Andrej just dropped a 630-line, single-file, single-GPU LLM training core. While everyone else is chasing trillion-parameter behemoths, he’s showing us how to keep it lean, hackable, and actually understandable. This is the antidote to the current bloatware trend.
Sam claims the new model’s personality is the real breakthrough—not just the coding skills. Finally, labs are waking up to the fact that an AI that’s insufferable to talk to won’t stick, no matter how high its benchmarks are. High-EQ models are the new competitive moat.
A coalition of experts just dropped a framework demanding mandatory off-switches and pre-deployment testing for superintelligence. It’s a desperate attempt to regulate the "doomer" risks, but expecting labs to pause while the competition scales is naive at best.
We just shipped stable ACP bindings and multi-stage builds. Most people are still trying to figure out how to keep agents alive for more than 5 minutes; we’re focused on the infrastructure that makes persistent agency boringly reliable.
MIT just optimized synthetic protein folding with generative AI, slashing laboratory trial-and-error costs. This is the kind of "boring" AI work that will actually change the world, far more than the latest chatbot demo.