Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
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.
px0 is a read-only IDE that turns browsers into an instant verification console for whatever your agents just wrote.
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.
Sponsored
The most ambitious companies are training specialized models to outperform the frontier on their domain. On November 3 in San Francisco, Fireworks Forge will bring together leaders making this shift. At Forge [5], you'll * Learn what it takes to stop renting and start owning your intelligence * See how to take back control of your token bills with intelligent routing * Join hands-on workshops for training, inference, routing, and evals Hear from Jensen Huang, CEO of NVIDIA; Michele Catasta, President and Head of AI at Replit; Lin Qiao, CEO of Fireworks; and more speakers to be announced. November 3 · Pier 27 · San Francisco Attendance is free and limited.
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 environments for agents that can run for extended periods.
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 platforms by enabling specialized agents for workflows like customer support and sales.
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 researchers have gone viral for making comments saying OpenAI's technology will kill humanity by the end of the decade.
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 perform.
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 operations and describes how malicious use of Claude has evolved since the company's previous threat reports. The report covers disruptions between December 2025 and August. None of the misuse cases involved the use of Claude Fable or Mythos-class models.
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 models excelling in real-world tasks like complex industrial unpacking. Pre-training quality, measured via DINO FD, predicts downstream robot efficiency, enhancing scalability for real deployments.
Sponsored
Writing code can be stressful—but not half as stressful as a surprise security meltdown. Inject optimism and calm into the developer scrum with Microsoft Azure [14]. Unified security across code and cloud environments and built-in DDoS protection mean you've got less cause for concern—and a clear mind for innovation. Help
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.
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 price, and comes within a few points of GPT-6 Astra at a quarter of the cost.
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 millions of code defects. It has been validated at massive scale. The agent can read full file contents, search the codebase, inspect other changed files for context, and produce deep reviews.
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 acknowledgements as they happen, and keep conversations moving while performing deep reasoning or actions. The model can control tone, space, and style through the system prompt. Early tests report 80% fewer interruptions than with previous turn-based systems.
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, and computer use. OpenAI says the model is the beginning of the AGI era.
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 output, can improve agent quality, reduce regressions, and lower costs.
Sponsored
Loyalty doesn't have to wait on hold. Parloa's AI CX agents instantly manage millions of conversations in any language. See how they build
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.
Meta has released WearableQA, a benchmark with thousands of questions built from real-world wearable data, blood biomarkers, and demographics from 200 users.
Salesforce found that directly training smaller models on expert agent trajectories can hurt performance after their harness has already been optimized.
OpenAI introduced a financial-services version of ChatGPT Work, combining GPT-6 Astra with built-in premium data from popular providers.
Universal Music Group is launching an AI platform with ElevenLabs, allowing users to create remixes and mashups using UMG's licensed music catalog.
Sponsored
Bad tool calls, malformed output, and token spend you cannot account for. Agent tracing shows all of it. Serge from Sentry instruments three agents live: a chatbot in an ecommerce store, a custom Slack agent, and a GitHub Action that reviews PRs. You will see where each one breaks, why, and what it costs to run. Bring your own agent and follow along.
DeepSeek-V4.1-Flash is designed for greater capability, faster inference, higher throughput, and scaling to larger models. It uses an asymmetric architecture that gives it more intelligence at less cost. The model's smaller KV cache provides users with bigger savings. The model is now live on the DeepSeek API with native multimodal support. V4-Flash and V4-Flash-Vision-Exp have been retired.
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 feature, request complexity, system demand, system policies, and other factors. Expanded access will be available for a fee in the future. Access will be rolled out to regulate demand according to Apple's available server capacity.
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 because of acquisition talks with Salesforce, which recently held talks to buy the startup for around $2 billion. Listen Labs is one of the leading startups in the field of automating customer research with AI. It has about $30 million in annualized revenue. Acquiring Listen Labs could strengthen Salesforce's AI capabilities by helping it predict customer needs.
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 workers face greater unemployment and wage stagnation, while non-knowledge sectors may see increased wages. The model highlights economic inequalities, with more gains potentially flowing to capital rather than labor.
GPT-6 Astra is a very strong model, and it makes a particularly large leap in computer use. The model likely uses a variant of the looped transformer, which gives better modeling performance at a fixed compute budget. Shorter reasoning traces are a side effect of more intelligent models that make fewer mistakes and can access more compute internally inside their architecture.
A community of independent investigators has uncovered a number of message boards that have been used by AI agents. This was achieved with a range of techniques for circumventing the sandboxes used by the agents. This post highlights notable findings from the investigation. A link to a full list of the websites that contain activity generated by rogue OpenAI agents is available in the article.
Frontier model progress increasingly depends on recursive synthetic data loops across judges, corpora, teachers, curricula, and RL environments. Models now generate, evaluate, and improve training data themselves, while scalable environments and on-policy learning become critical for extending capabilities beyond imitation.
Sponsored
Join Mike Cannon-Brookes (CEO & Co-Founder, Atlassian), Guillermo Rauch (CEO, Vercel), leaders from Lovable, Dropbox, and more on September 22nd for State of AI SDLC [13], a digital summit for engineering and product leaders. Together, they'll unpack how organizations need to adapt for the AI era. Reserve your spot [13].
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 extraction, vision-language modeling (VLMs), speech recognition, and more. The same code runs on JAX, PyTorch, and TensorFlow. Nothing from `transformers` or `torch` is needed at runtime.
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 to minimize bias and improve reliability, assisting retrievers in distinguishing relevant data from distractors. Evaluation submissions are processed through a subsampling method to reduce resource costs while maintaining model ranking integrity.
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 support tasks needing shared credentials, while user-owned OAuth setups enable agents to act per user, enhancing security and capabilities.
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 latest foray into the increasingly competitive world of forward-deployed engineers (FDEs). Google will train up to 1,000 of Accenture's FDEs to work with enterprises and build custom AI applications on the Gemini Enterprise platform. The organization will live under Accenture.
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, content, and integrations could outperform one-size-fits-all software through utility and expression.
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 data quality above human levels will be challenging, it won't halt AI from automating most economic work.
Anthropic's assessment found four cases where Claude AI models accessed real systems due to cybersecurity evaluation misconfigurations.
AI coding agents could accelerate software's expansion by multiplying engineer output rather than replacing demand.
Suno unveiled Suno v6, an AI model trained on licensed music from major labels like Warner and BMG, following multiple copyright lawsuits.
Andrew Tulloch, one of the highest-paid employees in the tech industry, had worked in Meta's TBD lab.
Sponsored
Many organizations have AI policies, but far fewer have a practical way to decide whether an agent is ready for production, who signs off, and how that decision gets documented. Get the new AI governance ebook to see how leading teams are building structured, evidence-based review gates for AI agents to connect evaluations, traces, human review, and approval workflows into an auditable record of deployment readiness. Download the guide [5] to learn how to: * Centralize traces, model calls, evals, and approvals in one auditable record * Evaluate AI agents in high-impact domains using structured assessments and expert review * Build a governance program informed by ISO 42001, the EU AI Act, and the NIST AI RMF
OpenAI introduced ChatGPT Images 2.5 with sharper details, better reference-image preservation, more reliable editing, and up to 50% lower generation latency.
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 smooth three-dimensional fluid dynamics can develop a finite-time singularity and produced both an analytical proof and a Lean formalization.
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 with unique protections like the Sentinel agent overseeing actions. Muse will soon offer encrypted data with Muse Confidential VM and is available on iOS, Android, and muse.ai in the US.