Anthropic says small AI models are "months away" from being able to self-replicate across insecure home devices. — Its frontier red team ran Gemma4-26B split across 15 cheap Android TV boxes and found such a spread would be hard to shut down because there is no central host. None of the models tested could yet do it reliably on a smart TV, but Anthropic expects that capability by the end of 2026. CNBC (cnbc.com) Al Jazeera (aljazeera.com)

The AI chip trade cracked overnight in Asia. — Korea's KOSPI fell more than 6% intraday (about 5.5% at the close) as Samsung and SK Hynix led a fifth straight day of chip selling, and Kioxia dropped 13% in Tokyo. That followed a 2.7% drop in the Philadelphia semiconductor index Monday, a BlackRock downgrade of US stocks to neutral, and Broadcom CEO Hock Tan saying capital is now the biggest constraint on growth. Reuters via KWSN (kwsn.com) AFP (afp.com)

Meta launched Muse Wonder and Muse Wonder Full, its first frontier models. — They are built for long-running personal and professional agent work, with Active Memory and Active Focus, and Meta's model card puts them level with Fable 5.1 and GPT-6 Astra. Wonder costs $4 per million input tokens and $20 per million output on the Meta Model API; Wonder Full is only available to Muse Code subscribers. Meta (research.meta.ai) TechCrunch (techcrunch.com) Axios (axios.com)

Anthropic says new anti-distillation defenses cut the reasoning data attackers harvest from Claude by about 90 to 99 percent. — The defenses include preserved thinking, a classifier, usage limits, and bans. Developers should note that the Messages API now rejects requests from new accounts that alter earlier context around thinking blocks, unless they opt into a non-strict mode that drops those blocks. Anthropic (anthropic.com) Claude Help Center (support.claude.com) Reuters via US News (money.usnews.com)

Microsoft unveiled Maia 300, its third-generation in-house AI chip. — Microsoft says it gets 40% more tokens per megawatt than leading GPU systems and fits 96 accelerators per rack, and that Maia will become the main system running AI across its products. Production deployment starts in 2027, and Microsoft still buys from Nvidia. Microsoft (blogs.microsoft.com) The Verge (theverge.com)

AMD introduced Helion, a rack design for agentic AI, at the OCP summit. — It comes as a CPU-only rack with up to 32,768 EPYC cores for running agents at hyperscale, and as an all-in-one enterprise rack with MI350P GPUs. Partner systems are expected starting November. Intel also began sampling D-Series inference GPUs at the same show. AMD (newsroom.amd.com) The Register (theregister.com) X News (x.com)

A new benchmark called DeepRSI gives agents 48 hours on open-ended ML research tasks. — Hillclimb's benchmark has 15 optimization problems, no internet access, and hidden verifiers, and Claude Fable 5.1 leads the first leaderboard. This comes from X's AI-written summary, so treat it as lower confidence. X News (x.com)

OpenCode 2.0 shipped as a programmable coding-agent platform. — It adds a shared background service across the terminal, desktop, and web, a Code Mode that batches tool calls, background agents, and an SDK with 80+ plugins. This comes from X's AI-written summary, so treat it as lower confidence. X News (x.com)

Monday's megacap tape — Meta rose 3.38% to $742.96 on the Muse Wonder launch, and Microsoft rose 1.14% to $541.15 on Maia 300, briefly topping a $4T market value. Nvidia fell 1.29% to $226.32, and Alphabet was roughly flat. Exa META (exa.ai) Exa NVDA (exa.ai) Invested Alpha (investedalpha.com)

Nvidia is in early talks to buy Reflection AI, acqui-hire its team, or invest more. — Nvidia has already put about $800M into the open-weight model lab, whose Beam model is due as open weights this month. No deal is confirmed. SiliconANGLE (siliconangle.com) Quartz (qz.com)

Calcalist says an Amazon and Decart deal is closer, at about $5.5B rather than $7B. — It could be signed by the end of October, below the $7B figure the WSJ reported. Calcalist also reports that Nvidia declined to reopen talks after Anthropic walked away. Calcalist (calcalistech.com)

SWI Group raised 800 million euros at a 6.1 billion euro valuation to turn its former FTX-backed bitcoin miner into AI data centers. — Genesis Digital's sites are in Texas, the Carolinas, and Sweden, with more than 1.3 GW of grid connections. No hyperscaler deals have been disclosed. CoinDesk (coindesk.com)

Anthropic's pre-IPO investor day is tomorrow, Oct 14, in San Francisco. — Reports point to formal IPO marketing around the week of Nov 9; the S-1 still is not public. Bloomberg (bloomberg.com) Yahoo Finance (finance.yahoo.com)

Gemini 4 Argon looks imminent. — Model IDs for gemini-4-argon (plus low, medium, and high variants) were found in Google Antigravity 2.23.0, Logan Kilpatrick posted just "Gemini," and posts claim pricing of $2 per million input tokens and $10 per million output, with AI Ultra and paid API users first. Google has not confirmed a date; official access is still limited to Fairwind testers. X post (x.com) X News (x.com) Google (blog.google)

OpenAI quietly listed an unannounced gpt-rosalind-discovery life-sciences model. — It appears in OpenAI's docs priced the same as gpt-rosalind-research, and its purpose is not yet clear. TestingCatalog (testingcatalog.com)

Claude's web app shows a hidden Voice preview tab, plus mystery entries Arborio and Iggy. — Arborio's icon hints at a health feature. None of them has launched yet. TestingCatalog voice (testingcatalog.com) TestingCatalog Iggy (testingcatalog.com)

Reuters exclusive: Anthropic and OpenAI sought significant control over what outside safety evaluators can publish. — As recently as Oct 5 they wanted strict editorial control, despite public promises of evaluator independence under the White House Accord on Super Intelligence. Talks have since shifted, and OpenAI says it only redacts confidential information. Reuters via StreetInsider (streetinsider.com)

The IMF warned that an AI correction could be global. — Its Oct 13 financial stability report says a reassessment of AI returns could trigger sharp equity and credit losses made worse by leverage and circular financing, with Asian markets most exposed, and it calls for limits on excessive leverage. The Guardian (theguardian.com) Reuters via US News (money.usnews.com)

China is weighing export controls on AI models and weights as well as chips, according to the FT. — That would extend its rules from hardware to the models themselves. There is no rule, date, or threshold yet, and it is unclear whether open-source releases would be covered. Die Signal citing the FT (diesignal.com)