Karpathy's notes after the Dwarkesh interview
AGI as a decade of agents, animals vs ghosts, RL as a straw, cognitive core, and why he does not want 1,000 lines of unsupervised code.
The interview went everywhere. This is Karpathy cleaning up the parts he botched by talking too fast: timelines, Sutton, RL, the cognitive core, nanochat, and a working relationship with models that is not 'go away for 20 minutes.' Bookmark-heavy on purpose.
Related reads
ThreadMuseAI
How Meta built safety into Muse
@AIatMeta
Official deep dive. Secure VM, Sentinel on the same box, prompt-injection quarantine, public bug bounty. First-party, long, and the reason they claim the agent can hold your context.
ThreadMuseAI
Introducing Muse
@MetaNewsroom
Meta's product note. Personal agent, Muse Spark 1.3, Secure VM, Stripe Link for purchases, free for most of what people need.
ThreadAgentsEngineering
The Agents CLI in 30 minutes
@svpino
Santiago on the Agents CLI. Build, deploy, and watch a small multi-agent system without the usual harness ceremony.
ThreadAIAgents
AI-native thinking
@dotey
Baoyu's Tencent talk. Find demand on the model capability line, kill bad ideas with prototypes, then redesign so the agent verifies itself.
