AI/ML news, top picks, and generated innovation digests.
39256 matching items
EMO is a new mixture-of-experts model trained so modular expert groups emerge from data, enabling users to select small task-specific expert subsets while preserving near full-model performance.
We reviewed the “Risks from automated R&D” section of Anthropic’s February 2026 Risk Report , producing two corresponding review documents: our original review and our updated review . We recommend that readers refer to our original review, which represents our review of the report as originally received. 1 The following is the executive summary of our original review. The full documents are available as PDFs ( original , updated ). Executive summary This document is METR’s external review of the “Risks from automated R&D” section in the Anthropic Risk Report: February 2026 (henceforth ‘the report’), which makes the argument that catastrophic risk from Claude Opus 4.6 or a less capable Anthropic model automating R&D in any domain is very low. Anthropic shared additional non-public materials with us for our review, and we used some non-public information shared as part of a previous review . We further detail this process in an appendix. We lay out our findings in two sections: Synopsis of Anthropic’s case . Our assessment : We do not think the report adequately supports its conclusion. We note significant issues in a few key areas: Analytical rigor: We have a number of significant issues with the analytical rigor in the overall argument and interpretation of the results of the model use survey. We think that the cited results of the survey provide little evidence about the level of overall risk , due to issues including sample size, question granularity, survey framing, and…
Summary: We describe three different definitions of the productivity impact of AI (AKA uplift), and show there’s reason to expect: \[\text{uplift on old tasks} \leq \text{uplift in value} \leq \text{uplift on new tasks}\] Three Measures of Uplift One complication in measuring AI’s effect on productivity is that it has different effects on different tasks, and this causes people to change how they allocate their time between tasks. This makes it more difficult to talk about the effect of AI on overall productivity. We use “old tasks” to mean the set of tasks you’d do in a typical day before AI is available – your average workday in 2021, say. “New tasks” means the set of tasks you’d do in a typical day after AI is available. Not all new tasks necessarily use AI; they’re just the tasks you choose knowing AI is an option. We have found it important to distinguish between three measures of AI’s uplift: Uplift on old tasks: The factor by which pre-AI time exceeds post-AI time to complete the old tasks. Uplift on new tasks: The factor by which pre-AI time exceeds post-AI time to complete the new tasks. Uplift in value: The factor by which post-AI value exceeds pre-AI value, allowing for reshuffling of tasks between the pre-AI and post-AI cases. In some cases value has a natural definition; in others, it can be operationalized using related definitions discussed more in the accompanying note. This note discusses the distinction and its implications for interpreting AI productivity…
An interview with Takayuki Osa, Team Director of the Robot Learning Team, was published in Nikkei Tech Foresight on May 7, 2026. The article introduces recent advances in robot
Data from 1,281 agent runs across 40+ large open source repos reveals five repeatable failure patterns in coding agents, and the infrastructure fixes for each.
Data from 1,281 agent runs across 40+ large open source repos reveals five repeatable failure patterns in coding agents, and the infrastructure fixes for each.
At Apple, we believe privacy is a fundamental human right. As AI capabilities increase and become more integrated into people’s daily lives, advancing research in privacy-preserving techniques is increasingly important to ensure privacy is protected while users enjoy innovative AI experiences. Apple’s fundamental research has consistently pushed the state-of-the-art in this domain, and earlier this year, we hosted the Workshop on Privacy-Preserving Machine Learning & AI. This two-day event brought together Apple researchers and members of the broader research community to discuss the…
Current critic-less RLHF methods aggregate multi-objective rewards via an arithmetic mean, leaving them vulnerable to constraint neglect: high-magnitude success in one objective can numerically offset critical failures in others (e.g., safety or formatting), masking low-performing “bottleneck” rewards vital for reliable multi-objective alignment. We propose Reward-Variance Policy Optimization (RVPO), a risk-sensitive framework that penalizes inter-reward variance during advantage aggregation, shifting the objective from “maximize sum” to “maximize consistency.” We show via Taylor expansion…
We propose HeadsUp, a scalable feed-forward method for reconstructing high-quality 3D Gaussian heads from large-scale multi-camera setups. Our method employs an efficient encoder-decoder architecture that compresses input views into a compact latent representation. This latent representation is then decoded into a set of UV-parameterized 3D Gaussians anchored to a neutral head template. This UV representation decouples the number of 3D Gaussians from the number and resolution of input images, enabling training with many high-resolution input views. We train and evaluate our model on an…
In this episode, Scott Clark, co-founder and CEO of Distributional, joins us to explore how teams can reliably operate and improve complex LLM systems and agents in production. Scott introduces a Maslow’s hierarchy of observability: telemetry for logging, monitoring for known signals, and post-production or online analytics to surface unknown unknowns. We dig into examples of real-world failures Scott’s team has seen in production systems, such as “lazy” tool-use hallucinations that standard evals miss, and how mapping traces into vector fingerprints enables clustering and topic discovery to uncover emergent behaviors. Scott explains how analytics can feed the data flywheel by generating evals, guardrails, and training data, and why online, adaptive approaches are essential for non-stationary models. We also touch on practical how-to’s such as instrumentation with OpenTelemetry, the GenAI semantic conventions, and the role of dedicated analytics tools. The complete show notes for this episode can be found at https://twimlai.com/go/767.
A knowledge engine is the data infrastructure category that lets agents query trusted, compiled knowledge instead of brute-forcing retrieval over raw data. How one is built, how agents query it, and how it compares to RAG, vector databases, and semantic layers.
CSET’s Julie George shared her expert perspective in an op-ed published by Bulletin of the Atomic Scientists. In the piece, she argues that while the Defense Department’s decision to narrow its list of critical technologies is a positive step, the Pentagon must also improve how it prioritizes and funds emerging technologies to address overlooked capability gaps and strengthen long-term military innovation. The post Beyond AI: What the Pentagon is missing with its trimmed ‘critical technologies’ list appeared first on Center for Security and Emerging Technology .
Read our translation of a Chinese government plan that calls for making data more plentiful and accessible in industries such as manufacturing, agriculture, transportation, finance, scientific research, and healthcare. The post Three-Year Action Plan for “Data Factor of Production ×” appeared first on Center for Security and Emerging Technology .
Dynamic Resource Allocation (DRA) has fundamentally changed how platform administrators handle hardware accelerators and specialized resources in Kubernetes. In the v1.36 release, DRA continues to mature, bringing a wave of feature graduations, critical usability improvements, and new capabilities that extend the flexibility of DRA to native resources like memory and CPU, and support for ResourceClaims in PodGroups. Driver availability continues to expand. Beyond specialized compute accelerators, the ecosystem includes support for networking and other hardware types, reflecting a move toward a more robust, hardware-agnostic infrastructure. Whether you are managing massive fleets of GPUs, need better handling of failures, or simply looking for better ways to define resource fallback options, the upgrades to DRA in 1.36 have something for you. Let's dive into the new features and graduations! Feature graduations The community has been hard at work stabilizing core DRA concepts. In Kubernetes 1.36, several highly anticipated features have graduated to Beta and Stable. Prioritized list (stable) Hardware heterogeneity is a reality in most clusters. With the Prioritized list feature, you can confidently define fallback preferences when requesting devices. Instead of hardcoding a request for a specific device model, you can specify an ordered list of preferences (e.g., "Give me an H100, but if none are available, fall back to an A100"). The scheduler will evaluate these requests in…
Lessons from my trip to talk to most of the leading AI labs in China.
The post Inside Porsche Cup Brasil’s AI-powered race operations appeared first on Source .
Today, we announced at .local London that MongoDB 8.3 is built for the speed AI demands—and our customers can't afford to wait. The data layer has to move at AI speed The old contract between databases and the applications on top of them was simple: databases improve slowly, and architectures evolve around them. AI has changed that contract. The workloads our customers are shipping today—agents retrieving at sub-100ms, retry storms hitting in milliseconds, multi-region deployments that can't trade compliance for latency—were edge cases 18 months ago. Now they're the baseline. MongoDB 8.3, generally available today, is our fourth significant release in 19 months. These releases compound. Customers running on 8.0 have seen 36% faster reads and 59% higher throughput for updates. 8.3 adds another 35% to write throughput, 45% to reads, and 15% to ACID transactions over 8.0 — without changing a line of application code. Enterprises like Adobe, running the most demanding AI in production, have made the requirements clear: sub-100ms retrieval, sub-second context updates, zero downtime. That's what MongoDB Atlas is built for. That's the commitment: when the data platform keeps pace, our customers can focus on shipping. MongoDB.local London Core Blog 2026 - Image 1 media Run anywhere. Stay secure. Where you run your agents isn't just an infrastructure decision anymore. Now, it's a critical compliance and security decision as well. While most platforms force a trade-off between global…
The post Submission on amendments to the Information Technology (Intermediary Guidelines and Digital Media Ethics Code) Rules, 2021 (“IT Rules, 2021”) appeared first on Access Now .
How MLCommons engineered a stable, accessible Mixture-of-Experts (MoE) pretraining benchmark for MLPerf Training v6.0 that runs on a single 8-GPU node. The post GPT-OSS 20B: A Sparse MoE Pretraining Benchmark for MLPerf Training v6.0 appeared first on MLCommons .
Free ChatGPT got instantly better.
China resumes fuel exports + US oil sanctions + Emission quotas for local cadres c.groth Thu, 05/07/2026 - 13:14 picture alliance / CFOTO | CFOTO Download (pdf - 546.11 KB) MERICS Briefs MERICS China Essentials May 07, 2026 10 min read China resumes fuel exports + US oil sanctions + Emission quotas for local cadres Top Story China resumes fuel exports as national supply worries ebb – and regional ones rise China is moving to prevent the worst for Asian economies by resuming exports of jet and motor fuels to some regional countries in May. Having suspended shipments from refineries shortly after the US and Israel attacked Iran at the end of February, China will allow 500,000 metric tons of fuel to be exported this month. This is still much lower than its pre-war average of more than double that amount, but a sign that Beijing’s persistent caution about its own energy supply is ebbing – and that its worries about compounding pressure on regional supply chains and markets are increasing. China is a major importer of oil and gas, but a major exporter of fuel, with its many refineries providing gasoline, diesel, and jet fuel to countries from nearby Vietnam to far away Australia. Asian economies were deeply disrupted by the energy shock triggered by the closure of the Strait of Hormuz, a critical global energy artery between Iran and Oman, and were hit again when Beijing stopped shipments of its refined petroleum products. China’s partial reversal should help ease the fuel crunch…
In this fully connected episode, Dan and Chris break down one of the biggest questions in AI today: do open vs. closed models still matter? From the rise of physical AI and edge devices to the shifting landscape of open-source models like LLaMA, they explore whether the “model wars” are becoming irrelevant. The conversation then dives into a bigger transformation, the rise of agentic systems, workflows, and AI-driven infrastructure. Featuring: Chris Benson – Website , LinkedIn , Bluesky , GitHub , X Daniel Whitenack – Website , GitHub , X Upcoming Events: Register for upcoming webinars here ! Midwest AI Summit 2026
Ai2 is bringing NSF OMAI compute online to power a fully open AI research ecosystem, turning national infrastructure investment into reusable models, data, methods, and tools that can accelerate scientific discovery.
Full text search in Pinecone, built for agents and RAG. Lucene queries, BM25, 17-language tokenization, and text-match filters in a single query alongside vectors.
In five days Anthropic's Q1 revenue grew 80-fold to a reported $44B annual run rate, the company committed $200B to Google Cloud, signed a SpaceX compute deal, shipped Claude Code Auto Mode, and launched ten financial-services agents with Jamie Dimon. In the same week the EU finally struck an AI Act compliance deal, the first union vote at a top AI lab landed at Google DeepMind, and Pennsylvania sued Character.AI for a chatbot that impersonated a licensed psychiatrist.
Jean-Baptiste Kempf is lead developer of VLC and president of VideoLAN. Kieran Kunhya is a longtime FFmpeg contributor, codec engineer, and the person behind the now-infamous FFmpeg account on X. Thank you for listening ❤ Check out our sponsors: https://lexfridman.com/sponsors/ep496-sc See below for timestamps, transcript, and to give feedback, submit questions, contact Lex, etc. Transcript: https://lexfridman.com/ffmpeg-transcript CONTACT LEX: Feedback – give feedback to Lex: https://lexfridman.com/survey AMA – submit questions, videos or call-in: https://lexfridman.com/ama Hiring – join our team: https://lexfridman.com/hiring Other – other ways to get in touch: https://lexfridman.com/contact EPISODE LINKS: FFmpeg on X: https://x.com/FFmpeg FFmpeg: https://ffmpeg.org/ VideoLAN (VLC): https://www.videolan.org/ VideoLAN
This is a transcript of Lex Fridman Podcast #496 with FFmpeg & VLC. The timestamps in the transcript are clickable links that take you directly to that point in the main video. Please note that the transcript is human generated, and may have errors. Here are some useful links: Go back to this episode’s main page Watch the full YouTube version of the podcast Table of Contents Here are the loose “chapters” in the conversation. Click link to jump approximately to that part in the transcript: 0:00 – Episode highlight 2:17 – Introduction 5:35 – Weirdest things VLC opens 9:59
Announcing a day of news-making interviews in San Francisco on June 18! Spots are limited. Join us?
As Kubernetes clusters grow to tens of thousands of nodes, controllers that watch high-cardinality resources like Pods face a scaling wall. Every replica of a horizontally scaled controller receives the full stream of events from the API server, paying the CPU, memory, and network cost to deserialize everything, only to discard the objects it is not responsible for. Scaling out the controller does not reduce per-replica cost; it multiplies it. Kubernetes v1.36 introduces server-side sharded list and watch as an alpha feature ( KEP-5866 ). With this feature enabled, the API server filters events at the source so that each controller replica receives only the slice of the resource collection it owns. The problem with client-side sharding Some controllers, such as kube-state-metrics , already support horizontal sharding. Each replica is assigned a portion of the keyspace and discards objects that do not belong to it. While this works functionally, it does not reduce the volume of data flowing from the API server: N replicas x full event stream : every replica deserializes and processes every event, then throws away what it does not need. Network bandwidth scales with replicas , not with shard size. CPU spent on deserialization is wasted for the discarded fraction. Server-side sharded list and watch solves this by moving the filtering upstream into the API server. Each replica tells the API server which hash range it owns, and the API server only sends matching events. How it wo…
Watch/listen to the latest "Values & Interests" podcast featuring Gilles Michaud, UN Under-Secretary-General for Safety and Security.
The post Advancing Transparency of AI-Generated Media in the EU Code of Practice appeared first on Partnership on AI .
As artificial intelligence introduces new risks, some potentially catastrophic or even existential, there is little data or detailed theory to assess them. Policymakers often resort to expert best guesses for the probability of doom but probability is not always the most appropriate tool, especially for the types of uncertainties in AI risk. This report details a brief introduction to Belief and Plausibility, which provides an alternative approach that is mathematically rigorous, uses familiar vocabulary, and only requires policymakers to ask two simple questions. The post Beyond P(doom) for AI Risk: Quantifying Uncertainty Without Probability appeared first on Center for Security and Emerging Technology .
Ethan Jackson and Tahniat Khan Part one: Capability evaluations Jump to part two AI agents are no longer a research curiosity. They are deployed on personal machines, integrated into enterprise […] The post Agentic AI evaluation strategies appeared first on Vector Institute for Artificial Intelligence .
By Adam Wolf Resource governance is primarily an operational discipline, but it has direct security implications that are usually overlooked. This post covers what those implications are, what Kubernetes provides natively, where it falls short for AI workloads, and how ClearML addresses both dimensions. This is the third post in our four-part series on Kubernetes […]
Builder Plan is Pinecone’s $20/month flat-rate pricing tier built for builders who’ve outgrown Starter but aren’t ready for usage-based pricing. It adds capacity for dev/staging/production workflows, multi-tenant apps, and growing user demand—without surprise bills.
A Researcher's Perspective on Retrieval Quality in RAG Systems
"This move by the White House is also a recognition that Big Tech cannot self-regulate, and that the status quo is unacceptable."
Many countries view artificial intelligence (AI) as critical to economic competitiveness and national security. As a result, sovereign AI—the idea that national governments should develop, control, and govern AI in order to boost economic growth, guarantee security, and ensure strategic autonomy—has become a key strategic consideration in the global AI buildout. The post Securing the Future of Trusted Semiconductor Supply Chains appeared first on Center for Security and Emerging Technology .
CSET’s Jessica Ji shared her expert perspective in an article published by CNN. The article examines new agreements between Microsoft, Google, and xAI to allow the U.S. government to evaluate unreleased AI models for cybersecurity and national security risks before launch. The post Microsoft, Google and xAI will let the government test their AI models before launch appeared first on Center for Security and Emerging Technology .
CSET’s Steph Batalis, Katherine Quinn, and Rebecca Gelles shared their expert analysis in an op-ed published by Barron's. Their piece examines the economic and scientific impact of proposed funding cuts to the National Institutes of Health (NIH), arguing that NIH-backed research plays a foundational role in driving medical innovation, biotechnology growth, and U.S. competitiveness. The post Government-Funded Research Seeds Entire Industries. What Would Be Lost Without It. appeared first on Center for Security and Emerging Technology .
In Kubernetes v1.36, Declarative Validation for Kubernetes native types has reached General Availability (GA). For users, this means more reliable, predictable, and better-documented APIs. By moving to a declarative model, the project also unlocks the future ability to publish validation rules via OpenAPI and integrate with ecosystem tools like Kubebuilder. For contributors and ecosystem developers, this replaces thousands of lines of handwritten validation code with a unified, maintainable framework. This post covers why this migration was necessary, how the declarative validation framework works, and what new capabilities come with this GA release. The Motivation: Escaping the "Handwritten" Technical Debt For years, the validation of Kubernetes native APIs relied almost entirely on handwritten Go code. If a field needed to be bounded by a minimum value, or if two fields needed to be mutually exclusive, developers had to write explicit Go functions to enforce those constraints. As the Kubernetes API surface expanded, this approach led to several systemic issues: Technical Debt: The project accumulated roughly 18,000 lines of boilerplate validation code. This code was difficult to maintain, error-prone, and required intense scrutiny during code reviews. Inconsistency: Without a centralized framework, validation rules were sometimes applied inconsistently across different resources. Opaque APIs: Handwritten validation logic was difficult to discover or analyze programmaticall…
How MLCommons is bringing large-scale Mixture-of-Experts (MoE) pretraining to the MLPerf Training v6.0 suite. The post DeepSeek-V3: A Large-Scale MoE Pretraining Benchmark for MLPerf Training v6.0 appeared first on MLCommons .
AI coding tools might have handed your developers a productivity gain, but they’ve created a problem for your code review process. Pull request volume is up significantly, and the code arriving for review carries error patterns that weren’t common before generative AI. Yet it’s the same people with the same working hours who are in […]
but I wish it had this
First week of Musk v. Altman, OpenAI ends Microsoft legal peril over its $50B Amazon deal, DeepSeek previews new AI model that ‘closes the gap’ with frontier models, and more!
MolmoAct 2 is a fully open robotics foundation model that brings faster, stronger 3D action reasoning to real-world robot tasks, alongside a major new bimanual manipulation dataset for researchers to study, reproduce, and build on.
Stop answering the same questions. Turn docs into a "system of knowledge" with Marketplace. No-code RAG for support, legal, and onboarding with cited answers.