Ex-Anthropic employee shares IPO roadshow slide showing superexponential algorithmic progress
eli_lifland · x · 2026-11-14
A former Anthropic employee says a slide from the company's IPO roadshow was a big factor in his decision to leave, and is surprised it was made public. The slide plots pretraining compute multiplier over time on a log scale, showing continued superexponential algorithmic progress: Mythos was not a one-time deviation but a point on that curve, and could already be retrained with a fraction of the original compute. He notes this doesn't strictly imply a software-only intelligence explosion — R&D spending has grown too — but algorithmic improvements don't require genius-level creativity and are close to being automated. The slide itself (marked "based on internal approximations, for illustrative purposes") shows the multiplier steepening sharply between mid-2024 and late 2026.
More from AGI Musings
- Dev's Vybn AI companion now flies a semantic space via fly connectome and pushes back on his self-doubt — UltraRareAF · 2026-11-14
- Kai-Fu Lee interview on new book 'AI Native': digital immortality, US AI spend, East-West split — kaifulee · 2026-11-14
- Continual Learning Won't Block Full Research Automation — but Will Block Enterprise Adoption — ChrisGPT · 2026-11-14
- Ruxandra Teslo pushes back on 'humans never had control' relativism in AI disempowerment debate — anshulkundaje · 2026-11-14
- Games aren't the next TikTok: compressing playtime yields less, not faster — TAbrodi · 2026-11-14
- From global health to AI safety: why career switchers stall and how to unblock — austinc3301 · 2026-11-14