OpenAI slowdown and safety skepticism

Why companies are becoming a series of loops | Anish Acharya (a16z)

Topic 10 of 19

Published Sep 6, 2026 · Clip 42:49–44:49 (2:00)

On OpenAI pausing the RL phase of its latest model, Anish is skeptical of the 'too dangerous to release' narrative, pointing to confounding factors: the extraordinary aura Anthropic gained from withholding a model, possible GPU shortages, and the incentive to keep a proprietary model internal to extend a lead. He takes offensive cyber capability seriously — systems should be hardened before they become trivial to penetrate — but thinks the dangerous-model story conflates marketing, inference capacity, and economic considerations. Industry trends contradict the feared proprietary-supermodel outcome: Anthropic looked unassailable, then OpenAI had an amazing six months and open weights are ripping, with Grok bot arriving out of left field.