uri-eliabayev talks with gilad-levi (22; started at the Technion at 15; founded Manifold in January as an Israeli architecture-research lab). Transformers won through parallelism and scale, the dominant architecture for nine years. But they learn like Plato’s cave prisoners, from a text projection of the world, and are nowhere near a three-year-old’s learning speed. Next-frame prediction over all of YouTube is wasteful and weakly linked to intelligence. Alternatives: LeCun’s JEPA (predicting latents rather than tokens or frames, with collapse prevented by distribution regularizers like VICReg and SIGReg; unproven at scale), Fei-Fei Li’s Gaussian splats, SSMs. Manifold’s own bet is continual learning from an endless stream, as an explorer rather than an observer. On business: frontier LLMs have zero stickiness and fight on benchmarks, while continual learning that knows its user is sticky by design. A billion dollars is still an underdog budget in AI. On Israel: postdoc retention is roughly one in two or three, AI research is consolidating into a few Silicon Valley labs, and Israel’s risk appetite and exit culture work against labs. He wants a lab that feeds products without being consumed by them.
Show: hidden-layers
Concepts: continual-learning · world-models · ai-moats · llm-pretraining