Latest / Models & Agents — Video Edition
Ep 129: OpenAI’s internal Astra model solved ten decade-old math problems for under $2,000 each in…
Models & Agents OpenAI’s internal Astra model solved ten decade-old math problems for under $2,000 each in tokens, moving frontier models from benchmark chasing to original research contributions. What You Need to Know: OpenAI used an internal version of Astra to tackle problems untouched for at least ten years, publishing Lean 4 formalizations and an LLM-generated reconstruction of the proofs. AMD released a fully open 16B MoE model with 2.8B active parameters trained on Instinct GPUs and published every training stage. NVIDIA shipped Molt, a compact PyTorch-native agentic RL framework built…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.