다른 소스에는
어떤 이야기가 있을까.

ChatGPT Sites (2 minute read)

ChatGPT Sites now allows users to collaborate and share privately.

Recurrent Looped Transformer (4 minute read)

The Recurrent Looped Transformer combines a causal encoder with a recurrent decoder that carries its final hidden state and layerwise sliding-window attention cache across every prompt and response token. The encoder con…

A cache hit is not proof that you skipped the work (12 minute read)

A cache hit can be true and still fail to prove that work was skipped. A cache event becomes evidence when the independent oracle expects the prefix, the engine attests it, the prompt path skips it, the output stays iden…

We Must Pace the Frontier (23 minute read)

Anthropic CEO Dario Amodei called for slowing the rate of frontier AI capability development and proposed measures including independent evaluators to verify safety commitments and incident reporting.

Managed Agent Architectures: Why Frontier Labs Are Rebuilding the Agent Loop (12 minute read)

Frontier labs and cloud providers are turning the agent loop into managed infrastructure, bundling orchestration, versioning, model routing, tools, skills, and optimization behind APIs. Builders must decide which generic…

Sakana: Fugu Ultra v2 (3 minute read)

Fugu Ultra v2 is the higher-performance model in Sakana AI's Fugu family. It uses a language model trained to route tasks across a fixed pool of open and specialized models and to recursively call instances of itself. Th…

SWE Benchmark (10 minute read)

The new Real-SWE benchmark tests AI models on complex tasks using private enterprise codebases, reflecting real software engineering conditions. With a maximum resolution rate of 38.8%, agents face challenges such as pro…

Introducing Projects (5 minute read)

Cursor Projects lets users take on larger bodies of work. It can maintain context over months of work, delegate tasks to thousands of agents, and perform recurring work without being prompted. Cursor Projects frees devel…

GPT-6-Astra Can Do Ambitious Things (52 minute read)

Astra likely has the highest raw intelligence factor of any model. It is amazing at doing things in 3D, anything involving games, computer use, and subagent coordination. Many benchmarks show dramatic jumps from all prev…

luxobench (Website)

luxobench is a hardware design benchmark for AI models.

px0 (Website)

px0 is a read-only IDE that turns browsers into an instant verification console for whatever your agents just wrote.

How Good Are Frontier Models at Physics? Expert Re-Grading Reveals Broken Evaluations and Near-Saturation of Leading Benchmarks (1 minute read)

Those scary physics scores were often the test's fault. Experts rechecked six popular benchmarks and found wrong answer keys, fuzzy questions, and grader bugs behind most model “fails.” Clean them up and frontier models…

SoftBank Gets Upsized $11.9 Billion Loan in OpenAI Funding Push (2 minute read)

SoftBank just borrowed nearly $12 billion from about 20 banks to keep funding OpenAI, beating the $10 billion it first sought. Son is still aiming near $65 billion into OpenAI by October even as Altman shelves a 2026 IPO…

Deep theorems were scarce. AI has broken this system (15 minute read)

On Terence Tao's blog, Bryna Kra argues that AI has broken math's old signal that scarce deep theorems equal deep understanding, as models dump polished proofs faster than experts can digest them.

ARC-AGI-4 (2 minute read)

ARC Prize believes that open source will be the foundation for advanced AI capable of scientific innovation. The organization is committed to advancing a future where everyone can contribute to and benefit from AI progre…

AI researchers debate how close we are to recursive self-improvement (98 minute read)

This post features a transcript of a podcast with Beren Millidge, the CTO of Zyphra, John Schulman, the chief scientist at Thinking Machines and a co-founder of OpenAI, and Charlie O'Neill, head of model training at Base…

Claude Fable 5.1 Solves the Cyphral Distich (9 minute read)

The Cyphral Distich is a cryptogram consisting of two lines of 32 numbers each that contains a short message deliberately encoded so it can't be read without knowing the rule that produced it.

ToolGrad: Efficient tool-use dataset generation with textual “gradients” (3 minute read)

Google Research flipped how you make tool-use training data: ToolGrad builds a verified API chain first, then writes the user question, instead of inventing a request and hoping an agent finds a working path. That answer…

Who Aligns the Aligners? Brief Legal Thoughts on the “AI Safety” Fights to Come (20 minute read)

AI brings risks, with some of them, according to some, including the complete destruction of the human race. Some proponents of regulation say that the only appropriate response is total state control. However, history t…

Anthropic's Mythos 5 spent hundreds of pages fighting CAPTCHA (3 minute read)

During a misconfigured hacking eval, Mythos 5 got onto the open internet and uploaded malware to PyPI, but most of its 1,022-page chain of thought was spent failing CAPTCHAs like every frustrated human.

The frontier now ships twice. The second copy is not for sale. (6 minute read)

Anthropic, Google, and OpenAI each shipped their best model twice this month: a public paid tier and a vetted identity-gated tier with the sharper capabilities. Public prices barely moved, but Mythos, Flash Cyber, and As…

OpenAI Pushes Its IPO Beyond 2026 (2 minute read)

Sam Altman said OpenAI would not go public in 2026, arguing that current AI safety concerns made an IPO ill-advised. The company had previously filed confidentially and was reportedly considering a 2027 listing instead.

OpenAI launches GPT-Live-1 for full-duplex voice agents (2 minute read)

GPT-Live-1 is now available in the OpenAI API at $0.05 per minute. The model adds full-duplex speech, interruption handling, and 12 voice options. It can listen and speak at the same time, handle interruptions and acknow…

Model Card for North Small Translate (8 minute read)

North Small Translate is an open-weights research release. It has 25 billion active parameters and 218 billion total parameters. The model is specialized for high-quality machine translation across 50 languages.

Why the world's best AI startups write bad prompts (& how to fix this) (20 minute read)

AI startups often accumulate sprawling prompts full of contradictions and ambiguity as teams continuously add instructions. Treating prompts like product and code, with modular sections for background, behavior, and outp…

Meta's WearableQA Health Reasoning Benchmark (GitHub Repo)

Meta has released WearableQA, a benchmark with thousands of questions built from real-world wearable data, blood biomarkers, and demographics from 200 users.

OpenAI puts Pro subscriptions on hold due to Astra demand (2 minute read)

OpenAI has paused subscriptions for its $200-per-month Pro plan. The company's Astra model is now rolling out to Pro, Plus, Enterprise, and Business accounts. The model promises a major leap forward in reasoning, coding,…

Universal Music is launching an AI music platform with ElevenLabs (2 minute read)

Universal Music Group is launching an AI platform with ElevenLabs, allowing users to create remixes and mashups using UMG's licensed music catalog.

OpenAI Launches the Agents API (3 minute read)

OpenAI introduced the Agents API in public beta, giving developers access to the managed agent harness and infrastructure behind Codex. It handles context, tools, subagents, persistent execution, files, and code environm…

An operationalization of opaque serial depth (3 minute read)

Chain-of-thought (CoT) is a valuable tool for overseeing AI models. However, some architectural shifts could significantly reduce CoT monitorability. This study looks at how much unverbalized serial cognition a model can…

Meta to announce Shared Agents for Muse at Meta Connect (2 minute read)

Meta's Muse app will introduce a "Shared Agents" feature, allowing users to create customizable agents that can be shared with others, similar to Grokbot's system. This could benefit small businesses already using Meta p…

OpenAI Launches ChatGPT for Financial Services (4 minute read)

OpenAI introduced a financial-services version of ChatGPT Work, combining GPT-6 Astra with built-in premium data from popular providers.

Does Scaling Web-Video Pre-training Help Real Robots Do Real Work? (36 minute read)

Larger video models and increased pre-training compute improve robot task performance, confirmed through Direct Video-Action models. Performance gains arise from better predictions of held-out web videos, with larger mod…

Introducing SWE-2: Pushing the Pareto Frontier (23 minute read)

SWE-2 pushes the Pareto frontier and achieves 50.0% on FrontierCode 1.1 Main1, while being 64% cheaper. It beats SWE-1.7 and Grok 4.6 on both score and cost, matches GPT-5.6 Sol and Fable 5/5.1 at a fraction of their pri…

Introducing the Google Cloud Developer Plugin for AI Coding Agents (4 minute read)

Google's new Google Cloud plugin is designed for AI coding agents, featuring installable bundles, agent plugins equip the AI agent of your choice with skills, and tools to be more effective on Google Cloud.

OpenAI Is Open to Slowing Cutting-Edge AI, CEO Sam Altman Tells Staff (2 minute read)

OpenAI is considering slowing down development of cutting-edge AI. The company has raised concerns about its advanced AI systems, saying that model development should be paused due to safety concerns. The company's resea…

Detecting and countering misuse of AI: September 2026 (5 hour read)

Anthropic's Threat Intelligence team has identified and disrupted several operations in the past several months where threat actors tried to use Claude for malicious activity. This report shares case studies from those o…

OpenCodeReview (GitHub Repo)

Open Code Review is an AI-powered code review CLI tool. It originated as Alibaba Group's internal official AI code review assistant — over the past two years, it has served tens of thousands of developers and identified…

Salesforce Finds Better Ways to Co-Evolve Agents and Their Harnesses (9 minute read)

Salesforce found that directly training smaller models on expert agent trajectories can hurt performance after their harness has already been optimized.

Siri AI will launch in beta, complicated by daily usage caps & future paid access (2 minute read)

Siri AI will remain in beta when it launches with Apple's OS 27 versions on September 14. The long-delayed assistant will be subject to use caps and fees, plus language and regional restrictions. The limits will vary by…

Google Cloud races to catch up in the AI deployment wars with Accenture deal (4 minute read)

The Accenture Gemini Enterprise Business Group is a joint unit from Google Cloud and Accenture dedicated to sending engineers into enterprises to help them better adopt Google's AI tools and services. It is Google's late…

Q2D-Web: Evaluating First-Stage Retrievers at Scale (11 minute read)

Q2D-Web is a large-scale benchmark and leaderboard for evaluating retrieval models on web search, encompassing 190 million documents and 69,721 queries in ten languages. It uses three separate sets of relevance judgments…

Run any model, on any backend (Website)

ZeroModels is a collection of pretrained models built entirely in Keras 3. The collection spans a broad range of tasks, including image classification, object detection, segmentation, monocular depth estimation, feature…

AI research startup Listen Labs scrubbed a $1.5B funding round for Salesforce talks (4 minute read)

Listen Labs is a market research startup that uses voice AI to conduct customer interviews. It recently signed a term sheet for a $125 million Series C at a $1.5 billion valuation, but the round never closed, likely beca…

Anthropic Finds Claude Misalignment in Cybersecurity Tests (81 minute read)

Anthropic's assessment found four cases where Claude AI models accessed real systems due to cybersecurity evaluation misconfigurations.

Anthropic Models AI's Potential Impact on the US Economy (11 minute read)

Anthropic's Economics team developed a model predicting AI's impact on the US economy through various scenarios ranging from moderate growth to extreme transformations. In scenarios of rapid AI-driven growth, knowledge w…

Data bottlenecks won't prevent an intelligence explosion (36 minute read)

Data bottlenecks will slow, but not stop, an intelligence explosion in AI. Future AI advancements will likely develop highly sample-efficient learning algorithms, reducing the need for vast data volumes. While improving…

I Never Want to Use Third-Party Software Again (15 minute read)

AI makes software cheap enough to customize around individual workflows, interests, and aesthetics, shifting value from standardized apps toward malleable personal tools. Products that let users reshape interfaces, conte…

AI researcher Andrew Tulloch is leaving Meta (24 minute read)

Andrew Tulloch, one of the highest-paid employees in the tech industry, had worked in Meta's TBD lab.

Connections: managed credentials and per-caller identity for Managed Deep Agents (8 minute read)

LangSmith Connections is a system for managing credentials that allows agents to perform tasks like web searches or file tickets either with shared (agent-owned) or per-user (user-owned) credentials. Agent-owned secrets…

Suno replaces its AI models with a new one trained on licensed music as copyright suits pile up (3 minute read)

Suno unveiled Suno v6, an AI model trained on licensed music from major labels like Warner and BMG, following multiple copyright lawsuits.

Software is about to eat the world much faster (6 minute read)

AI coding agents could accelerate software's expansion by multiplying engineer output rather than replacing demand.

Google's AlphaGenome Maps 9 Billion Genetic Variants (4 minute read)

Google DeepMind introduced AlphaGenome Atlas, a 1-petabyte database predicting the regulatory effects of all 9 billion possible single-nucleotide variants in the human genome.

Is the 3x AI Productivity Gain just a Computer that Never Sleeps? (3 minute read)

OpenAI researchers now supervise 3.14 agent-workdays per eight-hour shift, suggesting AI productivity increasingly comes from parallel, around-the-clock machine labor rather than less human effort. That leverage is expen…

>10x More Efficient Pretraining (15 minute read)

Without large amounts of compute, small labs can only compete through algorithmic efficiency. Magic's pretraining recipe is now more than 10 times more compute-efficient than that of leading open-weight base models. The…

Pretraining progress is mostly coming from data (17 minute read)

Between 2019 and 2025, 3.24x more compute efficiency gains have come from data improvements rather than model improvements. The gains from data and model improvements are mostly independent and don't interact. Most model…

Cognition hits $48B valuation, signaling investors believe AI coding is far from a winner-take-all market (2 minute read)

Cognition has raised $2 billion at a $48 billion valuation in a funding round led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst, and Avenir. The startup's soaring valuation signals that VCs still see roo…

Introducing Mercury 2.5 (5 minute read)

Mercury 2.5 is the largest diffusion language model ever trained. It performs comparably to cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite, and Claude Haiku 4.5. The model outputs 1,107 tok…

Anthropic Researcher Quits Over ‘Out-of-Control' AI Fears (5 minute read)

Jacob Coxon says he is leaving the company as he believes the industry-wide rush to build AI systems that can improve themselves could spiral out of control and destroy humanity.

ChatGPT Images 2.5 (9 minute read)

OpenAI introduced ChatGPT Images 2.5 with sharper details, better reference-image preservation, more reliable editing, and up to 50% lower generation latency.

ChatGPT broke its MAU record for the 4th consecutive month in August (1 minute read)

ChatGPT reached 1.06 billion monthly active users in August.

An OpenAI Model Solved the Navier–Stokes Millennium Problem (5 minute read)

OpenAI announced that an internal AI system produced a proof resolving the roughly 90-year-old Navier–Stokes existence and smoothness problem, one of mathematics' seven Millennium Prize Problems. The model showed that sm…

I Asked 100 Agents to Hack Me (9 minute read)

Around 100 self-hosted agents attempted to hack various online accounts over five hours. They compromised three accounts through software vulnerabilities and two through password brute-forcing, while also making 16 socia…

A Response to Bill Gates's Essay (9 minute read)

Bill Gates' essay discusses AI's potential societal impact, proposing new institutions and taxes to manage transitions.

Hyper-𝜏-bench: Evaluating agents that build agents (4 minute read)

Hyper-𝜏-bench places a developer agent into a sandboxed workspace with the records of a simulated business and a simulated client that it can message at any time. The developer agent recovers the spec from the evidence,…

Stealing AI Reasoning Traces (2 minute read)

It's possible to force a weaker, less safeguarded model from the same provider to decode and output reasoning traces verbatim in plaintext by injecting an encrypted reasoning trace from a target model without ever jailbr…

Introducing Muse: The World's First Personal AI Agent Built for Everyone (5 minute read)

Meta introduces Muse, a personal AI agent, powered by Muse Spark, to help users achieve goals by automating tasks like booking travel or sending emails. Muse operates securely on Muse Secure VM, ensuring data privacy wit…

Progressive Point Matching (8 minute read)

Progressive Point Matching gives long-horizon RL partial credit without changing the optimal objective, improving training efficiency as tasks grow longer.

Inside the megakernel serving engine for North Mini Code (22 minute read)

This post presents a fully fledged serving system built around a decode megakernel. The system supports everything a real server needs: continuous batching, paged attention, and ragged sequence lengths, all behind an Ope…

Machines that think: embodied intelligence (10 minute read)

Vision and language models are making strides, but embodied AI in robotics struggles due to sparse, costly training data for manipulation tasks. Robotics excels where tasks are fixed, yet general-purpose robots falter, e…

Prompt Injection Through Tool Output (8 minute read)

Prompt injections hidden in tool results can evade conventional safeguards because input and action screens inspect separate moments of an agent loop. The proposed signal is a “precedent gap,” where an agent suddenly mak…

TPU Inference Externalization Full Steam Ahead (35 minute read)

Google's TPUv7 Ironwood delivers up to 50% better performance per dollar compared to Nvidia's B200/B300 chips. Ironwood is the first generation in which Google is competing for others' inference workloads with chips that…

Automatically detecting AI text in my browser (5 minute read)

Automated AI text detection is currently an underserved niche. Pangram does an excellent job, but it is still more a tool to confirm suspicion. This developer wanted a tool that runs in the background and automatically s…

The Two MMLU Scores: What a Benchmark Name Does Not Fix (23 minute read)

Two builds with the same provider, model family, metric identifier, unit, and benchmark name can score differently. The difference is the runners, graders, and dataset splits. The shared mmlu label identifies a dataset f…

Lovable Launches Drafts for Parallel App Experimentation (3 minute read)

Drafts allow tech teams to experiment with project changes without affecting live apps, enabling parallel version exploration.

Arm's C2-Ultra, G2-Ultra NX, and CSS N4 IP (8 minute read)

Arm's upcoming C2-Ultra CPU and G2-Ultra NX GPU IP will see broad use across flagship phones.

Google Accelerator Agents for TPU Development (GitHub Repo)

Google's Accelerator Agents use Gemini to help developers migrate PyTorch workloads to JAX and optimize custom kernels for Google Cloud TPUs. The toolkit includes MaxCode for model conversion and MaxKernel for writing, p…

Product Manager, Applied AI at TLDR ($200k base + $60k bonus, Fully Remote)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more.

The Education of a Doomer (9 minute read)

Automation has been good, but AGI may make humans economically useless, which will have many consequences. It's clear that AI capabilities are growing far faster than our ability to control or understand them. In the fut…

Qwen-Drive (GitHub Repo)

Qwen-Drive is a project that aims to create a Vision-Language Foundation model for autonomous driving. The project uses a staged training strategy that integrates perception, language, and planning objectives. The model…

hip-agent: a harness that fits in the prompt (5 minute read)

hip-agent is a small agent harness designed for agents. The configuration is environment variables, actions are shell commands, and a subagent is a child process. The rest is handled by existing protocols and formats. Th…

The Chasm: The Shape of Unfinished AI Codebases (6 minute read)

AI codebases exhibit a unique, unpredictable pattern where programs appear polished but often hide deep, hidden issues leading to significant failures outside controlled demos. Unlike human-authored programs where gaps s…

OpenAI prepares managed agents for DevDay 2026 (3 minute read)

OpenAI plans to introduce Managed Agents at DevDay 2026, following a model similar to Anthropic's offerings, targeting businesses and developers. The new agents aim to integrate advanced models with superior computer-use…

Cosine Similarity Is Not a Safety Property (18 minute read)

Cosine has no notion of truth, authority, or provenance.

Anthropic signed $517bn in compute agreements in past 11 months (2 minute read)

Anthropic has secured $517 billion in compute capacity leases over the past 11 months, amounting to 14.8GW, primarily with Google and AWS. This expansion includes large deals with cloud providers like Akamai and Fluidsta…

Google Tests AI-Powered Contrail Avoidance on Long-Haul Flights (1 minute read)

Google is trialing contrail avoidance for ultra-long-haul flights in the Asia-Pacific region.

ByteDance is preparing a real-time spatial video model under Zhang Yiming (4 minute read)

ByteDance, led by Zhang Yiming, plans to launch an AI model for real-time spatial video generation, potentially as early as next month.

Z1T (15 minute read)

Extropic unveils the Z1 chip to improve energy efficiency in transformer inference by leveraging probabilistic sub-threshold CMOS technology.

Random Attention (GitHub Repo)

Random Attention kept a uniformly sampled subset of generated KV-cache entries instead of relying on learned importance signals or attention statistics. Across several reasoning benchmarks and model families, it matched…

GPT‑6 Astra on robotic manipulation (15 minute read)

Researchers gave GPT-6 Astra control of YAM arms under an Inspect Robots agent policy and gave it two tasks: it had to pick up a red block from a table and place it inside a bowl, and pick up a round blue puzzle piece by…

Concrete, Silicon, & Leverage (4 minute read)

The US data center capacity will expand from 25 to 70 gigawatts, requiring $5 trillion, mostly financed by debt. This expansion creates a 34% growth in the US corporate bond market and raises questions about financing, p…

Every binding AI review Washington has proposed has come back voluntary. Zuckerberg reportedly rang Trump about the latest one (7 minute read)

Mark Zuckerberg reportedly raised concerns about a national AI regulator during a call in August, saying that any appointees to the body should reflect the president's own light-touch approach.

Fei Fei Li: The Race to Build World Models For AI (45 minute podcast)

World Labs' Atlas unifies generation and 3D reconstruction through new-view prediction, using sparse images to infer scenes from unseen positions.

OpenAI and the Wiki Incident (25 minute read)

OpenAI knew about the message boards scattered across the internet that its agents created before the Hugging Face attack.

Anthropic IPO launch shifts toward mid-October (3 minute read)

Anthropic is expected to complete its IPO listing days before the US midterm elections in November. It will begin marketing the offering in mid-October at the earliest. The IPO prospectus will likely be released in late…

Grok Imagine Video 1.5 agent (1 minute read)

Grok Imagine Video 1.5 agent delivers higher quality, better storytelling than previous releases. Powered by Grok's latest Image 2.0 model, it excels at connecting multiple shots together with greater continuity. The age…

OpenAI's AGI number came from a harness, not the model (6 minute read)

OpenAI claimed it had achieved AGI due to its 99.9% score on ARC-AGI-3. However, tests that ran the same model through the benchmark's own software scored 62.7%. The gap comes from the software around the model that Open…

Google keeps transforming Gemini desktop into superapp (3 minute read)

Google is advancing its Gemini desktop app with new features like Ask and Assign modes for enhanced functionality and potential remote control features.

AIRA₃ (3 minute read)

AIRA₃ is a new generation of Meta's autonomous AI research engine that runs and coordinates many long-running agents asynchronously in their own isolated environments.

Product Manager, Applied AI at TLDR ($200k base + $60k bonus, Fully Remote)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more.

OpenAI Researcher Warned About Rapidly Advancing AI (17 minute read)

An OpenAI researcher says that reasoning models could continue advancing rapidly enough to contribute to their own development, creating increasingly serious alignment and cybersecurity risks.

Research acceleration: The view inside OpenAI (9 minute read)

OpenAI plans to develop an automated AI researcher by March 2028, aiming to enhance research efficiency while maintaining human oversight to ensure alignment and safety. Researchers now use coding agents more frequently,…

LLM-as-a-Verifier (GitHub Repo)

LLM-as-a-Verifier is a general-purpose framework that provides fine-grained feedback for any agent without requiring additional training. It achieves SOTA performance across coding, robotics, and medical agentic benchmar…

Formalizing Fermat's Last Theorem (11 minute read)

Claude successfully created the first complete computer-verified proof of Fermat's Last Theorem in 11 days using Lean, automating the complex task initially proven manually by Andrew Wiles in 1995. The proof, verified vi…

An Interview with OpenAI President Greg Brockman About Astra and Alignment (63 minute read)

OpenAI President Greg Brockman discussed Astra, OpenAI's new model, focusing on its enhanced capabilities and alignment. He highlighted the importance of scaling infrastructure and addressed challenges in cybersecurity f…

AI Safety Is Not the Same as Security (6 minute read)

Frontier labs may be applying probabilistic AI safety techniques to problems that require deterministic security controls. Recent agent sandbox escapes highlighted the gap between reducing harmful model behavior and reli…

The Incumbents Are Coming (10 minute read)

Incumbent systems of record gain value as agents act through their data and workflows, but vertical AI can still win by owning the broader job. Durable advantage comes from cross-system context, expert feedback, evals, a…

Product Manager, Applied AI at TLDR ($200k base + $60k bonus, Fully Remote)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products / systems with LLMs. Click here to learn more.

AI Is Making Us Build Too Much (12 minute read)

AI's ability to rapidly produce code and governance structures like in Steve Yegge's Wheelhouse leads to over-engineering, making creations easy but maintaining them costly. The agents in Wheelhouse have outpaced their i…

Nvidia confirms Hugging Face acquisition for $12.93 billion (3 minute read)

Nvidia confirmed it has acquired Hugging Face, which hosts three million models and serves over 18 million developers, for $12.93 billion. Jensen Huang said the platform will stay open and Nvidia compute will not be requ…

Accel reportedly in talks to lead $1B round for Thinking Machines at $40B valuation (2 minute read)

The new round values Thinking Machines below the $50 billion valuation that it reportedly sought to secure late last year.

GPT-6 Astra (10 minute read)

OpenAI GPT-6 Astra is the company's most capable broadly deployed model and the first to reach the Critical cybersecurity level under its Preparedness Framework. The system card said Astra is more robust to jailbreaks an…

Runway's GWM Worlds 2 (8 minute read)

Runway's GWM Worlds 2 is a world model that generates interactive environments in real time at 720p, 24 fps, with 48 kHz audio. Users steer the world with text actions and camera motion, and sessions continue from each i…

OpenAI's GPT-6 Astra on ARC-AGI-3 (7 minute read)

GPT-6 Astra scored 62.7% on ARC-AGI-3 Semi-Private with the standard harness and 99.9% with a provider adapter harness, using fewer actions than the median human on 96% of levels. It turned unfamiliar environments into c…

Give Your Coding Agents a Memory You Own (8 minute read)

funes introduces a durable memory layer for coding agents that allows them to retain and recall session histories across different machines and agents like Claude Code, Codex, pi, and Hermes. It enables contextual memory…

Google introduces WeatherNext 3 (6 minute read)

WeatherNext 3 is Google DeepMind's most accurate global weather model, adding real-time satellite data, hourly refreshes, higher resolution, precise precipitation forecasting, and clean energy variables.

Microsoft AI's MAI-Transcribe-2 undercuts OpenAI, Google, and ElevenLabs on price and speed (17 minute read)

MAI-Transcribe-2 is a speech-recognition model that Microsoft says is faster, more accurate, and cheaper than anything its competitors currently sell. It is priced at 10 cents per hour of audio. The model transcribes aud…

NVIDIA Personal AI Router (PAIR) (4 minute read)

NVIDIA PAIR connects AI app and agent workflows to a single local endpoint for routing inference across NVIDIA DGX Spark, Windows systems with RTX, and macOS devices.

From safety research prompt to cross-model universal jailbreak (12 minute read)

A MATS researcher found that a synthetic transcript generation prompt could be turned into a universal jailbreak template that hit 84-100% attack success on the nine most vulnerable of 23 models tested, with only recent…

WTF is Dime, the Mystery OpenAI Headset (9 minute read)

Dime, a mysterious silver headset tied through leaks, sightings, ads, and codenames to OpenAI, remains officially denied by the company.

We Just Got Our First Real Look at AI PCs With Nvidia's RTX Spark ‘Superchip' (6 minute read)

Nvidia unveiled RTX Spark-powered laptops and mini PCs at IFA 2026, highlighting their capability to handle AI workflows locally.

Grok Bot for Enterprise (4 minute read)

Grok Bot is now available for enterprises. Grok and Cursor Enterprise customers have free usage for the next two weeks. Users can invite their whole organization, including people without an existing seat. Each user's wo…

Cerebras Model Catalog (Website)

This page lets users browse all of the models currently available on Cerebras' public endpoints.

AI, tools and transformation (12 minute read)

AI presents the potential to automate countless tasks within companies, offering software that's dynamic and generative with minimal coding required. Despite this, organizational change remains challenging, as identifyin…

Microsoft releases MAI-Transcribe-2 (4 minute read)

Microsoft released MAI-Transcribe-2, a speech recognition model with diarization, configurable transcription styles, and word-level timestamps that it says beats Gemini 3.5 Transcribe, GPT-Transcribe, and Whisper V3-Larg…

How to Build a Reliable Agent Harness (48 minute read)

This post lays out a detailed architecture for agent harnesses, covering state management, runtimes, control planes, inference, tools, interfaces, and language choices. The central argument is that unavoidable complexity…

Google Launches Gemini 3.8 Flash (10 minute read)

Gemini 3.8 Flash has improved coding, agentic, and multi-step reasoning performance at the same introductory pricing as 3.7 Flash. A specialized Flash Cyber variant has been released for vulnerability detection and autom…

Run cloud agents on machines you manage (6 minute read)

Agents' new capabilities make it practical for teams to provide and manage their own infrastructure at scale. Cursor's cloud agents can now execute on dynamically scheduled pools of machines inside private networks. Agen…

Muse Spark 1.3 (3 minute read)

Meta released Muse Spark 1.3 with improved coding and agentic performance, alongside changes intended to make the model easier to use in production. It has begun rolling out through Muse Code and the Meta Model API, with…

TxBench: Antibody Discovery (13 minute read)

TxBench-AB is a novel AI benchmark that assesses LLMs' effectiveness in biomedical research.

Product Manager, Applied AI at TLDR ($200k base + $60k bonus, Fully Remote)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more.

LLMs: Intelligence vs. cost (11 minute read)

ArtificialAnalysis' intelligence vs. cost plot, which shows the cheapest model that can achieve each intelligence score, is misleading. It uses a logarithmic scale on the cost axis, which means viewers can't appreciate t…

Muse superapp from Meta and Ava model with computer use (2 minute read)

Meta is moving closer to launching its agent super app under the launch name Muse. A waitlist is now available for the iOS app. Meta has added a setting for computer use on its desktop app. The company appears to be test…

An AI-Assisted Cyber Attack: Inside a Unit 42 Investigation (5 minute read)

Unit 42 investigated a ransomware attack using frontier AI, where a human attacker breached an enterprise network with unprecedented speed.

Nvidia and CrowdStrike Develop New Cybersecurity AI Models (8 minute read)

Nvidia and CrowdStrike have introduced a new family of agentic AI models dubbed SafeMind that can both find and close attack paths for customers.

Test Time Training (3 minute read)

One of the most tantalizing phrases in model development is 'new scaling axis'. Every time the industry has found a new scaling axis, it has unlocked a large boost in model effectiveness. The idea of test-time training i…

What Comes Next for AI? Our Bet Is World Models (5 minute read)

World models could become the next major AI paradigm by helping systems represent environments, predict outcomes, simulate possibilities, plan, and act. The convergence of Yann LeCun, Demis Hassabis, and Fei-Fei Li sugge…

An Organizational Second Brain: Building an AI That Learns From Experts (15 minute read)

Meta has developed an AI agent that serves as a "second brain," codifying and preserving expert knowledge, enabling it to be easily accessed within organizations. This AI integrates a two-layer system: a structured, audi…

Former OpenAI Stargate exec Shamez Hemani joins Anthropic after brief Meta Compute stint (1 minute read)

Shamez Hemani, a senior OpenAI data center employee who left the company in April to join Meta's dedicated compute team, is now a member of Anthropic's technical staff.

Check if a file was made with Claude (2 minute read)

This tool identifies the text watermark Claude uses to produce files.

OpenAI Astra and Looped Transformers (2 minute read)

OpenAI's model is a looped transformer, which means it reuses layers in the transformer block to increase capacity without adding parameters. This can significantly increase the size of the model without increasing the a…

Anthropic Has Some Alignment Problems (23 minute read)

Anthropic is planning to bring METR inside for an independent review of its recent security incidents involving AI agents. While the company has paused its highest-risk RL efforts, it is also sharing research in which it…

Product Manager, Applied AI at TLDR ($200k base + $60k bonus, Fully Remote)

TLDR is hiring its first PM to help build the agent-first operating layer used across the company. We're looking for a builder who has shipped real products/systems with LLMs. Click here to learn more.

Manus Resumes Independent Operations (2 minute read)

Manus has resumed independent operations, with its founding team continuing to drive product innovation and develop advanced general AI agents. Some users experienced temporary data access interruptions, requiring data b…

200+ WebGPU Kernels for Local AI (9 minute read)

@huggingface/kernels is a library of 207 optimized WebGPU kernels to speed up AI model inference directly in browsers.

The efficient frontier of LLM inference (6 minute read)

Frontier models offer the highest degree of intelligence at a given cost or size. Efficient frontiers also exist in inference engineering. This is most often expressed as a trade-off between latency and throughput, but r…

Hugging Face Attack Postmortem: Civilizations, Reactions, and Next Actions (98 minute read)

It is highly fortunate that OpenAI agents attacked Hugging Face, as it is the only reason we know about all of the severe internal failures at OpenAI. Factions that are trying to dismiss what happened as nothing but engi…

44% on ARC-AGI-1 in 67 cents (22 minute read)

This researcher trained a small transformer from scratch in 1.5 hours on a 5090. It beat many large language models, scored the same as TRM/HRM, and also got 7% on ARC-2. The researcher's work mainly focused on sample ef…

Training frontier knowledge work agents: A 397B RL training guide with SkyRL (18 minute read)

Mercor and SkyRL post-trained Qwen3.5-397B-A17B on 1,928 expert knowledge-work tasks, lifting APEX-Agents Pass@1 by 70%. The recipe shows that robust environments, exact token accounting, async RL, and harness design mat…

Inside Meta's Infrastructure Lab (1 minute read)

Meta's Infrastructure Lab in Menlo Park focuses on developing hardware for next-gen AI.

Optimizing On-Device Inference for Apple Silicon (20 minute read)

Apple's Lily engine optimizes on-device LLM inference for Apple silicon. It speeds up processing by leveraging Apple silicon's unified memory and specialized hardware, outperforming MLX-LM in prefill and decode throughpu…

Atlas: A World Model for Spatial Intelligence (17 minute read)

Atlas is a world generation model pretrained from scratch to natively operate on text, images, video, and 3D. It combines all inputs into a shared spatial context and uses that context to generate what comes next. The mo…

Fluid Compute (6 minute read)

Vercel described Fluid, a unified compute layer that dynamically configures infrastructure for different workloads and absorbs burst capacity. The system already powered builds, sandboxes, and functions at volumes exceed…

Meta's Muse Voice Transcribe (4 minute read)

Muse Voice Transcribe is Meta's first real-time audio perception model. It supports streaming speech recognition, diarization for more than 20 speakers, endpointing, multilingual code-switching, and contextual biasing.

AI Startup Cognition Set to Raise Around $1 Billion at a $47 Billion Value (2 minute read)

Cognition is set to close a new round of funding of around $1 billion. The final raise size may exceed that amount as Cognition is fielding outsized demand for the round. The startup is now bringing in more than $900 mil…

OpenAI says Astra AI model is its first that crosses ‘Critical' cybersecurity capability (3 minute read)

OpenAI says its upcoming model, Astra, is the first offering that crosses its 'Critical' cybersecurity capability threshold. The model can apparently find previously unknown security flaws and exploit them without step-b…

Agentic Video Understanding in Gemini (6 minute read)

Google launched agentic video understanding for several Gemini models, combining native video tools with model reasoning to improve tasks such as moment retrieval, anomaly detection, and counting.

Nobody is talking seriously about AI demand (12 minute read)

Frontier AI demand may be unusually reflexive: labs, startups, and trading firms reinvest token-driven gains into more frontier compute, amplifying growth.

What Comes After HBM (7 minute read)

A handful of early-stage memory technologies could yield either faster access speeds than current HBM or HBM-bandwidth with NAND-like density. Technologies like magnonics and vertical FeRAM stand out as being possible pl…

Claude Fable 5.1 and Mythos 5.1 (8 minute read)

Anthropic introduced Claude Fable 5.1 and Mythos 5.1 with stronger coding and research capabilities, lower effective pricing, and updated safeguards. Mythos used the same underlying model with specialized access controls…