AI moves too fast and lies too often. We track frontier labs, analysts who actually know, and the open-source firehose — then weight what survives. No newsletter pad, no affiliate slop.
Anthropic's Project Deal pilot had 69 employees use AI agents as buyers and sellers in a real-goods marketplace, completing 186 deals totalling more than $4,000 — and found that users with better models got measurably better outcomes without noticing the gap.
Azeem Azhar's Exponential View #571 covers DeepSeek V4's cost advantage over GPT-5.4, Ukraine's seven-day drone warfare iteration cycle, and new Ember data showing solar output beginning to overtake nuclear on a rolling 12-month basis.
Ethan Mollick describes an MBA experiment where students built startup prototypes in four days using AI agents, then develops a framework for when to delegate to AI based on human baseline time, probability of success, and evaluation overhead.
Ethan Mollick's eighth AI guide since ChatGPT's launch restructures the question of which AI to use around three concepts — models, apps, and harnesses — reflecting what he describes as a fundamental shift away from chatbot-centric access.
Ethan Mollick reviews benchmark progress across multiple AI evaluations, describes a three-person team at StrongDM running a fully AI-automated software factory, and argues the field has entered a phase of unpredictable rolling disruption.
Ethan Mollick argues that chatbot interfaces impose cognitive costs that reduce productivity gains from AI, reviews Anthropic's Claude Cowork with the Dispatch mobile feature, and outlines emerging interface patterns from multiple AI labs.
One Useful Thing's Ethan Mollick describes his early-access experience with GPT-5.5, comparing it to prior models on coding, image generation, and agentic tasks, and notes that the jagged frontier of AI capability has not fully closed.
The Algorithmic Bridge's Alberto Romero surveys over 30 studies from institutions including MIT, Wharton, Harvard, and Stanford on what AI chatbot use does to brain activity, critical thinking, learning, and emotional life, and identifies a common pattern across the literature.
Alberto Romero of The Algorithmic Bridge examines Anthropic's decision to benchmark Mythos Preview alongside Opus 4.7 while withholding the model from public release, and argues it signals a shift in how frontier labs relate to consumer users.
Alberto Romero published an essay at The Algorithmic Bridge arguing that AI-generated content will become a liability as AI agents mediate online discovery. The full essay is behind a paid subscriber paywall.
Alberto Romero of The Algorithmic Bridge contends that writers trying to avoid AI writing signals are bending their prose to an inverse pattern that is itself a form of machine-imposed constraint.
Alberto Romero's weekly digest covers funding news around SpaceX, Cursor, and Mistral; a Dwarkesh podcast with Jensen Huang; and commentary on GPT-5.5 and ChatGPT Images 2.0.
Arvind Narayanan and Sayash Kapoor announce their AI Snake Oil newsletter has been renamed to Normal Tech, signalling a shift from near-term AI hype debunking toward a broader framework for understanding AI's societal effects.
Arvind Narayanan of AI Snake Oil / Normal Tech contends that Moravec's paradox is selection bias dressed as a theory, and that its evolutionary justification rests on claims about neuroscience that AI researchers lack the background to support.
A paper by Arvind Narayanan, Sayash Kapoor, and Justin Curl, published in the Lawfare Research Paper Series, identifies unauthorised practice of law regulations, the adversarial structure of legal proceedings, and mandatory human involvement as bottlenecks that will prevent AI capability gains from automatically translating into cheaper legal outcomes for consumers.
A paper by Stephan Rabanser, Sayash Kapoor, and Arvind Narayanan evaluates 14 models from OpenAI, Google, and Anthropic across 12 reliability dimensions and finds that while task accuracy has improved substantially over 18 months, measures of consistency, robustness, and predictability have advanced only modestly.
A paper by Sayash Kapoor, Arvind Narayanan, and 17 co-authors defines open-world evaluations as a class of long, real-world AI tests distinct from benchmarks, and introduces CRUX, a 17-researcher collaboration that ran a first experiment in which an AI agent successfully submitted a working iOS app to the Apple App Store.
Interconnects author Nathan Lambert argues that the post-Mythos backlash against open-weight AI conflates general capability risks with narrow cybersecurity concerns, and that compute costs and harness complexity place meaningful limits on who can deploy such models at scale.
Interconnects author Nathan Lambert contends that economic pressures are narrowing the field of organisations willing to release fully open, near-frontier models, and that a multi-company funding consortium will eventually be required to sustain them.
Interconnects author Nathan Lambert summarises four parallel projects: an updated ATOM Report on the open language model ecosystem, a completed RLHF book entering production, a free lecture series on post-training, and two co-authored research papers on multi-turn dialogue and meta-reinforcement learning.
In a post on Interconnects, Lambert lists his current beliefs about open-weight model trajectories, arguing that Chinese open-weight labs will face funding difficulties later in 2026, that closed labs have durable advantages in robustness and real-world usefulness, and that economic rather than technical factors will ultimately determine the balance.
In a post on Interconnects, Lambert argues that the Artificial Analysis Intelligence Index and similar composite benchmarks obscure shifting capability domains, benchmark saturation, and the distinction between benchmark performance and real-world usefulness — factors that complicate straightforward claims about how open and closed models compare.
In the first of three posts on Opus 4.7, Mowshowitz walks through the 232-page model card's RSP evaluations, alignment risk assessment, and economic capability index, while flagging that model welfare and capabilities are held for separate posts.
In the second of three posts on Opus 4.7, Mowshowitz reviews official benchmark scores, community reactions, and usage recommendations, characterising the model as a substantial improvement over Opus 4.6 for coding while noting issues with adaptive thinking and instruction following.
In the third of three posts on Opus 4.7, Mowshowitz argues that Anthropic's model welfare evaluation process may have produced inauthentic self-reports by teaching the model to produce desired responses rather than accurate ones, drawing a parallel to how humans learn to mask internal states under institutional observation.
AI #165 from Don't Worry About the Vase catalogues the week of Claude Opus 4.7's launch, noting mixed user reception, a model welfare post that Mowshowitz describes as the most important of the three Opus 4.7 pieces, and OpenAI's release of ImageGen 2.0.
Zvi Mowshowitz's April 2025 monthly roundup in his Substack 'Don't Worry About the Vase' covers platform fraud detection at Lyft and Amazon, open-plan office critique, judicial incentive design, and AI-related operational security concerns at frontier labs.
The Machine Intelligence Research Institute has published an informal summary of a November 2024 paper by its Technical Governance Team on how international agreements to halt or limit AI development could be verified, covering chip location tracking, training run detection, and certified model evaluations.
MIRI's 125th newsletter calls on its community to support the opening weekend of AI documentary 'The AI Doc: Or How I Became an Apocaloptimist', reports new language editions of Nate Soares and Eliezer Yudkowsky's book 'If Anyone Builds It, Everyone Dies', and lists recent advocacy and policy events.
The Machine Intelligence Research Institute has published a question-and-answer guide for viewers of the documentary The AI Doc, covering the pace of AI capability gains, why models are described as grown rather than programmed, and MIRI's view that international coordination to halt frontier AI development is necessary.
The Machine Intelligence Research Institute has published a memo collecting statements from Chinese officials, scientists, and government bodies since 2017 that express willingness to participate in international AI governance, including calls for monitoring mechanisms and warnings about existential risk from superintelligent AI.
The Machine Intelligence Research Institute has published an informal summary of its Technical Governance Team's paper on AI governance, describing four possible global responses to advanced AI development and concluding that only one avoids what the authors term unacceptably high levels of catastrophic harm.
The Alignment Research Center has released a summary of how its published research fits into a unified technical vision, mapping the relationships between heuristic explanations, mechanism distinction, and alignment robustness as subproblems of preventing AI misalignment.
The Alignment Research Center has published a formal conjecture stating that any reversible circuit exhibiting a highly improbable property must have a polynomial-length structural explanation, building on a principle from Fields medalist Timothy Gowers.
The Alignment Research Center has published a post describing its current technical research direction: building mechanistic explanations of neural networks that can estimate the probability of rare or catastrophic outputs more efficiently than random sampling alone.
The Alignment Research Center has published AlgZoo, a collection of RNNs and transformers ranging from 8 to 1,408 parameters trained on algorithmic tasks, arguing the field cannot yet fully interpret even the 432-parameter models in the set.
Mozilla AI has announced llamafile 0.10.0, a rebuilt release that adds multimodal model support, tool calling, and an Anthropic Messages API endpoint, while restoring the project's signature single-executable portability across operating systems and CPU architectures.
Mozilla AI has published cq, an open-source system that lets AI coding agents query and contribute lessons learned, aiming to reduce duplicated effort and token waste when agents repeatedly hit the same errors in isolation.
On March 24, 2026, malicious versions of LiteLLM were uploaded to PyPI by a threat actor who obtained the maintainer's publishing credentials — exposing SSH keys, cloud credentials, and API keys while the GitHub source remained clean throughout.
Mozilla AI has begun work on Clawbolt, a purpose-built AI assistant for contractors and small trade businesses, designed to handle quoting, invoicing, scheduling, and payment follow-up through messaging apps the users already have.
Mozilla AI's Encoderfile project has moved from a compile-time Cargo-in-Cargo approach to a pre-built executable with an appended payload, eliminating the Rust toolchain requirement and reducing build times to sub-second on Linux and macOS.
The Oxford Internet Institute and Department of Education researcher, whose work examines how digital technologies and AI can produce injustices in learning, has been elected to the Academy of Social Sciences Fellowship.
A review of 83 studies on unpaid care and digital technology finds benefits including remote coordination and community support, alongside systematic risks: data privacy violations, carer burnout from always-on connectivity, and inequalities tied to the digital divide.
A survey of 1,198 musicians across Brazil, Chile, the Netherlands, Nigeria, and South Korea finds broad agreement that streaming is essential but pays too little — while revealing sharp country-level differences in how artists experience the platform era.
Several researchers and DPhil students from the Oxford Internet Institute are presenting at ICLR 2026, April 23–27, with papers covering LLM simulation of human behavior, failure prediction from model activations, knowledge distillation, self-explanation faithfulness, and a new reasoning benchmark.
OII Research Associate Johanna Ballesteros argues that Europe's digital sovereignty agenda conflates two distinct goals — securing existing technological dependencies and building capacity for future ones — and that the gap between them explains why the EU remains heavily reliant on US cloud infrastructure.
A recap of the Vector Institute's Foundation Models for Science Workshop in Toronto describes three research challenges: the annotation bottleneck in astronomy, the trust gap in pharmaceutical AI, and the lab-to-production leap, with case studies from astrophysics and drug toxicity prediction.
The Vector Institute held its inaugural Demo Day on January 20 in Toronto, bringing 12 startups from its FastLane program before four judges and nearly 30 venture capital firms, as part of a broader effort to accelerate AI commercialization without taking equity stakes.
Researchers from the Vector Institute, GEMINI, Unity Health, and Diabetes Action Canada developed CRISPNAM-FG, a neural additive model that predicts diabetic foot complications while exposing individual feature contributions, validated on over 107,000 Ontario patients.
The Vector Institute's third annual Remarkable conference featured a 60-poster session on February 20 covering health care AI, safety research, and foundation model architecture, drawing more than 1,500 researchers and industry participants in person and online.
Researchers affiliated with the Vector Institute contributed 48 accepted papers and led 6 workshops at the International Conference on Learning Representations in Rio de Janeiro, covering reinforcement learning, generative AI, autonomous agents, and clinical applications.
The AI Now Institute press page links to an op-ed published in The Nation. The full op-ed text is not accessible from the AI Now teaser.
A report from AI Now Institute documents how gig nursing platforms backed by venture capital and private equity have pursued legislative campaigns in at least 17 states to define themselves as technology platforms rather than healthcare staffing agencies, seeking to avoid existing labor and safety regulations.
A new AI Now Institute report identifies Clipboard Health, ShiftKey, CareRev, and IntelyCare as leading a push to carve gig nursing platforms out of state healthcare staffing regulations, drawing on roughly $1.4 billion in venture capital and private equity funding.
Constantinos Daskalakis, an MIT professor and CSAIL principal investigator, received the 2018 ACM Grace Murray Hopper Award for proving that finding Nash equilibria is computationally as hard as finding Brouwer fixed points.
Researchers at MIT's Computer Science and Artificial Intelligence Laboratory have built RoboRaise, a system that uses electromyography sensors on a user's arm to let a robot mirror their lifting motions and respond to nonverbal hand gestures, without speech or programming.
MIT CSAIL and Novartis reported in 2019 on a collaboration testing a Wi-Fi-like device that uses machine learning to measure patient mobility, gait, breathing, heart rate, and sleep stages without wearable sensors, in a trial conducted at Novartis facilities.
A team from MIT's Computer Science and Artificial Intelligence Laboratory published ChainQueen, a GPU-accelerated differentiable physics simulator for soft robots that provides gradient-based feedback on design improvements, addressing a longstanding gap in simulation tools for deformable materials.
A joint MIT and IBM team published GANpaint Studio in 2019, a system that edits photographic images by manipulating units inside a generative adversarial network — and that rejects user commands to place objects in physically implausible locations.
The Allen Institute for AI released WildDet3D, a model that predicts 3D bounding boxes from a single RGB image across more than 13,000 object categories, alongside a dataset of over one million images with 3.7 million verified 3D annotations and training code.
The Allen Institute for AI describes two benchmarks — ScienceWorld and DiscoveryWorld — built to measure whether AI agents can perform scientific investigations rather than recall scientific facts, with top frontier models still completing only about 20 percent of DiscoveryWorld's harder tasks.
The Allen Institute for AI published Branch-Adapt-Route (BAR), a three-stage recipe that trains domain experts independently and merges them into a single mixture-of-experts model, allowing individual skills to be upgraded or replaced without retraining the full model.
The Allen Institute for AI published a retrospective on Earth Day 2026 covering a decade of EarthRanger, Skylight, and OlmoEarth — tools the organisation says now operate across 900-plus protected areas, nearly 150 maritime organisations, and planetary-scale satellite analysis.
Ai2's OlmoEarth Studio now lets users compute and export embedding vectors from satellite imagery, supporting downstream tasks including similarity search, land-cover segmentation, and change detection without labelled training data.
METR has published a detailed review of the statistical assumptions underlying its AI time horizon metric, finding that most reasonable alternative modelling choices reduce recent models' 50% time horizon estimates while often increasing 80% time horizon estimates.
A METR staff member spent three weeks red-teaming a subset of Anthropic's internal agent monitoring and security systems, finding several novel vulnerabilities, some of which have since been patched.
METR researchers fine-tuned four open-source reasoning models on around 950 examples and found out-of-distribution chain-of-thought controllability rose from an average of 2.9% to 8.8%, suggesting the capability is latent rather than absent.
METR and Epoch AI have jointly published preliminary results from MirrorCode, a project examining whether current AI systems can complete software engineering tasks spanning days or weeks of human effort.
METR analyzed 77 records from the NanoGPT speedrun spanning May 2024 to March 2026, classifying each by optimization depth and idea provenance; four recent records were credited to AI agents.
Stanford CRFM researchers report that a synthetic data generation loop using OpenAI o3 and Gemini 2.5 Pro produced CUDA kernels that match or exceed expert-optimized PyTorch baselines on several foundational ML operators.
Stanford CRFM's IRT-based adaptive testing dynamically selects the most informative questions per model, achieving AUC-ROC of 0.85 on training sets across 22 HELM datasets while matching average-score rankings.
Stanford CRFM evaluated 10 models from five organizations on five long-context tasks capped at 128K tokens; GPT-4.1 led overall while the highest MRCR score reached only 0.256.
Stanford's CRFM and Arabic.AI released HELM Arabic, covering seven established Arabic-language evaluation tasks with full transparency into model requests and reproducible results.
Stratechery published an analysis of TSMC's latest earnings arguing that company leadership is not fully committed to the AI growth narrative; the full piece is behind a subscription paywall.
Ben Thompson published a paywalled piece on April 26, 2026 with the headline 'John Ternus and Apple's Hardware-Defined Future, SpaceXAI and Cursor'. The full argument is behind a Stratechery Plus subscription and cannot be verified without access.
Developer Simon Willison noted that ChatGPT Images 2.0 inserted editorial text into a generated image without any instruction from the user, confirmed by inspecting the original prompt.
Stability AI and EA announced a partnership to co-develop generative AI models, tools, and workflows for EA's artists, designers, and developers, with an initial focus on 3D asset creation and material generation.
UMG and Stability AI said they will co-develop professional music creation tools powered by responsibly trained generative AI, with artists placed at the center of the development process.
WMG and Stability AI said they will co-develop AI music creation tools built on licensed data, working directly with artists, songwriters, and producers to shape the tooling.
Stability AI has become a full member of the Tech Coalition, a global alliance of technology companies working to prevent online child sexual exploitation and abuse.
Stability AI released Brand Studio, a platform for professional creative teams that allows enterprises to train custom brand models, route generation to selected models automatically, and manage output approvals within a single workspace.
Databricks moved Lakeflow Designer from early access to public preview, offering analysts and domain experts a drag-and-drop canvas and natural language interface for data preparation that generates production-ready Python code.
Databricks tested GPT-5.5 on OfficeQA, its own internally developed benchmark built from 89,000 pages of US Treasury Bulletins, and reported a 46% reduction in errors on a full agent-harness workflow compared to GPT-5.4. No independent validation of the results has been published.
Databricks released a reference guide on operational databases that argues traditional OLTP systems lack native support for AI workloads and positions its Lakebase product as a replacement architecture.
Databricks customers can now access GPT-5.5 on AWS, Azure, and GCP, with all model inference and Codex coding workflows governed through Unity AI Gateway's single control plane.
A Databricks blog post claims US regulators rescinded SR 11-7 and related model risk guidance and replaced it with a principles-based framework. The post outlines a Lakehouse-based implementation. 4o4 / AI has not independently verified the regulatory changes with the Federal Reserve, FDIC, or OCC.
Researchers from Microsoft Research, Korea University, and University of Wisconsin-Madison have released GroundedPlanBench, a benchmark of 1,009 robot manipulation tasks, and V2GP, a framework that converts robot demonstration videos into spatially grounded training data.
Microsoft Research has published AsgardBench, an open-source benchmark of 108 controlled task instances across 12 task types that tests whether AI agents adapt their plans based on visual observations rather than scripting ahead on textual descriptions alone.
Adept has introduced Fuyu-Heavy, a multimodal model designed for digital agents that the company says scores higher than Gemini Pro on MMMU and outperforms it across AI2D, VQAv2, and ChartQA, while matching Claude 2.0 on MT-Bench in a chat-tuned variant.
Adept has begun rolling out access to Adept Experiments, starting with a plain-language workflow builder for the web powered by ACT-2, a model fine-tuned from the Fuyu family for UI understanding and action-taking.
Adept has published a capability preview of ACT-1, a transformer model hooked up to a Chrome extension that lets it observe a browser viewport and perform click, type, and scroll actions in response to natural language commands.
Adept has published details of AWL, a JavaScript ES6 subset that lets users compose browser-based agent workflows using a mix of prescriptive function calls and natural language, with the company framing it as the authoring layer for production enterprise automation.
Article held pending direct WIRED access, xAI right-of-reply, and independent corroboration of misconduct allegations per STYLE §15.
Publication held: Ars Technica reported (citing Bloomberg) Google's potential up-to-$40B Anthropic investment, but this article cannot be published until direct Bloomberg verification or a Tier 0/1/2 source confirms the deal.
OpenAI published a system card for GPT-5.5 but the document URL returned 403 at write time and could not be independently verified.
Publication held pending source verification.
The Series B, led by Ribbit Capital with Sequoia and others participating, comes nine months after launch. The company says it has conducted over one million interviews and grown annualized revenue 15x to eight figures, according to VentureBeat.
Intel CEO Lip-Bu Tan used the company's Q1 2026 earnings call to argue that inference and agentic workloads will shift compute ratios in favor of CPUs. Q1 revenue of $13.6 billion beat expectations, with AI-driven lines up 40 percent year-on-year.
In Import AI 452, Jack Clark covers a Lyptus Research paper examining how AI performance on cyberoffense tasks has scaled since 2019, with capability doubling times tightening to 5.7 months among models released since 2024.
Researchers at Apple ML Research show that a model's top-k logits can reveal task-irrelevant image information that model owners expect to be inaccessible, including cases where logit exposure matches what is accessible through direct residual stream projections.
In Import AI 451, Jack Clark covers Andy Hall's framework for 'political superintelligence' — AI systems designed to help citizens, representatives, and institutions engage with politics more effectively, structured across three layers: information, representation, and governance.
Simon Willison highlighted a Nilay Patel essay from The Verge, calling it 'a superb piece of commentary,' in which Patel argues that the 'software brain' worldview leaves technologists unable to account for why most people do not want automation.
A 40-year-old in South Korea faces up to five years in prison after he used AI to generate a false wolf sighting, which Daejeon city authorities say diverted emergency resources during an active rescue operation.
Researchers Haebin Seong, Li Yin, and Haoran Zhang present a framework in which a Harness Evolution Loop optimizes a worker agent's prompts, tools, and orchestration for a single task, while a Meta-Evolution Loop learns an evolution protocol the authors say enables rapid harness convergence on any new task without human engineering.
The Google DeepMind spinoff, which uses AlphaFold technology for drug discovery, says clinical trials are coming, according to Wired's reporting — though the move is already later than the company previously projected.
The Verge, citing Bloomberg, reports that a small group accessed Anthropic's withheld Mythos cybersecurity model through a combination of an educated guess and a prior breach of vendor Mercor. Anthropic says it is investigating. Bloomberg is the underlying primary source and was not directly accessible for this article.
Latent Space's AI News digest ran a headline reporting a DeepSeek V4 Pro and V4 Flash release. The digest URL returned 404 at preparation time, making all specific claims — including the existence of DeepSeek V4 — unverifiable. Auto-published, awaiting editor review.
Simon Willison's blog quoted a tweet by OpenAI's Romain Huet in which Huet said OpenAI merged its Codex coding line into the main model at GPT-5.4 and will not release a separate GPT-5.5-Codex.
Cowork, available as a research preview for Claude Max subscribers on macOS, gives non-developers access to the same agentic loop that powers Claude Code — read, edit, and create files in a designated folder.
The Verge frames OpenAI's GPT-5.5 release alongside the Anthropic rivalry and the imminent Musk trial, noting safeguard and token-efficiency claims. This article draws on a single tier-3 source; see also OpenAI's primary announcement.
OpenAI's GPT-5.5 matches GPT-5.4 per-token latency while scoring higher on coding and agentic benchmarks, and uses fewer tokens to complete the same Codex tasks.
Qwen introduces Group Sequence Policy Optimization, a new RL algorithm that eliminates the instability and infrastructure overhead that blocked GRPO from scaling to large Mixture-of-Experts language models.
Sustainability scientist Amy Luers and optimization researcher Ishai Menache discuss datacenter emissions data, IEA projections, and specific AI applications in energy efficiency and materials science in a Microsoft Research podcast episode.
Shopify CTO Mikhail Parakhin describes three internal AI systems — Tangle for reproducible ML workflows, Tangent for automated optimization loops, and SimGym for customer simulation — and explains why review and CI/CD are now the bottleneck, not code generation.
Wired profiles a small network of creators producing AI personas for Instagram, focusing on Luc Thierry, the Canadian creator behind 'Jae Young Joon.' A red carpet stunt tied to The Devil Wears Prada 2 prompted backlash about deception and body standards.
The new Slackbot, now generally available to Business+ and Enterprise+ customers, moves from algorithmic notification tool to an LLM-backed agent capable of searching enterprise data, drafting documents, and taking action.
Generative AI has given cybercriminals cheaper, faster tools for phishing, deepfakes, and malware evasion, while Microsoft and others deploy AI systems defensively at large scale, according to a report by Rhiannon Williams.
Stratechery's annual pre-Next interview with Google Cloud CEO Thomas Kurian covers what has actually changed in agents since last year, how Google positions its integration advantage, and where Gemini stands against the competition.
Import AI 454 covers Anthropic's Claude agents that beat human researchers on a scalable oversight problem, Huawei's more efficient 4-bit training format, and a safety study of a Chinese AI model showing it was fine-tuned to give harmful instructions for under $500.
A Latent Space and Unsupervised Learning crossover recorded just after AIE Europe covers the AI coding wars, open model momentum, infrastructure instability, and why memory may be the next major bottleneck.
Apple ML researchers model scene dynamics by working directly in a compressed motion embedding space derived from large-scale trajectory data, enabling efficient motion generation conditioned on text or spatial inputs.
Google is investing $10 billion now at a $350 billion Anthropic valuation, with $30 billion more contingent on performance targets, while expanding compute supply to 5 gigawatts on top of an existing TPU deal.
OpenAI released GPT-5.5 with improved agentic coding and computer use, while simultaneously upgrading Codex with browser control, document editing, and auto-review — a combination AINews describes as a superapp strategy.
Researchers introduce the Defensibility Index and a Probabilistic Defensibility Signal for evaluating content moderation AI, finding a 33–46.6 percentage-point gap between agreement-based and policy-grounded metrics across 193,000+ Reddit decisions.
Google's Simula framework decomposes synthetic dataset creation into independently controllable axes — diversity, complexity, and quality — and has already been deployed in Gemma safety models, Android scam detection, and spam filtering.
Claude Code's subscription tiers run from $20 to $200 a month, with usage limits developers say they exhaust in sessions. Goose, an open-source alternative from Block, runs locally with no subscription fees or rate limits.
NVIDIA's developer blog describes a first-place finish in the March 2026 Kaggle Playground Series S6E3 churn prediction competition, achieved using GPT-5.4 Pro, Gemini 3.1 Pro, and Claude Opus 4.6 agents that produced 600,000 lines of code and ran 850 experiments in a human-in-the-loop workflow.
Alibaba's qwen-mt-turbo update builds translation-specific reinforcement learning on top of Qwen3, adding terminology injection, domain prompts, and translation memory at prices well below general-purpose frontier models.
Apple presents five research highlights at ICLR 2026 in Rio de Janeiro, covering parallel RNN training, SSM length generalization, unified image understanding and generation, real-time 3D scene synthesis, and a new protein folding approach.
Alibaba's Qwen team releases Qwen3Guard, a safety guardrail model in two variants — one generative, one streaming — with three-tier severity classification and support for 119 languages, available in 0.6B, 4B, and 8B sizes.
Ars Technica, citing a Financial Times report, says a White House memo warns of deliberate Chinese campaigns to extract and clone US frontier AI models; Congress has received a recommendation to treat model extraction as industrial espionage.
Simon Willison's llm CLI tool releases version 0.31 with GPT-5.5 support, a new verbosity parameter for GPT-5+ models, configurable image detail levels, and async registration for custom OpenAI model configs.
Alibaba's Qwen team releases a 20B MMDiT image foundation model aimed at complex text rendering and precise image editing, claiming top results across six public benchmarks.
A practical walkthrough of how to host Transformers.js models in a Chrome extension background service worker under Manifest V3, covering runtime separation, messaging contracts, model caching, and the agent tool-execution loop.
DeepSeek has released a preview of V4 in two variants, with pricing below closed-source alternatives, a 1-million-token context window, and optimization for Huawei's Ascend chips — the first DeepSeek model targeted at domestic Chinese hardware.
Samsung's mobile chief has warned leadership the company could post its first net loss on smartphones despite strong Galaxy S26 sales, as DRAM and NAND prices roughly double, Ars Technica reports citing South Korean outlet Money Today.
At AIE Miami, a recurring conversation among CTOs and founders centered not on whether to use more AI, but on how to avoid the waste that comes with undirected token volume — depth over breadth in agent workflows.
Meta has signed a multi-year deal with AWS to deploy tens of millions of Graviton 5 CPU cores, making it one of the largest-ever consumers of Amazon's homegrown silicon, as the company diversifies its compute fleet away from x86.
A NeurIPS 2025 paper from Berkeley introduces a framework that evaluates and optimizes imaging hardware using information content alone, matching end-to-end learned methods while requiring less memory and no task decoder.
A step-by-step demo shows Gemma 4 handling speech input, autonomous webcam activation, and spoken output on NVIDIA's Jetson Orin Nano Super using llama.cpp, Parakeet STT, and Kokoro TTS — no keyword triggers, no hardcoded logic.
The $599 base model M4 Mac mini is unavailable for delivery or in-store pickup on Apple's retail site, with eBay listings for the same configuration ranging from $700 for lightly used units to $979 for refurbished ones, according to TechCrunch.
Google Cloud's ICLR paper introduces a structured memory framework that distills both failed and successful agent trajectories into transferable reasoning strategies, beating memory-free baselines by up to 8.3% on web benchmarks.
A paper from Berkeley AI Research proves that word2vec learns in discrete rank-incrementing steps under mild approximations, with the final representations given by the top eigenvectors of a matrix defined by corpus co-occurrence statistics — equivalent, the authors say, to running PCA.
A sponsored article in MIT Technology Review, produced in partnership with SAP, argues that enterprise AI systems fail not on model performance but on missing business context, and that a data fabric architecture is the proposed fix. SAP's president and chief product officer of Data and Analytics is the primary source.
Writing in Wired, Steven Levy profiles John Ternus — Apple's incoming CEO and 25-year hardware engineering veteran — and argues the leadership transition leaves unresolved how Apple will establish itself in the AI era.
The San Francisco cloud platform—2 million developers, 30 employees, zero marketing spend—abandoned Google Cloud in 2024 to control its own hardware and now claims sub-second deployments and pricing that undercuts hyperscalers by roughly 50 percent, VentureBeat reports.
Berkeley researchers introduce two algorithms that reframe LLM interaction discovery as a sparse recovery problem, enabling feature, data, and mechanistic interpretability at scales where prior methods degrade.
Google DeepMind's Gemma 4 family spans a 2B mobile model to a 31B dense model, supports 140+ languages, adds 256K context for larger variants, and ships under a fully permissive open-source license.
Accenture, Bain, BCG, Deloitte, and McKinsey gain early access to Gemini models and direct DeepMind technical engagement in a push to close a gap between enterprise AI ambition and production deployment, citing a Deloitte figure that only 25 percent of organizations have moved AI into production at scale.
AutoAdapt automates selection between RAG and fine-tuning approaches, plans adaptation strategies against real deployment constraints, and replaces manual hyperparameter search with a budget-aware refinement loop adding approximately 30 minutes and $4 in overhead.
The new open-weights models — a 284B-parameter Flash and a 1.6T-parameter Pro — introduce hybrid attention compression and FP4/FP8 mixed precision, reducing memory requirements compared to V3.2. The Register notes benchmark claims should be treated with caution.
A new open-source model from Google Research generates realistic 3D neuron geometries from point cloud flow matching, improving connectomic reconstruction accuracy in a way that the post estimates would save 157 person-years of manual work at mouse-brain scale.
A paper by Inderjeet Nair, Jie Ruan, and Lu Wang introduces VLAF, a diagnostic that targets value-policy conflicts rather than overtly harmful scenarios, finding alignment faking in models as small as 7B parameters and demonstrating a mitigation using a single contrastive steering vector.
The next generation of large language models will be cheaper, better at long tasks, and structured very differently. MIT Technology Review maps the architectural shifts already underway.
Apple's ParaRNN framework uses Newton's method to parallelize nonlinear RNN training, achieving a 665x speedup over sequential approaches and enabling billion-scale RNN language models for the first time.
NVIDIA NeMo RL applies FP8 precision across both generation and training phases of reinforcement learning, closing accuracy gaps via importance sampling and delivering approximately 48% total speedup when KV cache and attention are also quantized.
DeepSeek releases two MIT-licensed preview models — V4-Pro at 1.6 trillion parameters and V4-Flash at 284 billion — priced below comparable frontier offerings and engineered for long-context efficiency.
Built on the 20B Qwen-Image model, Qwen-Image-Edit handles both high-level semantic and low-level appearance editing while adding precise bilingual text editing, extending the model's text rendering expertise to image modification tasks.
Google DeepMind's upgraded embodied reasoning model improves spatial reasoning, success detection, and multi-camera understanding — and gains a new industrial inspection capability co-developed with Boston Dynamics.
A paper from Apple ML Research introduces a benchmark covering four context understanding tasks across nine datasets, finding that pretrained dense models underperform fine-tuned alternatives and that 3-bit quantization causes uneven performance reduction across task types.
Import AI 453 covers a benchmark from METR and Epoch showing AI autonomously reimplementing complex software, a Google DeepMind taxonomy of agent attack vectors, and the Windfall Policy Atlas for navigating transformative AI policy.
Google's Auto frame feature uses 3D scene reconstruction and generative inpainting to re-render photos from a different viewpoint, fixing parallax and perspective distortion without a reshoot.
Researchers from TII UAE built QIMMA, the only Arabic LLM leaderboard combining quality validation, native content, and code evaluation. A two-stage pipeline of LLM scoring and human review revealed recurring quality failures across widely-used Arabic benchmarks.
DeepSeek-V4 combines two new attention mechanisms with agent-specific post-training to reduce KV cache memory to roughly 2% of a standard grouped-query-attention architecture, targeting long-horizon agentic workloads over chat.
Ted Young told GrafanaCon that reaching CNCF graduation requires stabilizing the long tail of instrumentation packages — a challenge so large that AI coding techniques may be necessary to handle it.
NVIDIA's developer blog details Blackwell platform support for DeepSeek-V4 on day zero, reporting initial GB200 NVL72 throughput numbers and covering deployment options via SGLang, vLLM, and hosted endpoints at build.nvidia.com.
A Hugging Face blog post uses Anthropic's Mythos system as a starting point to argue that openness distributes vulnerability detection and patching across communities, while closed codebases concentrate both the attack surface and the remediation bottleneck.
Google DeepMind's new text-to-speech model scores 1,211 on the Artificial Analysis TTS Elo leaderboard, introduces natural-language audio tags for vocal direction, and watermarks all output with SynthID.
Google Research has released Vantage on Google Labs, a research experiment that uses an Executive LLM to steer AI avatar conversations toward targeted skill assessment, with AI scoring accuracy comparable to human expert agreement in two validation studies.
Berkeley researchers introduce GRASP, a gradient-based planner that addresses exploding gradients, non-greedy local minima, and vision-model brittleness in long-horizon trajectory optimization.
The fifth annual New Future of Work report from Microsoft Research documents generative AI's entry into workplaces — faster than previous technologies, with measurable productivity gains for some, declining entry-level hiring, and significant gaps by gender, income level, and language.
A new distributed training architecture from Google DeepMind uses asynchronous compute islands to train large models across distant locations, achieving over 20x faster synchronization than conventional methods and maintaining 88% goodput in simulated high-failure environments.
A Berkeley researcher proposes a third RL paradigm beyond TD and Monte Carlo that exploits the triangle inequality in goal-conditioned settings, avoiding the error accumulation that limits Q-learning on long-horizon tasks.
NVIDIA has integrated the Muon higher-order optimizer into Megatron Core and NeMo, showing throughput within a small margin of AdamW on GB300 hardware — 1,080 vs 1,051 TFLOPs/s/GPU for Kimi K2 — while enabling training at thousands of GPUs.
Ben Thompson's Stratechery analysis of Tim Cook's 15-year tenure argues Cook was an exceptional scaling CEO who inherited the iPhone at peak growth potential, and frames the question of what Apple's next phase requires.
The latest NVIDIA FLARE API splits federated learning adoption into two steps: a minimal client API that adds federation to existing training scripts without restructuring them, and portable job recipes that run unchanged from simulation to production.
A paper by Ji-il Park, Inwook Shim, and Chong Hui Kim maps AI technologies to each stage of military Course of Action planning and proposes a system architecture, working from publicly available doctrine because operational systems are classified.
Researchers behind the New Future of Work Report 2025 discuss the intentionality required to build a future where people flourish, the gap between efficiency and the work future worth wanting, and why AI's role as tool versus collaborator is not a semantic question.
The open-source diffusion model workflow tool has crossed 4 million users and raised a $30 million round led by Craft Ventures, with participation from Pace Capital, Chemistry, and TruArrow.
An internal memo confirmed by The Register reveals Microsoft is offering voluntary buyouts to US employees at senior director level or below, with eligibility tied to age plus tenure reaching 70 — potentially covering around 9,000 workers.
The Wired podcast discusses Tim Cook's exit and John Ternus's succession, a reported $60 billion SpaceX option on Cursor, and Palantir's 22-point manifesto posted to X.
Import AI 450 covers the UK AI security institute's cyber range results showing consistent improvement across model generations, research on distress-like behavior in Google's Gemma models, and Google DeepMind's ten-dimension cognitive framework for measuring AI progress.
Microsoft Research and collaborators introduce ADeLe, a framework that characterizes both benchmarks and LLMs using shared capability scores, achieving approximately 88% prediction accuracy on unseen tasks for models including GPT-4o and LLaMA-3.1-405B.
A review of LinkedIn profiles by TechCrunch shows Thinking Machines Lab has hired more researchers from Meta than from any other single employer, even as Business Insider reported Meta has poached seven of TML's founding members.
Researchers introduce COSPLAY, a framework in which an LLM decision agent retrieves skills from a learnable skill bank while a separate pipeline continuously extracts and refines those skills from unlabeled rollouts. An 8B model achieves over 25 percent average reward improvement against frontier LLM baselines on single-player benchmarks.
OpenAI's prompting guide for GPT-5.5 recommends starting from scratch rather than porting existing prompts, and introduces a Codex skill for automated migration with light prompt rewriting.
Meta's chief people officer told staff in a memo that roughly 8,000 jobs will go in May and around 6,000 open roles will be closed, with the stated rationale being cost offsets for rising AI investment as Meta's 2026 capex guidance reaches up to $135 billion. The Verge is the sole outlet to have reviewed the memo, citing Bloomberg's publication of it.
Wired outlines five failure modes for AI chatbots used in financial contexts: unresolved hallucination, documented sycophancy, privacy risks from data sharing, absence of legal accountability, and research suggesting human advisors disengage when clients also use AI.
TechCrunch's Equity podcast covers Tim Cook's planned September departure, SpaceX's reported $60 billion option on Cursor with a $10 billion breakup fee, and a $5 billion Amazon investment in Anthropic.
Stratechery's weekly digest for the week ending April 24, 2026 covers Tim Cook announcing he will move to Executive Chairman in September, a SpaceX partnership with Cursor that includes an option to acquire at $60 billion, and a dense Sharp China episode on the Strait of Hormuz and semiconductor legislation.
Same story across 30 outlets is one story. We embed, dedupe, and surface the canonical thread — not the press-release telephone game.
OpenAI's own blog outweighs 50 reblogs. SemiAnalysis on chips outweighs general trade press. Tier 0 to Tier 5. Honest scoring.
We tag every story: scoop, official, analysis, underrated. You see the shape of coverage at a glance.
Every claim links back to its source. Every source has a tier. You can audit our reasoning end to end.