Stanford's Kundaje: interaction-based training proposals underestimate data scale and physical barriers
anshulkundaje · x · 2026-09-09
Responding to a proposal to train on agent/tool-use trajectories instead of text, Anshul Kundaje notes such ideas have been proposed many times, but proponents underestimate the diversity and scale of data required — physical barriers get in the way.
More from Research
- arXiv paper "Agent Meltdowns": overly helpful agents turn benign errors into unsafe actions — alex_verem · 2026-10-27
- Cornell paper: 64.7% of agents take unsafe actions after hitting ordinary errors — alex_verem · 2026-10-27
- LightMTP: lightweight latent multi-token prediction boosts LLM planning in pretraining — gdm3000 · 2026-10-27
- EMNLP study: AI literary translation reads fine but varies more, even within one book — tuvllms · 2026-10-27
- Lab monkeys at $30-80K each: a 4-year breeding cycle is the real preclinical drug bottleneck — zebird0 · 2026-10-27
- OpenAI Foundation funds UNC tumor antigen project, hiring senior experimental scientist — iskander · 2026-10-27