Karpathy at Sequoia Ascent 2026: Three New Frontiers LLMs Open Beyond Speed
Original: Karpathy at Sequoia Ascent 2026: Three New Frontiers LLMs Open Beyond Speed View original →
LLMs as More Than Accelerants
Andrej Karpathy shared highlights from a fireside chat at Sequoia Ascent 2026. His central argument: LLMs are not just tools for doing what we already do faster, they unlock categories of functionality that either should not need to exist anymore or were previously impossible.
Three New Horizons
1. LLM-native apps (e.g., menugen)
Apps where the LLM handles all computation natively, input an image and output an image, with no classical code required. The app becomes just a prompt, not software.
2. .md skills instead of .sh scripts
Why write a complex bash install script when you can write the installation out in natural language and hand it to an LLM? The LLM reads English as a high-level interpreter, targets your specific setup, and debugs inline.
3. LLM knowledge bases
Computation over unstructured data from arbitrary sources and formats was fundamentally impossible with classical code. LLMs make this a first-class capability.
Explaining LLM Jaggedness
Karpathy addressed why the same model can coherently refactor a 100,000-line codebase and also give nonsensical answers. The answer lies in verifiability and economics: RL training distributions follow revenue and TAM, so models excel on tasks well-represented in training data and struggle off-distribution. Understanding this is key to practically harnessing LLMs.
The Agent-Native Economy
Karpathy's third theme: products and services decomposing into sensors, actuators, and logic across all computing paradigms. The emerging core skill in agentic engineering is making information maximally legible to LLMs. His longer-term vision: mostly-neural computing with classical CPUs as coprocessors.
Related Articles
Google unveiled Gemini 3.5 Flash at I/O 2026, combining frontier performance with agentic capabilities. The new Managed Agents API provisions a full sandboxed agent environment with a single API call.
The useful number in the Reddit report was not the hardware spec; it was a reported 12% tool-call formatting error rate.
NVIDIA is packaging a 550B-parameter MoE model with agent tooling instead of treating the model as a standalone release. The pitch is concrete: up to 5x faster inference, up to 30% lower cost, and availability beginning June 4.