METR spent $400k in API credits just to probe the HF hack, fueling AI swarm cost debate
lfschiavo · x · 2026-09-07
An X thread highlights the extreme cost of AI "swarms": @mentalgeorge notes METR's investigation of the Hugging Face hack transcripts alone cost $400k in API credits, with the actual swarm likely one-to-two orders of magnitude more expensive. He predicts a bifurcation of AI workloads between everyday coding tasks and million-dollar moonshot swarm programs affordable only to a few orgs — and argues future "warning shots" will likely come from the latter. The thread also boosts the phrase "ecologically valid evals" as a rising key concept in evaluation. Implication: cost barriers will stratify AI applications and shape where serious risks originate.
More from AGI Musings
- tenobrus: Human-centric code conventions may not survive superhuman coding models — ricklamers · 2026-09-07
- Deep dive: grading 42 funds behind the AI private-equity rollup wave — curious_vii · 2026-09-07
- OpenAI Publishes 'An Alien Mind' by Its Chief Scientist — bparrish · 2026-09-07
- OpenAI Chief Scientist: No Lab Has Solved Alignment Enough to Keep Scaling at Max Speed — Tinac4 · 2026-09-07
- LLMs as a Cognitive Virus: New Paper Models AI Dependence Tipping Points — serrjoa · 2026-09-07
- Privacy Outrage Is Just the Lag Between Tech Progress and Cultural Adaptation — signulll · 2026-09-07