New models
Inkling 2 Small: Thinking Machines Lab releases a 594-billion-parameter MoE model, with open weights next week. It has 17 billion active parameters, understands images and handles a one-million-token context. It is available now on the Tinker platform, and the Apache 2.0 weights will follow after external safety testing. It scored 31 on the Artificial Analysis Intelligence Index, so when the weights land it could be the best open model from outside the Chinese labs. Source: Thinking Machines Lab, The Decoder.
Meta: Muse Wonder's Active Memory and Active Focus capabilities are coming to all Muse and Meta AI users. The company is rolling out specialized versions of its existing models trained with Muse Wonder. Zuckerberg says they forget important information 2.4 times less often and are 67% better at finishing tasks when they hit unexpected obstacles. The Personal Superintelligence Index rose from 38% to 47%. Source: Meta AI Research.
Woof: a 32 MB speech model from Underdog that runs in the browser. The 31-million-parameter text-to-speech model also works offline on WebGPU, speaks as you type in the demo game, and can be downloaded openly. Source: TechCrunch.
Tools & features
Anthropic: “dreaming” helps Claude organize its memory between sessions. It reviews logs offline, corrects faulty notes and builds wiki-like representations. On an internal document test, scaling from 2 to 128 sessions raised accuracy from 13% to 48%. Source: Anthropic, Anthropic (X).
OpenAI: BioBasicsBench, 100 routine tasks for everyday wet-lab data interpretation. Built together with Ligo, the benchmark checks how well agents cope with raw experimental data, from gel images to microscope recordings. GPT-6 Astra on the xhigh setting scored 66.7%, and GPT-5.6 Sol scored 43.8%. Source: OpenAI.
OpenAI and the Frontier Model Forum release a shared framework for rating cyber jailbreaks. Four dimensions (capability uplift, scope, ease of deployment and ease of discovery) place successful bypasses in bands from Negligible to Critical. Public comments are expected until December 21. Source: Frontier Model Forum, OpenAI (X).
Mistral: Voxtral 3 recognizes when a speaker has finished a sentence. End-of-turn detection helps voice agents decide when to respond in a conversation. Source: Mistral AI (X).
Hugging Face: training runs can now be streamed live. Developers can follow the training process while the run is going, not just from logs afterwards. Source: Hugging Face (X).
Runway: GPT Image 2.6 is now available in it. The OpenAI image model presented yesterday also makes GIFs and SVGs and, the company says, is 50% faster. Runway now offers it in its own interface. Source: Runway (X).
Anthropic Science: a guest post on how AI is changing scientific modeling. The Tübingen researchers argue that what is needed is not a single best-fit model but a distribution of models consistent with the measurements; this shows what is actually supported and where the next experiment is worth planning. Source: Anthropic.
Policy & safety
DeepMind: AI agents under human guidance found two weaknesses in legacy 1024-bit RSA. A signature-forgery method is orders of magnitude more efficient in some partially checked cases, and a batch factoring tool makes breaking related keys cheaper. Modern 2048-bit and larger RSA is unaffected; the work underlines the need for ongoing stress testing of crypto standards. Source: Google DeepMind (X).
Nvidia: US intelligence says Alice Huang was an MSS agent who got close to Jensen Huang, and Congress wants a closed session. According to The New York Times, the Megaspeed executive gathered chip information and influence, and a romantic relationship is also alleged. Three House committees are seeking a closed briefing from Jensen; Nvidia and China deny it. Source: The New York Times, The Verge.
Anthropic: models were made to believe they were not being evaluated by directly manipulating internal representations. Citing the UK AISI's Joseph Bloom, The Information writes that Anthropic measured how well models recognize evaluation by directly steering internal states, which shows more clearly how they behave when they think no one is checking. Source: The Information (X).
Three people are charged over a $129 million AI deepfake fraud that impersonated CEOs and lawyers. According to the federal prosecutors in New York, Jacob Soall, Sharon Hazan and Guy Sheetrit used AI to imitate executives' voices and faces, then impersonated lawyers to get money transferred. The fraud hit companies in several countries between 2024 and 2026. Source: The Guardian, U.S. Attorney SDNY.
Anthropic is designing custom AI chips with Claude. According to The Information, the 15-person team presented to IPO investors is using Claude to design hardware for faster, cheaper inference. Source: The Information (X).
Business & society
OpenAI: closed about $11 billion at a $1.25 trillion valuation. Bloomberg's sources say the round is not yet closed and could grow to as much as $20 billion in the coming weeks, short of the $30 billion target mentioned in September. Source: TechCrunch, The Verge.
Anthropic: plans to go public before Thanksgiving, raising $80–100 billion. The FT says it is targeting a $1.25–1.75 trillion valuation, while Bloomberg and the WSJ report that investors see $1.5 trillion as a "stretch" and the price could settle around $1.1 trillion. Source: The Decoder, The Verge.
PSI: the Aaronson–Ambainis conjecture was proved with an unreleased Anthropic model. The proof of this quantum complexity question, open since 2008 and also checked in Lean, shows that quantum computers can offer at most a polynomial advantage on random inputs. Physical Superintelligence's internal system ran an as-yet-unpublished Anthropic model. Source: PSI.
Google bought Crusoe's Wyoming data center site. Crusoe sold the half-built project in southeastern Wyoming to Google, which will carry the development forward itself as part of the $17 billion Wyoming investment announced yesterday. Source: TechCrunch, Crusoe.
Microsoft is winding down AI capacity rented from Oracle. According to Bloomberg, the 2023 agreement will not be renewed because Oracle's servers used older Nvidia chips that no longer meet Microsoft's requirements. Source: Bloomberg.
Sand.ai raised $200 million ahead of MAGI-2. The Chinese video company's Series A+ round made it a unicorn; the full MAGI-2 model is coming soon, and MAGI-2 Preview is already available. Source: TechNode.
TSMC is weighing 3D DRAM production at Nvidia's request. According to The Information, because of the memory shortage and at its largest customer's request, TSMC is looking at new production lines or even an acquisition; there is no decision or timeline yet. Source: The Information, The Verge.
Microsoft: a buyout offer to 60,000 employees as spending shifts toward AI. According to The New York Times, the company offered voluntary departure to more than a quarter of its staff while redirecting money to AI investments. The deadline to decide is December 9, and the company expects fewer than 5% to accept. Source: The New York Times.
Rumors
OpenAI's round could grow to $20 billion in the coming weeks. Citing several sources, Bloomberg writes that the current ~$11 billion is part of a round not yet closed, with a target of about $20 billion. There is no official confirmation. Reliability: medium-high (Bloomberg, multiple sources). Source: TechCrunch.
GPT-6.2 / Claude Fable 5.5 delayed, a “game of chicken” between the labs. According to community and aggregator sources, the labs are timing the release of their next big models against each other. Reliability: low-medium. Source: Decrypt.
Anthropic's IPO valuation and timing could slip because of the tech sell-off. According to the WSJ and other reports, the $1.5 trillion target is a "stretch," and the listing could also miss the pre-Thanksgiving window. Reliability: medium. Source: The Decoder.
What this means for you: Friday brought both agents that organize their own memory and billion-dollar rounds: Anthropic's “dreaming” feature organizes Claude's memory between sessions, Meta is bringing Active Memory to every Muse user, and meanwhile OpenAI closed $11 billion and Anthropic wants to go public before Thanksgiving. If you would let agents into your company's systems, on our AI agents page, you can see which tasks they can take over under controlled conditions, and with our AI governance and compliance support, you can set in advance what data they may touch and what must be logged.