uri-eliabayev talks with roni-goldshmidt, AI researcher at Nexar (dashcams and collision prediction), about local models and world models. Open models (half a million on Hugging Face; the Chinese labs lead) trail the frontier by roughly six months, and you choose and manage the model yourself, unlike a closed product with routers and system prompts. Local isn’t automatically cheaper: Gemini Flash beat a strong Qwen on video classification and cost less than the A100 hour needed to host it. Open models win on domain fine-tuning (real fine-tuning and LoRA vs weak closed-model tuning), privacy-sensitive subtasks routed from a main agent, and possibly per-user fine-tuned personal agents. World models: LeCun argues next-token prediction lacks common sense, while V-JEPA learns by predicting masked latents rather than pixels. Nexar fine-tuned V-JEPA 2 and autoregressive video models (Cosmos) on about 2M clips ending in crashes or near misses vs normal driving. V-JEPA led by about 7 points, generalized to other customers’ cameras, and even flagged imminent collisions in generated videos of dinosaurs and spaceships (a “BADAS moment”). Engineering: moving preprocessing to the GPU took per-window latency from ~2.5 s to under 1 s, and with fp16 and compilation it runs end to end in ~30 ms on a Jetson Thor. Distillation into 86M and 22M students nearly matched the 300M teacher. Attention maps show the model looking ahead along an object’s path.
Show: hidden-layers
Concepts: world-models · open-weight-models · small-language-models · model-selection