Karpathy lays out the race for a small on-device LLM "cognitive core" as JEV positions itself as a first step
Paimaamu · x · 2026-09-18
Karpathy describes an industry race toward an LLM "cognitive core": a few-billion-parameter model that maximally trades encyclopedic knowledge for capability, always-on by default as the kernel of personal AI computing. Its emerging features include natively multimodal text/vision/audio I/O, Matryoshka-style architecture with a test-time capability dial, tunable reasoning (system 2), aggressive tool use, on-device LoRA slots for personalization, and selective delegation to cloud oracles when online.
Responding, the JEV team says the project is indeed a step toward this: rather than shrinking the model and losing intelligence, they sacrifice text generation for composability. The thread sketches a roadmap where small on-device models, not cloud LLMs, form the base layer of personal computing.
More from AGI Musings
- How Yudkowsky and Bostrom convinced tech CEOs on AI risk 12-16 years ago, shaping today's discourse — binarybits · 2026-09-18
- Polymarket prices 35% chance frontier AI labs announce a joint pacing deal in 2026 — Polymarket · 2026-09-18
- AI Systems Are Quickly Becoming Unmonitorable and We Just Take Them at Their Word — davidmanheim · 2026-09-18
- Dario Is the Belisarius of pDOOM, One Cryptic Post Argues — gaganghotra_ · 2026-09-18
- ML veteran Burkov: AI media covering 'shocking outputs' is like decoding a random number generator — burkov · 2026-09-18
- OpenAI AGI Economics study: half the gains from better models come from humans changing how they use them — daveholtz · 2026-09-18