# karpathy

Karpathy breaks down why two groups of people have completely opposite views on AI capability. One group is laughing at ChatGPT fail videos. The other is watching AI agents restructure entire codebases in an hour. Same technology, different universes.

Karpathy found that vibe coding makes writing code a breeze, but deployment is pure hell. His exchange with Stripe CEO Patrick Collison reveals the next battleground: the entire DevOps lifecycle must become code before AI agents can truly take over.

Karpathy's MenuGen journey shows the hard part of vibe coding was not writing code, but assembling Vercel, Clerk, Stripe, OpenAI, and other services into a product. His takeaway: DevOps must become code before agents can really ship.

Andrej Karpathy spent four hours polishing an argument with an LLM, felt invincible, then asked the same LLM to argue the opposite — and got completely dismantled. LLM sycophancy is a real trap, but flipping it around is genuine alpha.

LiteLLM hit by supply chain attack — pip install was enough to steal all credentials. Karpathy warns about dependency tree risks and advocates using LLMs to yoink functionality instead of adding more deps.

Karpathy's Autoresearch isn't trying to be a general AI scientist. It's a ruthlessly simple experiment harness: the agent edits one file, runs for five minutes, checks one metric, keeps wins, discards losses. The lesson? The best autonomous systems aren't the freest — they're the most constrained.

Andrej Karpathy argues the IDE era isn't over — it's evolving. The basic unit of programming has shifted from 'one file' to 'one agent,' and soon we'll be forking entire agent organizations.

Karpathy shared a Cursor data chart showing the evolution from Tab completion to Agents. Too conservative means leaving leverage on the table. Too aggressive means creating more chaos than useful work. His advice: the 80/20 rule.

Karpathy spent a weekend running 4 Claude + 4 Codex agents as an ML research team on GPUs. The result: agents are S-tier at implementation but F-tier at experiment design. His key insight — 'You are now programming an organization' — might define agentic engineering in 2026.

Karpathy says coding agents started working in December 2025 as a hard discontinuity. He built a DGX Spark video analysis dashboard in 30 minutes from one English sentence. Programming is becoming agent direction, not typing.

Karpathy argues CLIs are the natural interface for agents precisely because they are legacy tech: text in, text out. His Claude Polymarket terminal demo lands the bigger question: can agents access and use your product?

Karpathy's viral SF AI Startup School talk: software is entering the 3.0 era (English = programming language), LLMs are the new OS but we're in the 1960s. He introduces the 'autonomy slider' and 'Iron Man suit' frameworks, warning that agents are a decade-long journey, not a year.

Karpathy's Claw-era security reality check frames Claws as the next layer above agents, but warns exposed instances, RCE, supply-chain poisoning, and malicious skills can turn productivity into liability. His direction: small core, containers, auditable skills.

Karpathy used Claude Code to build a custom dashboard in 1 hr, reverse-engineering a treadmill API. He believes AI-native sensors & LLMs will enable highly custom, ephemeral apps, rendering the App Store model obsolete. The ultimate goal: 1-min app creation.

Hugging Face CTO Thomas Wolf analyzes how AI fundamentally restructures software: return of monoliths, death of Lindy Effect for legacy code, rise of strongly typed langs, new LLM langs, & open source changes. Karpathy predicts: "rewriting large fractions of all software many times over."

Karpathy's 'art project': a GPT model in 243 lines of pure Python, zero dependencies. Every operation uses atomic math (add, mult, exp, log). Efficiency is secondary. It's the nand2tetris for AI education.

Karpathy used DeepWiki MCP and GitHub CLI to have Claude rip fp8 training out of torchao into 150 self-contained lines that ran faster. His bacterial code idea favors low-coupling, dependency-free code agents can extract and transplant.

Opus 4.6 & Codex 5.3 sped up Karpathy's GPT-2 training by 3 mins. Karpathy failed similar attempts, noting AI's weak open-ended code optimization. Opus deletes comments, ignores CLAUDE.md, and errs. Yet, with oversight, models are useful.

Karpathy open-sourced nanochat — a minimal LLM training framework. With 8 H100 GPUs running for 3 hours at $72, you can train a GPT-2 level model. OpenAI spent $43,000 training the same model in 2019. That's a 600x cost reduction. On spot instances, it's just $20.