This edition of Import AI covers four main topics: (1) An analysis of fast16.sys, a pre-Stuxnet virus that subtly corrupted precision engineering calculations in tools like LS-DYNA, potentially targeting weapons programs. (2) A teardown of the Muon optimizer revealing it causes neuron death in MLP layers, along with the introduction of Aurora, a leverage-aware replacement that outperforms Muon on 1.1B-parameter models. (3) A position paper from researchers at Oxford, Google DeepMind, OpenAI, and Anthropic advocating for 'positive alignment' — designing AI to actively support human flourishing beyond just preventing harm. (4) Prime Intellect's experiment showing LLM agents (Codex and Claude Code) can autonomously optimize AI training runs, beating human baselines on the nanoGPT speedrun, though they struggle with original ideation.
Sort: