Anthropic suspends foreign access to Fable 5, Mythos 5
Today's top 14 insights for PM Builders from X and YouTube.
Anthropic suspends foreign access to Fable 5, Mythos 5
#1 đť•Ź
Anthropic has been ordered by the US government to suspend all foreign-national access to Fable 5 and Mythos 5, so both models are now disabled for all customers. Access to other Claude models remains unaffected, and they’re working to restore service as soon as possible.
Also covered by: @Armin Ronacher
#2 đť•Ź
Sebastian Raschka introduces Cohere’s new 30B open-weight model, a parallel-transformer successor to Command A+ that doubles layers while halving size.
#3 đť•Ź
There's An AI For That Netflix’s new AI model, VOID, not only removes objects from video footage but also automatically corrects the scene’s physics afterward.
#4 ▶️
Claude Fable 5 is BANNED. What to do?
Greg Isenberg
The video details running a 12-billion-parameter AI model locally on a 16 GB RAM machine using LM Studio or Ollama, applying Q4 quantization, and connecting it to the Hermes agent for an always-on, offline private AI layer.
- LM Studio offers a GUI with a model browser and one-click runs, while Ollama provides a single-command CLI, both enabling local model deployment in approximately 10–20 minutes without internet or API keys.
- A 12-billion-parameter model fits on a 16 GB RAM system; 27–35 billion-parameter models require ≥30 GB RAM or a dedicated GPU, and 70 billion-parameter models need ≥128 GB unified memory (e.g., Nvidia DGX Spark).
- Quantizing models to Q4 roughly halves their memory footprint with minimal quality loss, allowing server-scale models to run smoothly on laptops.
#5 đť•Ź
Ali Ghodsi launched Omnigent, an open-source harness that plugs into Claude Code, Codex, OpenCode and pi, lets teams collaborate via Slack/Teams, CLI or WebUI, and enforces a fine-grained security model to tightly control what agents can do.
#6 đť•Ź
Teresa Torres highlights how Musubi built a moderation pipeline that flags AI-human disagreements and sends the content, both decisions, and the customer’s policy to a reasoning model as a tiebreaker.
#7 đť•Ź
clem 🤗 argues that current guardrails for frontier model APIs are shallow, easily jailbroken smokescreens, and calls for a new paradigm in AI safety.
#8 đť•Ź
Logan Kilpatrick says Google AI Studio’s long-term goal is to remove friction from building AI, then bring that ease to businesses to unlock economic opportunity for everyone. He sees it as early innings but is increasingly excited about the progress.
#9 đť•Ź
Lenny Rachitsky hosts a deep-dive with Benedict Evans on AI’s future, covering where value will actually accrue in the AI stack, why labs are buying consulting firms, the rise of anti-AI sentiment, distribution as the ultimate moat, and reframing the AI-job question from “wha...
#10 đť•Ź
Madhu Guru notes that launching a frontier LLM is like shipping a black box with infinite use cases and failure modes, demanding tough trade-offs via extensive evals and red-team testing.
#11 đť•Ź
There's An AI For That open-sourced Netflix’s VOID-model on GitHub, providing the full codebase, configs and pretrained checkpoints for their new retrieval-augmented LLM.
#12 đť•Ź
Sebastian Raschka notes that large sparse models require only about 40–60 GB of memory and run fast enough for practical use, whereas equivalently sized dense models remain too slow.
#13 đť•Ź
Logan Kilpatrick unveiled a free platform that takes you 80% of the way from idea to a working, shareable app.
#14 đť•Ź
Cognition is removing access to Claude Fable 5 across its products following Anthropic’s announcement and a US government directive. All other models—Claude Opus 4.8 and GPT-5.5—remain in Devin, and Devin Ultra will keep using the most powerful available models.
Also covered by: @Armin Ronacher