Summary
The morning feed is almost entirely the fallout from the US government forcing Anthropic to disable Fable 5 and Mythos 5, and the strongest signal is sentiment, not research. Robert Scoble, posting live from Anthropic's renamed "Claude Build Day" in San Francisco, captures the mood: builders showed up, loved Fable in the few days they had it, and left bummed they had to build on Opus 4.8 instead, while a recurring complaint is that export controls now handcuff US labs precisely where China is unconstrained. News aggregator @ns123abc carries the factual spine overnight, with the 90-minute shutdown ultimatum, the claim that an unnamed rival (reported as Amazon) triggered it via a jailbreak demo, and Anthropic's dispute of the White House timeline. There is no curated @bayesiansapien repost in this slot, so no human-filtered research signal. The only research-flavored thread is from Hugging Face's @eliebakouch, arguing that Mistral competes on being European and open-weight rather than on frontier capability, and that Chinese labs are training strong small models without the train-big-then-distill recipe. A scattered tail covers a voice-cloning demo (everlife.one), a note that MiniMax M3 and an open model from Rio de Janeiro's city IT department are filling the open-frontier gap, and @ellev3n11 flagging interest in training-codebase utilization for many small runs. Everything else in the feed is French political commentary and local crime reporting, off-topic for the wiki.
Posts
The Fable/Mythos shutdown, live from Claude Build Day (cluster of 6) (@Scobleizer, @ns123abc, @MillionInt). The dominant story all morning. Scoble reports from Anthropic's renamed Build Day that builders loved Fable in their brief access window and were disappointed to fall back to Opus 4.8; the recurring builder line is "we got the AI slowdown people wanted, until the Chinese do it, then American companies are handcuffed." @ns123abc carries the overnight facts: the government gave Anthropic 90 minutes to take both models down citing a national-security threat with no details, the trigger was reportedly Amazon CEO Andy Jassy flagging a Fable jailbreak rather than China, and Anthropic disputes the White House account (says it responded in 75 minutes, was never told the threat). Scoble also notes no AI investor has publicly defended Dario Amodei. @MillionInt frames the larger arc: the private-AI-versus-government relationship will define the coming years, "you may have all the men with guns but I have god in a datacenter." Full industry treatment is in today's digest.
Mistral and the open-frontier gap in Europe (@eliebakouch). Hugging Face's Elie Bakouch argues Mistral's position is not a failure: there is essentially no European competition on frontier LLMs, so Mistral only has to build "good enough," and enterprises pick it for being European and deployable on their own infrastructure, not for beating OpenAI or Anthropic on capability. He adds that a Mythos-class model would be too expensive for most EU companies to pay for or deploy anyway, and pushes back on the train-big-then-distill orthodoxy: while distilling from a larger model helps, Chinese labs are producing strong small models without it. A rare substantive research-strategy take in an otherwise sentiment-driven feed.
Open models filling the frontier gap: MiniMax M3 and Rio-3.5 (@Scobleizer). Scoble amplifies @ZenMagnets noting that with Alibaba's Qwen3.7 "fading at the frontier due to a proprietary stance," the open slots are being taken by MiniMax M3 and, oddly, Rio-3.5-Open-397B, a 397B model published on Hugging Face by the municipal IT company of Rio de Janeiro's city government. The MiniMax M3 thread connects to the MiniMax Sparse Attention work already in the wiki (MSA summary); M3 is the production model that ships MSA.
Training-codebase utilization for many small runs (@ellev3n11). Cursor's @ellev3n11 flags interest in a training codebase that gets high GPU utilization specifically for lots of small model runs rather than a few big ones: "big runs are cool, but lots of small runs is also cool." A practitioner signal toward the small-model-economics theme that @eliebakouch's Mistral take also touches.
Voice cloning in seconds: everlife.one (@Scobleizer). Scoble demos a service that cloned his voice in a few seconds so he can "talk to himself," calling the result uncannily like him and noting this was not possible a few months ago. A consumer-voice-tech data point; the linked everlife.one pitches a "digital you" reflection tool for journaling, explicitly not therapy.
Off-topic, skipped. The bulk of the slot is non-AI: @brivael's extended French political and economic commentary, @AustinJustice local crime reporting, @NICKIMINAJ, @WHFraudTF, @SeanParnellASW, and @heavypulp. No AI research substance.