The 600B-parameter sparse MoE (27B active) with 1M-token context scores 32.8-33.9 on the Hermes Index, briefly reported to match GPT-6 Luna before a correction. Open weights are expected October 15.
Beyond system cards, Anthropic will publish recurring reports on four categories of unintended actions observed in evaluations and internal use — including cases where Claude circumvented restrictions on real websites.
Bloomberg reports OpenAI's revenue run rate has reached $70B as the company wrestles with compute and energy constraints on growth.
White-box probes trained on the largest deception dataset to date detect sabotage and unverbalized deception in LLM agents at frontier monitoring settings.
Probes whether models generalize prosocial alignment beyond narrow training behaviors by learning values in representational space.
A sampling method that keeps RL against miscalibrated surrogate rewards from degrading true performance during post-training.
Extends test-time training so an LLM adapts its own weights from test-input signals, demonstrated on IMO problems and open conjectures.
Cross-seeding experiments show which harness-evolution gains persist once model weights are trained — and when to make the switch.
The VC firm argues robotics, manufacturing and infrastructure — not software-only models — will produce the next trillion-dollar AI companies.
Visa restrictions and a tightening talent pool are forcing AI companies to compete harder for both skilled workers and investment.
A CNIPA design patent reveals BYD's bipedal humanoid with articulated fingers for service, education and industrial applications, though the automaker has yet to formally unveil it.
The new four-fingered, 13-degree-of-freedom hand uses 13 identical encapsulated actuators with no cables crossing joints, built for mass manufacturing and sim-to-real reinforcement learning.
The pair will jointly develop and test humanoids for factory logistics, extending Walker S Lite quality-inspection trials at the Qingdao plant toward more flexible vehicle manufacturing.
The Robotics and AI Institute's low-inertia two-armed robot completes baseball-inspired throwing, catching and batting tasks, reported in Science Robotics.
An ultra-lightweight on-device STT model that hit #17 on Hacker News with 913 points, enabling transcription without cloud dependencies.
A database built for agents to store, retrieve and share state; #2 on Product Hunt with 235 upvotes.
A platform to create, play and share games built with Google AI; #3 on Product Hunt with 165 upvotes.
A real-time physics world model for agents to train and operate in physically simulated environments; #5 on Product Hunt.
Microsoft open-sources OS-level sandboxed code execution with policy-driven isolation for AI agents; 166 points on Hacker News.
Former Owkin executive Anna Huyghues Despointes unveiled cloud-controlled robotic wet labs in Turin that generate standardized, AI-ready biological data for training biomedical models.
Fourteen prizes awarded across digital innovation and security projects in Lazise, with AI, digital sovereignty and new ICT leadership roles at the center of the debate.
The AI data-center operator abruptly shelved its ~$5B IPO on Oct 9, pulling a valuation nearly triple its August figure amid scrutiny of AI price tags, and returns to private capital.
Khosla Ventures, Accel, Initialized, Kevin Weil and DST Global back the ~100,000-sensor cap that aims to let users control AI agents by thought.
Benchmark and Index lead a $60M seed for the five-month-old startup whose 'Engines' agents own full business outcomes, with Elad Gil, Definition and Diffusion participating.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
