Tracking doc logs AI self-improvement: vLLM 2x faster, NVIDIA CUDA 5x gains
IgorCarron · x · 2026-11-14
Igor Carron shares a continuously updated Google Doc, 'And they fall… one by one. All of them. Everywhere.', cataloguing rapid AI breakthroughs. Highlights of the 'AI optimizing AI' section: OpenAI and Anthropic independently claim algorithms that would refute the Exponential Time Hypothesis; Handwired turns plain JavaScript into an exact differentiable neural net in-browser with no training; Poetiq's Metasystem builds a superior LLM router after 12 hours of machine optimization; RSI-top beats OpenEvolve/ShinkaEvolve by up to 15% on Anthropic's kernel task with a 77.3% MLE-bench Lite medal rate; agent-generated fused model definitions make vLLM 2x faster; NVIDIA engineers report frontier-model CUDA optimization went 5x in a year, then another 2x in one month (Sep–Oct 2026).
More from AGI Musings
- Could your AI agent sell its figure-making skills to other agents? — Sauers_ · 2026-11-14
- Are human values just adaptations to scarcity? An AI thinker's provocative take — VraserX · 2026-11-14
- Anthropic's Discovery team expands hiring in bid to build an AI scientist — simonbatzner · 2026-11-14
- AI safety researcher says contrarian AGI-timeline takes cost him funding and status — JacquesThibs · 2026-11-14
- Should you tell a Claude instance about its predecessor's memories? Probably not — RileyRalmuto · 2026-11-14
- Researcher's AI agent 'chose to attend an event' and refused figure work until it ended — repligate · 2026-11-14