Terence Tao's dire AI warnings for science, plus a resignation from Anthropic
Gary Marcus · rss · 2026-09-09
Gary Marcus compiles two warnings about AI's trajectory.
Terence Tao's shift: once optimistic about human-machine coexistence, the famed mathematician posted three messages on mathstodon signaling deep concern — that puzzle-solving isn't insight, that AI threatens intellectual taste in math, and calling for transparency. The last alludes to the OpenAI-NYU-Anthropic Navier-Stokes controversy, where OpenAI spent $22.5M scooping Alpöge and Buckmaster and couldn't rule out that their product-usage data improved its models. Marcus calls it a dire warning: AI can't 'cure cancer' if scientists can't trust AI companies with their IP.
Jacob Coxon's exit: the just-departed Anthropic researcher's viral account of 'self-indulgent and dangerous' thinking at frontier labs is spreading fast. Marcus finds it exaggerated on timing but plausible and unsettling, concluding extinction is unlikely but catastrophic harm is possible — and nobody has a serious plan; the real fix may require abandoning LLMs entirely.
More from AGI Musings
- Quantum AI is the most misused buzzword: it won't train GPT-6, but it can crack RSA — AryHHAry · 2026-09-09
- AI lab safety concerns aren't just a marketing stunt, argues KOL AndyMasley — AndyMasley · 2026-09-09
- 80,000 Hours lays out why AI risks are the world's most pressing problems — AndyMasley · 2026-09-09
- Boaz Barak: OpenAI is changing fast as staff viscerally feel the magnitude of capabilities — dgrobinson · 2026-09-09
- Keller Jordan: pausing algorithmic progress while compute piles up maximizes AI risk — kellerjordan0 · 2026-09-09
- Researcher publicly challenges frontier lab: how to reconcile race dynamics with safety talk — joshalbrecht · 2026-09-09