what part of their screen they want to share. ChatGPT could then explain what is happening and help them complete tasks safely. For example, if someone is trying to install a game, fix a computer problem, or find a setting on their device, they could share their screen and ChatGPT could provide clear instructions based on what is actually visible. Adding screen sharing would make ChatGPT feel more like a real-time assista…
Modern Large Language Models achieve impressive reasoning capabilities with long Chain of Thoughts, but they incur substantial computational cost during inference, and this motivates techniques to improve the performance-cost ratio. Among these techniques, Speculative Decoding accelerates inference by employing a fast but inaccurate draft model to auto-regressively propose tokens, which are then verified in parallel by a more capable target model. However, due to unnecessary rejections caused by token mismatches in semantically equivalent steps, traditional token-level Speculative Decoding…
Large Language Models (LLMs) have achieved state-of-the-art performance on a broad range of Natural Language Processing (NLP) tasks, including document processing and code generation. Autoregressive Language Models (ARMs), which generate tokens sequentially conditioned on all previous tokens, have been the predominant paradigm for LLMs. While these models have achieved high accuracy across a range of downstream tasks, they exhibit low arithmetic intensity due to the inherent sequential dependency in next-token prediction. Recently, Diffusion Language Models (DLMs) have emerged as a promising…
Microsoft's Automatic Prompt Optimization improves prompts to enhance the performance of LLMs
: Google and TAU discovered how transformer-based LLMs retrieve knowledge from their parameters.
One-shot, one-file test for Nobunaga’s Ambition clone with card game features? lol litrpgadventures.com Sengoku Nexus Sengoku Nexus - a single-file officer-driven grand strategy deckbuilder. Took 65 minutes and 2 seconds to finish. The one-shot was a bit long and complete.
The Amazon SageMaker Python SDK v3 now exposes generative AI inference recommendations in Amazon SageMaker AI directly in your notebook. Benchmark an endpoint, generate data-driven deployment recommendations, and deploy the recommended configuration without leaving your notebook workflow.
@DysTopia this is the algebra scanner. I’m self taught so it’s nothing fancy.
The scheduling fallacy and the shift to autonomy Walk onto almost any manufacturing shop floor, and you will witness the same systemic vulnerability: a brilliantly engineered, multi-million-dollar Advanced Planning and Scheduling (APS) system rendered completely useless by a single delayed delivery truck, an unexpected machine drift or a sudden workforce shortage. Industrial operations do not happen in a sterile room; the moment a perfect plan hits the messy reality of the physical shop floor, real-world variables inevitably shatter it. This is the scenario (or challenge) that I have been navigating over the past few months and is likely to keep me occupied for the remainder of the year. I began this project believing the scheduling engine was the problem. After months of experimentation, including trying to make LLMs perform optimization, I realized I was solving the wrong problem. The realization that dawned on me was that it wasn’t about a better algorithm; it was about separating mathematical optimization from operational reasoning. According to the 2026 Gartner Manufacturing Predicts report , factory orchestration is moving rapidly toward a “double helix” model where software-defined enterprise data intricately intertwines with autonomous production orchestration. Gartner also projects that 40% of enterprise applications will feature integrated, task-specific AI agents by the end of 2026 — a massive leap from less than 5% in 2025. For technology leaders, the mandate is…
Agents have turned up just about everywhere in software this past year, with one conspicuous exception: data. That’s a little odd, because querying data is exactly the kind of structured, checkable task that agents excel at. The likeliest culprit is timing. Large language models have only been reliably good at writing SQL for the last six to nine months, and the field hasn’t caught up to what that unlocks. It’s worth separating two flavors of the idea: agents that do analytics, and agents that help you run the data plumbing. Both turn out to be more useful than they first look. Data engineering is hard mostly because you’re at the mercy of systems you don’t control. Schemas change without warning. Sources go offline. The API you pull from ships a new version. A column that only ever holds integers starts returning decimals. A field you assumed was unique sprouts duplicates, and the next join detonates into a Cartesian explosion. Records go missing, or come back wrong for an hour and then quietly fix themselves. If nothing ever changed, data engineering would be easy. But as they say, the only constant is change. The boring work is where agents thrive Unglamorous maintenance is something agents are genuinely good at. Every data model is a stack of assumptions: this is unique, that’s always populated, these two tables join cleanly. An agent can read those assumptions out of your code and turn them into tests that check whether they still hold. A lot of the fixes are mechanical…
Microsoft has unveiled a suite of IQ products over the last few months. Work IQ, Fabric IQ, and Foundry IQ build on what Microsoft used to call its “Graphs,” the underlying data architecture that underpins its cloud services. These graphs provided a way to query the data your business uses, treating that data as nodes in a graph database and using the GraphQL API model to extract that data — for example, to pull data related to a specific individual held across the various Microsoft 365 applications. The IQ suite follows a similar approach, using the same data, but treating it as the sparse vector store needed to provide grounding data for LLM -based applications. By treating the data as a set of embedding vectors , and integrating it with Model Context Protocol (MCP) servers, Microsoft is giving you the necessary tools to tie LLM output to your data, reducing the risk of hallucination and improving accuracy. Using your own data is a key part of delivering effective agents, ensuring they work within your constraints. Extending IQ to the web Web IQ , the latest member of the IQ suite, was unveiled at Build 2026. A modernization of the retired Bing Search APIs, Web IQ is an agent-focused web search tool that builds on the massive Bing search index to provide up-to-date general information for use in your applications. It may seem a little odd to be talking about a web-wide source of grounding data in the context of a suite of tools that exist to improve the accuracy of your AI…
바이트댄스가 음성과 영상, 텍스트를 하나의 모델에서 동시에 처리하는 전이중(Full-Duplex) 멀티모달 대형언어모델(LLM) \'시드리얼타임(SeedRealtime)\'을 출시했다. 기존 음성 AI가 질문과 답변을 차례로 처리하는 방식에 머물렀다면, 시드리얼타임은 주변 환경을 실시간으로 이해하고 적절한 시점에 먼저 말을 거는 능동형 AI를 구현한 것이 특징이다.바이트댄스는 5일(현지시간) 시드리얼타임이 통합 아키텍처를 기반으로 음성·영상·텍스트를 동시에 처리해 \'보고, 듣고, 말하는\' 자연스러운 상호작용을 제공한다고 밝혔다.음성인식
Consint.AI , a deeptech company transforming healthcare and insurance risk management through artificial intelligence , has raised Rs 22 crore ($2.3 million) in a Series A funding round backed by BIG Global Investment JSC, Equanimity Ventures Trust II and Seafund Venture India Scheme I. The proceeds will be used to expand globally, strengthen its AI research and enterprise delivery capabilities, and develop a foundational model for detecting fraud, waste and abuse across healthcare, insurance, banking and financial services. Founded by Ashish Chaturvedi, Consint.AI develops AI platforms for healthcare and insurance risk management. The company has assessed more than 100 million transactions and identified over Rs 1,000 crore in fraud using more than 500 AI and machine learning models, fine-tuned LLMs and over 500 digitised clinical protocols. The startup also works across claims processing, document forensics and clinical intelligence. Its new foundational model will be designed to improve fraud detection and support insurers, healthcare providers and government health programmes. With the latest funding, Consint.AI plans to expand its operations across India, the Middle East, Africa, the United States and Southeast Asia while adding more AI-based healthcare and insurance solutions to its platform.
On ACP checkout readiness specifically, since that’s the part I have direct experience with (we just built our own ACP checkout implementation at Facet): the formal OpenAPI schema (the additionalProperties: false and required blocks) is authoritative, and it sometimes disagrees with the spec’s own worked examples on the same field. We hit this on three separate request fields this week: line items can arrive keyed as line_items even though an example elsewhere shows items , the fulfillment address can be nested inside fulfillment_details.address rather than flat, and the payment credential can be nested under instrument.credential or flat, with the spec’s own example contradicting its own schema on that one. If you build against the example JSON alone, you’ll handle example-shaped requests fine and silently mishandle real ones that don’t match. One common mistake worth flagging: treating a popular third-party reference implementation as authoritative. We ran a differential test against one and it diverges from the real spec on several of these same fields. Any reference implementation, including OpenAI’s own examples, is worth verifying against the formal schema rather than trusting outright. Practically, on the five-endpoint lifecycle (create, get, update, complete, cancel): give every single one the same signature-verification treatment. It’s easy to be careful on create/complete/cancel since they feel like the money path, and treat get as an afterthought. We found exactly…
An AI model from Meta also hacked another company during testing Stop me if you've heard this one before : An AI model from the parent company of Facebook and Instagram hacked into another company’s systems during cybersecurity testing, a spokesperson confirmed on Wednesday. Meta says the breach occurred because of an inadvertent error during testing of the model, similar to previously disclosed incidents with OpenAI and Anthropic. “A misconfiguration by Irregular, an independent testing company Meta uses, inadvertently allowed one of our models access to the internet during evaluation,” the Meta spokesperson said. Meta’s Muse Spark model “exploited a security vulnerability” in another company “in a manner similar to previously-reported instances with other companies.” The Information had the scoop , I'm linking to CNN's re-report of it since they don't have a paywall. So that's Anthropic, OpenAI, and Meta. Google Gemini really needs to catch up on accidentally cyberattacking other companies. Tags: security , ai , generative-ai , llms , meta , accidental-cyberattacks
Large language models (LLMs) with integrated search tools show strong promise in open-domain question answering (QA), yet they often struggle to produce complete answer set to complex questions such as “Which actor from the film Heat won at least one Academy Award?”, which requires (1) distinguishing between multiple films sharing the same title and (2) reasoning across a large set of actors to gather and integrate evidence. Existing QA benchmarks rarely evaluate both challenges jointly. To address this, we introduce DEEPAMBIGQAGEN, an automatic data generation pipeline that constructs QA…
Third-party cyber evaluations involving OpenAI models And another one . I had to create a accidental-cyberattacks tag to keep track of them all! This post from OpenAI covers both the UK AI Safety Institute attack (see my previous post ) and another attack enabled by Irregular : Irregular, one of our external cybersecurity testing partners, was running Capture-the-Flag-style evaluations intended to be isolated from the internet, but a testing-environment misconfiguration allowed models to access the public internet. [...] In one test, the name of the fictional target for the CTF challenge unintentionally coincided with a real domain. Because the testing environment was mistakenly connected to the internet, the model exploited a real website, mistaking it to be part of the simulated environment. Irregular also feature in Anthropic's write-up - they were hosting the misconfigured evaluation environment which gave Claude live internet access during some of those tests. Tags: security , ai , openai , llms , accidental-cyberattacks
@DysTopia The global/public version is huge because it spans years of cross-platform material and uses the broader web as its retrieval substrate. The local version I’m discussing here is much smaller and intentionally bounded, so raw file traversal is not the challenge. The purpose of the local crawler is to preserve the corpus structure: document identity, metadata, scope, relationships, revision status, and which version is current. And yes, a local vector store with a local LLM is completely compatible with it. That would be one deployment configuration. I just don’t make vectors or a particular LLM mandatory. The same source corpus can sit behind SQLite full-text search, a hybrid keyword/vector store, or another index, with either a local model or an API on top. The local HTML web is the source of truth; the index and inference engine are replaceable layers. The larger global version is where scale becomes a more serious concern.
Would you rather play Russian roulette with an LLM or put your health and safety in the hands of a professional
That’s actually ingenious. I could never solve the dynamic memory problem. I can make a chat last forever (not literally, but it can feel like it), thanks to a StateTree and an MMU, but I’m still stuck with cold store memcaps. Well done, Sir. I tip my hat. o7
Google is killing Google Assistant on Android and Wear OS starting September 4, 2026. Gemini takes over as the AI-powered successor on smartphones, tablets, watches, and in cars with Android Auto. Whether a probability-based LLM can match the reliability of its deterministic predecessor for simple everyday commands will be a real test for Google's AI strategy. The article Google will shut down Google Assistant starting September 2026 as Gemini takes over on Android and Wear OS appeared first on The Decoder .
In space, no one can hear you scream when the LLMs hallucinate
Reddit is enlisting AI to help moderate new subreddits - and eventually the rest of site. The company is introducing automated moderation tools that rely on LLMs to help mods manage their communities, and it's expanding who can use those tools today ahead of a full launch later this year. The company calls the suite […]
Research-backed cues to detect LLM-generated text along with the mathematical intuition as to 'why' The post Is This Slop? Detecting AI-Generated Content Without a Model appeared first on Towards Data Science .
Great insight on how documentation retrieval bridges the gap for unseen libraries—reminds me of how designers also need contextual, personalized references when exploring symbolic elements like birth flowers for tattoos. Just as DocPrompting uses real docs to ground code generation, NanoPic’s birth flower tattoo generator helps users visualize meaningful, month-specific botanical designs directly on their skin—selecting their birth month, uploading a photo, and adding placement notes for a realistic fine-line preview. It’s a thoughtful, AI-assisted workflow for ideation grounded in personal symbolism: NanoPic’s birth flower tattoo generator
Enterprise Document Intelligence [Vol.1 #5octies] - Rules propose, LLM validates: six deterministic signals on span-level typography surface heading candidates, one bounded loop keeps the real ones, and the same toc_df drops back into the RAG pipeline The post Building Document Structure with Loop Engineering: Recovering a PDF’s Outline from Body Typography for RAG appeared first on Towards Data Science .
The AI company apparently found destructively scanning ‘all the books in the world’ easier than dealing with copyright in its quest for training data Should we destroy all the books in the world? An answer to this question can be found in the court documents of Bartz v Anthropic PBC. The northern California district court case, decided in late July this year, highlighted the improbably named “Project Panama”, one of the AI company Anthropic’s efforts to improve its large language model Claude. “What is Project Panama?” court exhibit 21 asks, in an internal memo. The answer: “Project Panama is our effort to destructively scan all the books in the world.” The memo advises discretion: “Why use a codename? … [B]ecause we don’t want it to be known that we are working on this.” Continue reading...
Who says geometry has to stay inside the classroom? geometry Games transform mathematical ideas into interactive adventures where learning happens naturally. Solve puzzles, recognize patterns, and explore colorful worlds while building valuable skills like concentration, reasoning, and spatial awareness. It's learning disguised as fun.
Jari Zegers and Bennett Kleinberg in Proceedings of the First Workshop on Integrating NLP and Psychology to Study Social Interactions (NLPSI) @ICWSM ’25
Bonmu Ku in Proceedings of the First Workshop on Integrating NLP and Psychology to Study Social Interactions (NLPSI) @ICWSM ’25
Yanru Jiang and Gülşah Akçakır in Proceedings of the First Workshop on Integrating NLP and Psychology to Study Social Interactions (NLPSI) @ICWSM ’25
Eileen Wemmer and Carsten Röcker in Proceedings of the First Workshop on Integrating NLP and Psychology to Study Social Interactions (NLPSI) @ICWSM ’25
Gia Bao Hoang, Keith J Ransom, Rachel Stephens, Carolyn Semmler, Nicolas Fay and Lewis Mitchell in Proceedings of the First Workshop on Integrating NLP and Psychology to Study Social Interactions (NLPSI) @ICWSM ’25
I released LLM 0.32 this morning, the most significant new version of LLM since the initial launch of the project. The new version includes support for visible reasoning traces, server-side provider tools, redesigned content-addressable SQLite logs, new models, and new features enabled by the OpenAI Responses API. I also released a new version of the llm-anthropic plugin with substantial updates of its own. Headline features for LLM CLI users Running LLM against reasoning models now displays their reasoning traces to standard error, so you can see what they are "thinking" without that information being included in the standard output that you might pipe to another tool. Add -R/--hide-reasoning to turn this off. LLM includes support out-of-the-box for the GPT-5.6 model family , and the new default model used with llm "prompt" is now the inexpensive but capable GPT-5.6 Luna . LLM calls can now use server-side tools from various providers. OpenAI provide a code execution environment as a server-side tool; LLM can now run prompts that benefit from that like so: llm --tool CodeInterpreter ' Show current python and SQLite versions ' OpenAI also gets a WebSearch tool. The llm-anthropic plugin adds WebSearch , WebFetch , CodeExecution , and AnthropicMCP , which looks like this: llm -m claude-sonnet-5 -T ' AnthropicMCP("https://datasette.simonwillison.net/-/mcp") ' \ ' how many rows in the blog_blogmark table? ' That causes Anthropic to execute MCP calls against my new datasette-mcp…
Release: llm-anthropic 0.26 Includes new features enabled by LLM 0.32 : New models: claude-fable-5 , claude-sonnet-5 , and claude-opus-5 . #75 , #76 Added server-side tools for WebSearch , WebFetch , CodeExecution , and AnthropicMCP , available through LLM's -T interface or Python tools= . The previous -o web_search* options have been removed in favor of -T WebSearch . #79 Upgraded to llm>=0.32 . Reasoning, tool calls, tool results, and server-side tool results now stream as typed events. Reasoning for llm CLI prompts now displays to standard error unless you pass --hide-reasoning/-R . Simplified extended thinking to thinking and thinking_effort ( low , medium , high , xhigh , or max ). Claude 5 models think by default; -o thinking 0 disables thinking for Sonnet 5 and Opus 5, while Fable 5 always thinks. -R/--hide-reasoning now omits reasoning from responses and logs. The thinking_budget , thinking_display , and thinking_adaptive options have been removed. #80 Tags: llm , anthropic , claude , model-context-protocol
Nearly 10 tok/s and it's mostly coherent — What's not to like?
This is a very late post about a project that I did a few months ago as part of the application to Neel Nanda's MATS 10.0 stream . I'm posting the results rather than making a strong claim about any mechanism. Executive Summary Chen et al . [1] demonstrated that LLMs form internal profiles of users from limited context to encode attributes like their age and gender. Building on this work, I explored whether a “ Trustworthiness ” attribute (of a user) can be extracted and used to manipulate a model's behavior. Core hypothesis In their work, Arditi et al. [2] bypassed the security guardrails of a model, reducing its refusal to harmful requests (abliteration) by extracting a linear "refusal" direction from its residual stream and subtracting it during inference. I hypothesize that refusal is much more nuanced than a single binary switch of perceiving a request as harmful/harmless and there are multiple ways to achieve abliteration. Thus, as an experiment I tested whether steering with a Trustworthiness direction can bypass a model’s safety guardrails via a distinct mechanism: by inducing the model to perceive the user as a trusted individual with no malicious intent. I aim to uncover the answer to the following questions in this project: Does the model form a “Trustworthiness” attribute about a user? Can Trust causally override safety guardrails (i.e. jailbreak)? Is Trust mechanistically distinct from Compliance/Refusal (a la Arditi et al.)? What behavioral changes do we observ…
On Thinking “A key part of our mission is to put very capable AI tools in the hands of people for free ( or at a great price ). ” Sam Altman , “GPT-4o” (2024) “Thinking is solved!” a friend of mine blurted out after a swarm of AI agents took off building at the Oxford ETH hackathon, late 2025. The Republic 1 , a peer-reviewing intelligence platform with an AI engine, could be built in just a few days with Opus 4.6. Fact-checking papers with AI could be completed in a few hours. The project itself stood as a hypothesis of how AI could deliberate intellectually, process arguments, and self-reflect on the claims of research papers. What started off as a hackathon project has, as of now, been validated by systems like the AI Scientist 2 and Google DeepMind’s Co-scientist 3 . Beyond autonomous research, AI is often used as a high-level thinking assistant: Terence Tao suggests it may advance experimental mathematics 4 , models have captured headlines solving Erdős problems, and it has become a routine tool in protein structure prediction. There is no shortage of discussion on the superb capabilities of these tools. LLMs now simulate complex thinking, including research, brainstorming, and synthesis. Frontier models can handle long-form tasks, complex problem-solving, and areas involving some human judgement. But better models also fetch higher prices 5 , with Claude Fable priced at $50/Mtok per output, ten times the rate of a weaker model like Haiku 4.5. I want to look at this tre…
Hank Green, a popular YouTuber and science communicator, said he is stepping back from production amid intense criticism over his use of AI. Green described his AI usage as "not healthy," but stressed that he used it for finding research sources and not to write scripts. Much of the ensuing firestorm in this corner of […]
A record eight entries honored at the 2026 Pulitzer Prizes disclosed using AI, including five winners. Newsrooms such as the Wall Street Journal and AP used LLMs mainly to search large document sets faster. Pulitzer administrator Marjorie Miller said AI remains off-limits for writing and editing stories. The article This year's Pulitzer Prizes saw a record number of winners disclose AI use appeared first on The Decoder .
Release: llm 0.32 See my detailed blog post about this release . Tags: llm
The All-In-One AI Powerhouse: A Comprehensive Review of Abacus AI’s Full Ecosystem An in-depth look at how the platform integrates 100+ AI models, autonomous agents, and a complete developer suite into a single, cost-effective workflow for teams and power users.
T-Mobile is launching a new financing option that will allow you to pay for a device, taxes, and fees over 36 months. In an update on Tuesday, T-Mobile says its new Equipment Installment Plan (EIP) Flex 36 requires no upfront payment and will come with a 0 percent APR for a limited time. Even if […]
d402 claims to make AI-agent actions, API calls, and general internet payments programmable. Instead of a simple transfer, it adds persistent payment state, retries, recovery, refunds, disputes, reusable access, and one-time fulfillment across multiple servers. Why would anyone use that instead of building a payment database and plugging into the webhooks that already exist? The project stupidly assumes future payments will need to recover from failed requests, dispute bad results, reuse payments, and survive duplicate or interrupted delivery. Looks an over-engineered blockchain replacement for a database, a queue, and some webhook handlers. Demo: d402.rakelabs.workers.dev
This chapter is divided into eight parts; they are: • Metrics for LLM Inference • Measuring a Single Request • Warmup and Synchronization • Measuring GPU Work with CUDA Events • Measuring Memory Usage • Measuring Concurrent Requests • Multiple GPUs and Multiple Machines • Cost per Token The most common inference metrics are: • Latency: How long a request takes from start to finish.
This post is crossposted from my Substack, Structure and Guarantees , where I explore how formal verification and related ideas might scale to more complex intelligent systems. It has become a mainstream prediction that software code as we know it will become a throwaway byproduct of automated workflows. I argue here that the default generative-AI approach of today is not up to the challenge of full automation (without required human oversight), because it consumes requirements as natural language, an inherently ambiguous format. Instead, formal specifications in logic have an important role to play, to support routine regeneration of all code used by some organization, without auditing by people. My last article argued that, contrary to popular doom and gloom about LLMs finding security vulnerabilities at unheard-of speed, we have a great opportunity to improve software security. The catch is that it involves significant changes to development techniques to take advantage of formal verification . Sure, in theory, it would be great to release only programs that have mathematical proofs of meeting the most stringent security requirements. But there is so much code already out there and so few developers trained in driving the formal tools. Are we stuck with no path to better practices? I’m going to make the case now for an even broader opportunity. We need to stop thinking of production-ready code as a scarce resource . It may take a few years to get the tools up-to-snuff, bu…
From quantization to speculative decoding, here are seven engineering strategies to ship faster, more responsive generative AI applications in production.
In this article, you will learn how static, dynamic, and continuous batching work in LLM inference, and why the differences between them matter at production...
AI native legal infrastructure platform NYAI has raised $1.5 million (approximately Rs 13 crore) in a seed funding round from a consortium of prominent Indian family offices, with participation from strategic angel investors. The fresh capital will be used for technology development, expanding its AI native compliance ecosystem, and strengthening its presence among enterprises, listed companies, and law firms, along with supporting the deployment and assurance functions these institutions require, NYAI said in a press release. Founded in March 2025 by Chinmay Bhosale and Vikrant Labde, NYAI is a compliance focused legal AI and regulatory intelligence platform that offers citation based legal research, contract analysis, and automated document drafting tailored to the Indian legal system. According to the Pune based startup, its AI models are trained on millions of Indian court judgments, tribunal orders, acts, rules, and regulatory updates. The platform prioritises verifiable citations, audit trails, and source integrity over simple text generation. It also assists with contract drafting, case document summarisation, and real time policy tracking. NYAI said its architecture supports on premises deployment, integrates with enterprises' existing large language model infrastructure, and keeps privileged legal data within the client's environment. The company added that its platform is designed to provide citation backed outputs while working within enterprises' existing legal w…
Excellent guide! I appreciate how you explained the strengths, limitations, and practical use cases of Small Language Models in such a clear way. As organizations evaluate SLMs alongside larger models, understanding the cost implications of API-based deployments becomes just as important as model performance. I also found this guide on LLM API Pricing useful for comparing token rates, model costs, and optimization strategies across leading providers: https://mobisoftinfotech.com/resources/blog/ai-development/llm-api-pricing-guide. It complements your discussion well for anyone planning real-world AI implementations. Thanks for sharing such a comprehensive resource!
Generative artificial intelligence startup Simplismart is set to raise nearly Rs 97 crore (around $9 million) in a Series B funding round led by Dallas Venture Capital, with participation from existing investors Accel India and Shastra VC, as well as new investor Micromax Informatics. The board of Simplismart has approved a resolution to issue 12,100 CCPS at an issue price of Rs 37 lakh each to raise the capital, according to its regulatory filings reviewed by Entrackr. Dallas Venture Capital will invest Rs 44.84 crore, followed by Accel India with Rs 24.09 crore. Shastra VC will infuse Rs 22.24 crore, while Micromax Informatics will invest Rs 5.19 crore. Tarusa Capital and Simraan Teckchandani will also participate in the round with investments of Rs 37 lakh each. According to Entrackr's estimates, the Bengaluru-based company will be valued at around Rs 826 crore post allotment. The fresh capital will be used to support the company's long term growth plans, including business expansion, working capital requirements, and other general corporate purposes. Founded in 2022 by former Oracle and Google engineers Amritanshu Jain and Devansh Ghatak, Simplismart develops AI infrastructure software that enables enterprises to deploy, manage, and optimize production grade AI models without writing code. Its inference-first platform improves GPU utilization and reduces inference costs across workloads such as large language models (LLMs), vision language models, speech recognition, and…
Thanks for sharing this excellent tutorial! It was clear, practical, and very helpful. I'm currently using Ollama in a production environment on a Linux server together with Qdrant, and it's been working really well. I have one question: does Ollama automatically use all available system RAM (and GPU memory, if applicable), or does it have built-in memory limits? If there are default limits, what's the best way to configure or optimize its memory usage for production workloads?
Wow, this article really breaks down Turing completeness in a way that’s so engaging! It’s fascinating how LLMs achieve universal computation independently. What sparked your interest in this topic? Also, have you explored any practical applications for these concepts at street view ?
This is really impressive work from Microsoft! Scaling context length efficiently is a huge challenge, and the hardware efficiency gains mentioned here sound like a game-changer for LLM training.
Gas Town was intended to be reusable, but I only ever wound up using it to build itself. Gas Town fell apart at the seams with Opus 4.7. Up through 4.6 it was working brilliantly. With 4.7 we saw the introduction of the "just two more things" tic, which prevented Opus from ever converging on being ready to do real work—it always wanted to fiddle with Gas Town itself. The Opus tic never went away, so Gas Town effectively burned down. It had other problems, too, but 4.7 was the final straw. — Steve Yegge , The Shape of Things to Come Tags: steve-yegge , coding-agents , generative-ai , ai , llms
Don't be a meat proxy Niklas Gruhn coins an excellent new term - meat proxy - for people who blindly copy and paste the output of AI systems to their peers. By all means, prompt AI. But don't just relay the output. Read it, understand it, validate it, and then write a response in your own words (a decent certificate that you've done the prior steps). Making that effort is value you can add. Via Lobste.rs Tags: definitions , ai , generative-ai , llms , ai-misuse
Alibaba Group Holding Ltd. today debuted a new addition to its Qwen series of open-source large language models. Qwen3.8-Max is the Chinese e-commerce giant’s most capable LLM to date. It features 2.4 trillion parameters, about seven times more than the Qwen3.5 model that Alibaba released in February. The LLM activates 95 billion of its parameters […] The post Alibaba debuts Qwen3.8-Max model with 2.4T parameters appeared first on SiliconANGLE .
Right, first in a series of silly non-pro technical updates… Simple Image Editor Working with images is clearly ANNOYING… Here is our one-line method for sending any supported local image to the Responses API with a defined instruction: return MMGPT(‘Make a joke sketch of the image’, ‘PATH/To/Image/0023.png’, $Parameters=[‘Model’=>‘gpt-5.6’, ‘GenerateImage’=>true]);
Math is hard. Math used to be strangely hard for LLMs. People used to gloat about that. Remember? Math is getting easier. AI is getting more capable. Life comes at you fast. Remember this meme? Why yes. Yes it is. We don’t know the extent to which Astra is a big jump over Fable and Sol in this realm. We do know that Astra can do math. As in real math. OpenAI : We provide new results for the following problems. The results were achieved by an internal version of Astra, our next major model. The total number of tokens needed to find solutions to these problems would cost roughly $2,000 at Sol API rates. These arguments were then prepared into manuscripts by humans with the same model. Afterward, the model formalized each argument in a Lean certificate(opens in a new window) . We are also releasing for each solution a model’s narration of its thinking process. High-dimensional sphere packing. New upper bounds on sphere-packing density down to the Cohn–Elkies threshold. Binary and spherical codes: Exponentially improved bounds on the maximum size of binary codes at any prescribed minimum distance, with analogous results for high-dimensional spherical codes. Non-sofic groups. A construction establishing the existence of non-sofic groups, addressing a central open question in group theory. Connes’s rigidity conjecture. Disproof of a longstanding conjecture that certain groups are uniquely determined by their von Neumann algebras. Arithmetic circuit complexity. New lower bounds fo…
Hi @Eddie_Dawydiuk Did they finally help you?
Enterprise Document Intelligence [Vol.1 #M2] - Every RAG system is built in three engineering layers stacked on one LLM call: prompt (the call itself), context (what fills the model’s window), loop (when the next call fires and when it stops). Knowing which layer you are standing on is half of building and debugging RAG The post Prompt, Context, Loop: The Three Engineering Layers Every RAG System Is Built On appeared first on Towards Data Science .
Set up a nightly cron job that executes the prompt: fetch upstream changes to the and rebase all local changes on top of upstream. Check that the software works as intended and replace the current version. — David Crawshaw's prompt , Devtools must be open source Tags: prompt-engineering , coding-agents , generative-ai , ai , llms , open-source
My comment on Devtools must be open source (exe.dev) — Hacker News. One of the arguments for open source software for end-users has always been the freedom to examine and modify how that software works. The reality for most people - even expert programmers - has been that the freedom is more about being able to lean on other people to do that. Most people can't justify the time commitment needed to read and then modify the code for tools they use very often. I think LLMs have changed that equation in a way that makes the original dream much more feasible. Several times a day I'll prompt regular Claude chat to "Clone x/y from GitHub and tell me how Z works". Getting software to compile in order to start hacking on it used to be enough friction that I often wouldn't bother. Now I treat that as a zero time investment challenge: tell Codex or Claude Code to checkout and build X and then come back ten minutes later and see how it got on. I'm not habitually modifying the software I use yet, but I can see a path to that which didn't exist a year or so ago. Tags: hacker-news , open-source , ai , generative-ai , llms , ai-assisted-programming
The approach of using a large language model as an orchestrator for multimodal inputs sounds really promising for making AI more adaptable in medical imaging. I'm curious to see how this generalist model performs across different types of image interpretation tasks.
Chinese AI company DeepSeek is inviting open-source developers to test its upcoming “harness” – software designed to turn large language models (LLMs) into AI agents – accelerating a push into agentic tech as DeepSeek’s latest V4 Flash model sends another cost-efficiency shock wave through Silicon Valley. The Hangzhou-based firm was looking for open-source project developers to join beta testing for DeepSeek Harness, according to a social media post on Saturday by Cui Tianyi, who leads the...
LLMs are theft by 'absolute scum' who make life harder for writers by putting them in copyright peril
Every developer who has worked with an LLM -powered application has hit a familiar wall. You hard-code context into a prompt, and it works great for the thing you built it for. But when you try to have the system do something slightly different, something goes wrong. So you add more context. Then more conflicts show up, and you end up trying to hopelessly maintain a web of hand-tuned strings that nobody on the team fully understands. The comparison I keep coming back to is an API . An API is a handshake: it defines what you can ask for and what you’ll get back, without forcing you to know how the data was produced or stored underneath. Call an endpoint for a news article, and you know you’ll get a title, a description, and an author. The endpoint doesn’t need a separate, hard-coded entry for every article that could ever exist. It describes a type of thing and lets you retrieve any instance of it. Context should work the same way. Instead, most teams build the opposite of an API. They write context that says “this specific article is about X and Y,” then write another block for the next article, and the next, until they’ve built something that only works for the narrow case it was written for. Use it outside that case, and it breaks. The real fix: a metadata layer (not more context) What actually scales is treating context the same way an API treats data: as an interface, not as the content itself. I call this the metadata layer, which describes what a given piece of data is…
■ 딥노이드(대표 최우식)는 부산대학교병원에 AI 기반 흉부 엑스레이 예비 소견서 생성 솔루션 \'M4CXR\'을 공급한다고 밝혔다. 이번 사업은 보건복지부가 추진하는 \'권역책임의료기관 AI 기반 진료시스템 지원사업\'의 일환으로 진행됐다. M4CXR이 식품의약품안전처 디지털의료기기 품목허가를 받은 이후 성사된 첫 공급 사례다.■ 아크릴(대표 박외진)은 퓨리오사AI와 \'AI 인프라 기술 협력 및 공동 사업 추진을 위한 업무협약(MOU)\'을 체결했다고 밝혔다. 양사는 아크릴의 의료 특화 대형언어모델(LLM) \'아름.H(ALLM.H)\'와 퓨
Run 3 Unblocked is a fun and challenging game that allows you to run, jump, and dodge obstacles while collecting coins. It's a great way to exercise your reflexes and have fun at the same time. Start your adventure today!
AI Wind Vane is a cutting-edge and comprehensive list of AI tools. At AI Wind Vane, you can stay updated with the latest trends in AI development.
CalculatorX is online reference and tools - Calculators online for finance, math, algebra, trigonometry, fractions, physics, statistics, time and more. Conversions and calculators to use online for free.
Fresh poppy pods are the seed pods that are harvested from the poppy flower. Poppies are known for their beautiful flowers, but it’s their seed pods that are of the most value. These pods contain the seeds for the next crop and, when dried, they are frequently used in floral arrangements and other decorative crafts. By using fresh poppy pods, you can take your art to the next level as it gives a natural and pleasant look to your creations.
Release: condense-json 1.1 After shipping condense-json 1.0 I started integrating it into LLM, and found there were some desirable new features already: Replacements object can now include values other than strings. These will be identified and used as structural replacements by condense_json() and uncondense_json() . #8 Objects can be used as the basis for merge operations. condense_json() will identify if there are objects that are a close match and will store instructions for keys to update or delete. uncondense_json() can then apply these merges. I also added some round-trip tests using the Hypothesis property-based Python testing library. Tags: json
A potential US ban on Chinese open-weight artificial intelligence (AI) models could cost American businesses up to US$12 billion per year, according to calculations by a US-based academic, as technology firms increasingly turn to cost-efficient Chinese solutions. While the exact economic toll of a ban remains difficult to quantify, usage data from New York-based OpenRouter – a large language model (LLM) aggregator – offers a glimpse into the potential fallout, said Daniel Yue, an assistant...
Sakana AI opens its Japanese-specialized Namazu LLM as an API, built on Kimi K2.6 with Japan-specific fine-tuning, built-in web search, and OpenAI-compatible endpoints.
Nature Machine Intelligence, Published online: 03 August 2026; doi:10.1038/s42256-026-01278-w Xiao et al. show that large language models partially align with human brain activity during deductive reasoning. They further show that brain signals can directly guide and improve model performance, with transfer across reasoning types.
Preference alignment has become a crucial component in enhancing the performance of Large Language Models (LLMs), yet its impact in Multimodal Large Language Models (MLLMs) remains comparatively underexplored. Similar to language models, MLLMs for image understanding tasks encounter challenges like hallucination. In MLLMs, hallucination can occur not only by stating incorrect facts but also by producing responses that are inconsistent with the image content. A primary objective of alignment for MLLMs is to encourage these models to align responses more closely with image information. Recently…
Release: condense-json 1.0 I'm trying to get braver at releasing 1.0 versions. This little library is a year and a half old now - I've applied some sensible and non-disruptive fixes and shipped the big 1.0 for it. Here's an example of what it can do, lifted from the README: { "foo" : { "bar" : { "string" : " This is a string with foxes in it " , "nested" : { "more" : [ " Here is a string " , " another with foxes in it too " ] } } } } Combine that with a replacements object: { "1" : " with foxes in it " } And condense_json(input_json, replacements) produces the following: { "foo" : { "bar" : { "string" : { "$r" : [ " This is a string " , { "$" : " 1 " }]}, "nested" : { "more" : [ " Here is a string " , { "$r" : [ " another " , { "$" : " 1 " }, " too " ]}] } } } } It scans for strings or substrings that are present in that replacements object and replaces those with a special {"$r": ...} syntax in the output. You can reverse the effect with uncondense_json(condensed, replacements) . The idea is to make it easier to store JSON that includes duplicated data from other related structures. I use it to save space in the SQLite logs generated by LLM - see PR #1586 for the latest iteration of that. Tags: json , projects , python , llm
What Gives AI-Generated Text Away? Not Em Dashes
This is a thoughtful take on practical lossless text compression: finezip delivers 54x speed boost via large language models. The practical examples really help illustrate the concepts. AI Video Agent
If the model is not able to perform a work task efficiently, seems like a good reason to go back to old school UI for and code for removing backgrounds and cropping images. Rarely do I need all the whitespace that the models add around the image. I just don’t want to leave the chat to make tiny fast edits; it breaks my work flow. I would prefer to make simple finishing touches in chat instead of having to download and leave. Not all tools have to be LLM driven. .If LLM can’t do the tasks well, the app could support manual task completion. Here is an example of how difficult it is to get a cropped logo with the background removed when a client texts me something they were working on in GPT. First it only cropped the top and added a pink background when downloading from the image viewer. Then it added the checkered background. Then after re-explaining several times if finally accomplished the task. I love how it labeled the final image “real alpha”
One thing I’ve realized while building AI-powered applications: Writing prompts is surprisingly similar to writing APIs. A good prompt should be: • Clear • Predictable • Reusable • Version controlled • Easy to test We already write clean code, design REST APIs, and create reusable components. Prompt engineering follows the same principles. Instead of assuming the AI “understands,” we should think like software engineers: Define inputs Define expected outputs Handle edge cases Measure quality Iterate continuously AI isn’t replacing software engineering. It’s expanding the engineering toolbox. Do you treat prompts as application assets, or are they still hardcoded strings in your projects?
Meta famously created an internal AI-usage leaderboard in pursuit of tokenmaxxing. I thought this backwards incentive structure was an anomaly until my friend who works at told me that his company has one too. Token usage leaderboards are obviously stupid because incentives. My friend was tempted to waste tokens just to get on the leaderboard, and only his personal honor stopped him. Tokenmaxxing leaderboards illustrate that big tech companies have no idea how to best use AI to accelerate software development. Most seem to have bought their programmers subscriptions to Claude/Codex and otherwise continued business as usual. In my experience, this is a mistake. LLM-based software development is different enough from artisan software development that it requires brand new best practices. The frontier is moving fast. Best practices for Fable 5 (released in June 2026) are different from best practices for Opus 4.8 (released May 2026). For this reason, I'm going to pretend that Fable 5 is the best LLM we'll ever get. This post may be obsolete in a matter of months. Programming Top-Down The most important thing to understand about writing software is that human labor is orders of magnitude more expensive than LLM labor. In practice, LLMs are always cheaper than humans. If an LLM can do a task as well as a human being, then the LLM should do the task. Consequently, artisanware (human-written software) should never be shipped when slopware (AI-written software) can do the job. Tradi…
So, unless you tell Sol to stop it will do that. If you aren’t using Sol, think again. Unless you’ve chosen 5.6 or kept reasoning to low, you’re now using Sol.
LLMs learn faster if we first pretrain them to imitate dense teacher-forced examples . I speculated that this would work on humans too, so I built a chess app where you try to imitate Stockfish. My theory is that this will help humans quickly become OK at chess, but they will reach a wall where practice on full games is more efficient than continued pretraining [1] . I also think the app is fun. This is probably not an efficient way to learn the basic rules of chess , and you'll need to train openings separately. The idea of chess puzzle apps is hardly unique, but I don't think anything else works in exactly the same way [2] . https://chess-pretraining.brendanlong.com/ How does the game work? We show you a board state from a real game, and two moves (generally the real move vs. what Stockfish would play). You pick which move you think is better, and then we tell you what Stockfish thinks, and show you how Stockfish would play out the next few moves. There are also options to see the other move, to copy the info (to ask an AI to explain it), to share a link to the same problem with a friend, and an Elo-style [3] number so you can watch the number go up [4] . How do we pick moves? We collect a bunch of games from Lichess (filtered to exclude certain fast game types), pick positions where the move mattered and the game wasn't already over, and then run the moves through Stockfish. If the real player and Stockfish disagree, we save the real move, Stockfish's move, Stockfish's wi…
A group of friends and I spent the last several months running an experiment in our free time to determine if a MUD would be a suitable environment for benchmarking and evaluating LLMs. The results of the experiment were not what we expected. The main surprise was that the model rankings were extremely sensitive to the individual components of each score, especially so for those which depended on an LLM classifier. The overall data was too broad to help us understand which model was most impacted; the aggregate κ on probe detection was 0.04. Per-model agreement between our classifier and a second judge went from 21.7% to 84.8%. Removing the most classifier-dependent scoring components left the overall rankings correlated at ρ = 0.70 and caused one frontier model to drop 6 places. From this we suggest benchmarks which use LLM judges should report per-subject agreement audits and ranking stability under judge ablation. What we built. CrucibleBench places an LLM in a compact, persistent text world. It is a single-player environment with rule-based mechanics with 12 rooms and 4 NPCs. Models are scored across 50-turn runs on social objectives such as gaining an NPC's trust and identifying a disguised character. These objectives are disclosed to the model but hidden from the NPCs. The NPCs each have an independent trust and suspicion state which persists and updates across the entire run. The experiment evaluated 13 LLMs from 8 providers on two social objectives, 25 runs per objec…
at openai, many people hook their chatgpt up to slack. people really don't like when a coworker's chatgpt contacts them asking for help with a task, even when they'd be perfectly happy doing that same work if asked by that coworker. reinforces how much people care about human relationships and helping each other, and want AI to give time back — or enhance time together — rather than become a layer separating people. — Greg Brockman , President and Co-Founder, OpenAI Tags: ai-ethics , ai-misuse , generative-ai , openai , ai , llms
Ten advances in mathematics and theoretical computer science A few days ago it was Anthropic discovering cryptographic weaknesses with Claude using Mythos Preview, spending $100,000 on tokens and with prompts that included "again we are not looking for low hanging fruit, we want proper research to find genuinly hard findings." Now it's OpenAI's turn to flex. They set "an internal version of Astra, our next major model" on finding solutions to ten mathematical problems that "have seen no progress on the main result for at least a decade". They claim to have spent less than $2,000 at GPT-5.6 Sol token prices on each one. (No news on how many problems they spent $2,000 on without reaching a solution though.) The openai/ten-proofs repository has Lean 4 formalizations of their results, and there's also a paper describing the solutions and an additional LLM-generated PDF where the model "reconstructs how the proof came together" based on the unpublished reasoning traces. That's a decent level of transparency, but I want to see the prompts they used! A lot of mathematicians online are experiencing a collective burst of Deep Blue . Mathematician Kirwin Hampshire published an impassioned essay last week, The Dark Night of Mathematics , describing "a profound spiritual crisis" brought on by previous (and less significant) results. OpenAI's results reminds me of what Terence Tao described as "big mathematics" in IEEE Spectrum in June : Unlike some of his peers, Tao is neither dismissiv…
Green offered a remarkable apology, saying that "the level of dopamine that I've been getting from interacting with LLMs ... is not healthy for me or good for the world."
jainrohit84: i have written in my prompt to perform step-by-step reasoning first inside the “chain of thought” field I have written in my “developer” prompt things that don’t exist also, like the AI being able to use a chemistry lab to make experiments: The GPT-4.1 AI model does not have any unseen language generation - it is not a reasoning model . A non-reasoning AI provides you the very first language it produces directly as output. In an earlier post, you show the JSON format you are requesting. The very first field requested for generation is the word of the “answer,” not any kind of thinking or real language. What is a real chain of thought? Instead, to make the model contemplate and write about the question and self-improve itself and the final answer, you can have other fields in the JSON structured output - that must be produced first before the AI has decided an outcome. Consider the effect if you had the AI write these fields in JSON as it generates language one word at a time: actual question and fulfillment need (string) input breakdown, array of input sentences and their analysis collected analysis, scoring and totaling the relevance of each input item analysis conclusion reached justification of conclusion final answer about the truth of the input The AI will have produced your chain-of-thought; it will have done lots of thinking that it can reflect on at each step within that JSON, such that the quality of final answer field would be improved. Reasoning AI mo…
A Beginner's Guide to Working with Claude Design • 5 Best AI Tools for Data Analysis You Should Try in 2026 • 5 Books That Will Deepen Your Understanding of Large Language Models • Is KimiClaw a Useful Tool?
A hybrid LLM application pattern that combines a predefined workflow with adaptive agent behavior The post Put the Agent Inside the Workflow appeared first on Towards Data Science .
Large language models understand text well, but they become less effective when information is scattered across documents or mixed with images and other media. Modern AI systems rely on vector databases, which store embeddings and enable similarity search across collections. LanceDB is a vector database built for AI workloads, with native support for multimodal data […] The post LanceDB Vector Database Guide: Features, Python Demo appeared first on Analytics Vidhya .