ChatGPT's new model now rivals Claude for math animations, demo shows
PTenigma · x · 2026-09-06
- The author tested the new ChatGPT model Astra for generating statistical teaching animations and found it now matches Claude, which used to be clearly better at math animations.
- Demo: 2-class LDA in p dimensions where data is "heated" with individual and site-specific noise, classifier AUC degrades; a learned spectral filter suppresses the site effect and AUC recovers.
- Framed via heat semigroups: corruption and the spectral filter connect to diffusion and learning to reverse it (site effect ≈ diffusion tensor); in SDE terms the discriminant dies off under heating, a "Gamma flow."
- Gives the identity (d/dt)D²(t)=−2ΓQ(ℓt), where Gamma is the "carré du champ" of the LDA score; the author will discuss using Gamma calculus to improve encoders in a MICCAI talk and leaves fMRI motion-removal as an exercise.
More from Models
- GPT-6 Astra made free on Experiential platform alongside free tiers of Qwen3.8, DeepSeek V4 Flash and more — Roger_M_Taylor · 2026-09-06
- GPT-6 Astra and Claude Fable 5.1 appear free on third-party platform Experiential Labs (unverified) — Roger_M_Taylor · 2026-09-06
- NVIDIA gives free year-long access to 140+ AI models including GLM 5.2 and MiniMax M3 — Roger_M_Taylor · 2026-09-06
- Andrew Carr: now is the day for huge, well-documented proprietary datasets — andrew_n_carr · 2026-09-06
- With $20 to spend: Gemini Pro vs GPT Plus for Go and Python engineering — BrejeiroKiller · 2026-09-06
- GPT-6 Astra autonomously beats Portal, echoing OpenAI's 2016 game-solving goal — scaling01 · 2026-09-06