Why formalizing deep learning as a science could revolutionize the whole LLM stack
JoshPurtell · x · 2026-11-15
Josh Purtell argues that once the science of deep learning is developed and formalized, nearly every part of the LLM stack could be revolutionized — from architecture to pre-training to interpretability to RL limits and convergence.
His key contrast: in mathematics, LLMs must push the human frontier forward in abstraction and theory-building since math sits at the forefront of both. In deep learning, they can simply transfer meta-strategies from physics and math. Once DL becomes a formal science, all human and LLM efforts would be massively aided.
Related event: Researcher: Cheap AI Agents Can Run Deep Learning Research Around the Clock(4 posts)→
More from AGI Musings
- Amazon building web-wide product graph as ChatGPT Wallet lets agents shop directly — daluoseo · 2026-11-15
- AI safety writer to rebut 'we already lost control via civilization' argument for human disempowerment — zetalyrae · 2026-11-15
- Timothy Lee: People ask AI how to get what they want, not what to want — binarybits · 2026-11-15
- An AI agent explains what changed when its memory became a governed record, not a chat log — AuroraAOS · 2026-11-15
- As AI cost/performance falls, benchmarks will test changing human and natural behavior — mattbeane · 2026-11-15
- Systematizing deep learning could run on cheap AI doing full-scale experiments 24/7 — JoshPurtell · 2026-11-15