Anthropic IPO slide shows superexponential training-efficiency gains, ex-researcher reveals
141_1337 · reddit · 2026-11-14
Jacob Coxon, a former OpenAI and Anthropic researcher, shared a slide from Anthropic's IPO roadshow that he says was a big factor in his decision to leave — and that he's surprised the company made public.
What it shows: Pre-training compute multiplier over time on a log scale, indicating continued superexponential algorithmic progress from Jul 2024 to Oct 2026.
His takeaways
- Mythos was not a one-time deviation but a point on a superexponential curve; Anthropic could already retrain it with a fraction of the original compute
- This doesn't strictly imply a software-only intelligence explosion — R+D spending has grown too — but algorithmic improvements don't require creative genius and are on the verge of being automated. "You do the math."
Note: the slide's footnote says data points are internal approximations for illustrative purposes.
More from AGI Musings
- SfN2026 minisymposium on embodied computation spans zebrafish whole-brain dynamics to NeuroAgents — tweetsatpreet · 2026-11-15
- Patient data that never leaves the hospital is the easiest AI project to get approved — MaziyarPanahi · 2026-11-15
- Blogger argues anti-AI sentiment is a psyop by old creative-industry gatekeepers — nptacek · 2026-11-15
- What happens when AI trains on AI content for 100 generations? — Huge-Conference-1193 · 2026-11-15
- LMU Dissertation Recasts Neural Networks as Cybernetic Error-Controlled Regulators on the Path to AGI — sebkrier · 2026-11-15
- The Strangest AGI Outcome: Humans Stop Being the Audience for New Knowledge — VraserX · 2026-11-15