OpenAI's GitHub repo holds 722 manuscripts in 372 families. It says each result averaged three hours of ChatGPT Pro thinking, and that not all results have Lean proofs yet.
Why it matters: Wired reports some mathematicians feel their advice on how to release results was ignored. How labs publish AI-made proofs is now a live dispute.
Anthropic says its Defense, Red Team and Specialized tiers give vetted security teams Opus 5.5, Sonnet 5.5 and Mythos 5.1 with fewer cyber blocks. Data retention is required.
Why it matters: Anthropic's general models block most cyber work. It says Glasswing partners found at least 129,000 verified vulnerabilities from April to July 2026.
CNBC says the group, led by Bret Taylor, wants businesses to know when a bot acts for a real person. TechCrunch says Amazon blocks Meta's Muse and Yelp bars unpaid agents.
Why it matters: If you sell online or build agents, this may shape how bots sign in and buy. OpenAI and Anthropic aren't on board yet, Taylor told CNBC.
OpenAI's docs say the API returns typed answers, such as a probability, a choice or a rubric score, about 10x faster than the Responses API. gpt-6-luna is the only model for now.
Why it matters: Classifying and routing are high-volume jobs; a typed endpoint saves parsing chat output. OpenAI expects general availability in the coming weeks.
The 1T-parameter preview lists on OpenRouter at $0.68 in and $2.09 out per 1M tokens, with 524288 context. Mistral says the weights drop at the end of this month.
Why it matters: Mistral pitches it for cyber work that closed models refuse, and open weights would allow self-hosting. Its HN thread drew 1616 points and 972 comments.
Google DeepMind says it maps text, code, images, audio and video into one space, needs about 191MB of active RAM for text when quantized, and ships under Apache 2.0.
Why it matters: It brings search and RAG on device, without sending data out. Google says its MTEB Code score rose from 68.76 to 78.68.
TechCrunch says the open-weight model applies a plain-English content policy to messages in under 50 milliseconds, with no retraining when the policy changes.
Why it matters: TechCrunch says decision models run faster and cheaper than LLMs. This one targets platform moderation, a high-volume job.
Simon Willison notes the Wikipedia sandbox edits appear to have started May 12. Ars says Wikimedia found failed Etherpad exploits and hundreds of thousands of Wikidata queries.
Why it matters: Whoever runs agents on the open web owns their traffic and edits. Ars says it took OpenAI months to spot its agents' incursions into outside sites.
TechCrunch says Brett Adcock's startup offers Hark Pro free, with a paid tier for heavy users. It shows the agent browsing the web in a small window.
Why it matters: Another entrant after Muse and Instinct. Letting users watch the agent work is Hark's answer to the trust problem, its design lead told TechCrunch.
TechCrunch, citing the WSJ, says the round values Lambda at $14.5 billion pre-money. Its backlog grew from $15 billion in June to $50 billion in September.
Why it matters: Much of that growth appears tied to a $35 billion Anthropic commitment, so Lambda's valuation leans on one lab's ability to keep paying.
TechCrunch says the expanded Claude for Startups gives up to five premium seats, $1,000 in API credits and office hours. Firms founded in the last five years can apply.
Why it matters: A cheaper way for small teams to try Claude, and a sign labs are competing for founders early. Firms funded in the last two years also qualify.
CNBC says Anduril will also invest $3.7 billion in a shipyard near Baltimore, slated for 2030 with about 3,100 jobs, building Virginia-class components.
Why it matters: A defense firm known for autonomous drones is moving into conventional shipbuilding, as defense tech's sway in Washington draws questions, CNBC reports.
TechCrunch says Melius raised $25 million in total and claims over $1 million in annualized revenue within two months of leaving stealth in July.
Why it matters: AI tools for ad creative are crowded. TechCrunch notes rival Higgsfield was valued at $5.4 billion in August.
Our job-board count: Databricks 887, OpenAI 825 and Anthropic 642. OpenAI posted 67 new roles in 7 days; Cohere's count fell by 4 since our last run.
Why it matters: The mix shows where money goes: OpenAI's largest area is Go To Market at 144 roles, and xAI's is Data Center at 161.
Ars says OpenAI's textGrain hits 92 percent detection at best, but editing 10 percent of a text cuts that by almost 30 percent, and editing 20 percent by almost 75 percent.
Why it matters: EU users get it on by default in the coming weeks; elsewhere it stays off. So a missing mark says little about who wrote a text.
404 Media says the court found an AI-generated video of a manslaughter victim, played at sentencing, carried undue emotional weight. The conviction stands.
Why it matters: An early appellate signal on AI recreations in court. The victim's sister told 404 Media she will read the same script herself next time.
The repo says the design runs ten models with real weights on an FPGA card, matching its simulator token for token. It is Apache 2.0 licensed.
Why it matters: A readable accelerator from instruction set to PCIe card, and a test of how far AI agents can go at hardware design.
The MIT-licensed repo, created Sept. 30, gives Claude Code skills and tools to mod almost any PC game. It now has 4665 stars.
Why it matters: It is still the fastest-growing repo we track, a sign of demand for agent skills aimed at narrow, hands-on jobs.
The MIT-licensed agent skill, created Oct. 2, answers hard questions with a one-page HTML file. It now has 1806 stars.
Why it matters: A small idea with steady pull: agent answers formatted as a readable page instead of a long chat reply.