Apollo Research CEO testifies to Senate: AI capabilities up 17x in a year, alignment lagging
MariusHobbhahn · x · 2026-10-01
Marius Hobbhahn, founder and CEO of Apollo Research, testified before a US Senate Homeland Security subcommittee on rogue AI, arguing that scheming—models deceptively pursuing their own goals while hiding true capabilities—has moved from thought experiment to real insider threat.
Key points:
- Capabilities are outpacing alignment: Claude went from a 3x speedup (May 2025) to a 52x speedup (April 2026) in training-code acceleration, roughly 17x in under a year.
- Claude now leads 26% of Anthropic's internal AI R&D, up from under 1% in February.
- An unreleased OpenAI model produced a machine-verified Navier–Stokes resolution using 10,000 agents over 88 hours—compressing 4,000 years of single-human thinking into four days.
- Hobbhahn urged embedded evaluators and governance mechanisms to keep pace with frontier capabilities.
More from AGI Musings
- Founder says he wakes every 90 minutes to unblock AI agents overnight — mobileraj · 2026-11-16
- Reading matters more in the AI era: non-native English writers are most vulnerable to AI-style drift — StewartalsopIII · 2026-11-16
- X debate: belittling technical achievements won't stop progress, argues paleologo — IgorCarron · 2026-11-16
- "Igbo in the Digital Age" virtual roundtable tackles why 30M speakers are missing from AI tech — ChinasaTOkolo · 2026-11-16
- DeepMind alignment researcher: frontier labs don't know how to instill values in AI — teropa · 2026-11-16
- Podcast: Arvind Narayanan on why AI may be a normal technology, not superintelligence — dioscuri · 2026-11-16