Transformers are trained in the lab on static data but deployed in the real world where distributions shift; they cannot learn at test time beyond limited in-context learning or inefficient…

Current chain-of-thought scaling spends compute one token at a time, making inference costly and limiting frontier model access to a subset of users; architectural changes that increase com…

Current separate pre-training (perplexity minimization) and RL (chain-of-thought) pipelines are suboptimal; combining them with second-order optimizers like Shampoo and architecture-optimiz…

Human brains build custom circuits during development; matching biological efficiency likely needs analog compute with error correction, not just scaling current GPU/TPU architectures.

Building labs where AI agents write kernels, run experiments, and iterate architectures daily — targeting 10-200 experiments per day vs. current human-limited pace — will dramatically short…

Current coding agents already give individual researchers 10x leverage; a natively automated lab targeting 10-200 architecture experiments per day could outpace large labs constrained by or…

Novel architectures require custom high-performance kernels (e.g., 60x speedup for QR factorization), but current models cannot write them; automating kernel generation via AI-assisted sear…

Writing high-performance kernels (e.g., QR factorization) requires rare human expertise and months of effort; automating this with AI models would unlock orders-of-magnitude faster architec…