michalstrnadel/source-to-skill / examples/karpathy-agi-decade-away
karpathy-agi-decade-away
VerifiedAndrej Karpathy on the Dwarkesh Podcast ("AGI is still a decade away", Oct 2025, 2h25m) distilled. Covers why agents need a decade (no continual learning, cognitive deficits), "ghosts not animals" and pre-training as crappy evolution, the cognitive core (~1B params), why RL is terrible ("sucking supervision through a straw"), LLM judges being gamed, silent model collapse, why coding agents failed on nanochat, the autonomy slider, AGI blending into 2% GDP growth, loss of control, the evolution of intelligence and LLM culture/self-play, self-driving's "march of nines", and Eureka's approach to education (ramps to knowledge, Eurekas per second). Load when reasoning about AI timelines, agent capabilities, RL limitations, coding-agent usage, AI's economic impact, or how to teach and learn technical material.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 8.3 KB |
| cheatsheet.md | markdown | 11.9 KB |
| evals/evals.json | config | 5.7 KB |
| evals/last-run.json | config | 2.7 KB |
| segments/01-decade-of-agents-ghosts-not-animals.md | markdown | 7.9 KB |
| segments/02-missing-brain-parts-and-nanochat.md | markdown | 5.5 KB |
| segments/03-coding-agents-and-autonomy-slider.md | markdown | 6.3 KB |
| segments/04-rl-is-terrible.md | markdown | 5.8 KB |
| segments/05-how-humans-learn-collapse-cognitive-core.md | markdown | 8.7 KB |
| segments/06-agi-progress-jobs-and-why-coding-first.md | markdown | 6.3 KB |
| segments/07-asi-loss-of-control-and-2-percent-growth.md | markdown | 6.8 KB |
| segments/08-evolution-of-intelligence-and-culture.md | markdown | 5.6 KB |
| segments/09-self-driving-march-of-nines.md | markdown | 7.4 KB |
| segments/10-future-of-education-eureka.md | markdown | 12.5 KB |
| source.json | config | 1.3 KB |
segments/03-coding-agents-and-autonomy-slider.md
View on GitHubhttps://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1785
# 03: LLM cognitive deficits, seen through coding (29:45–40:05)
Chapter "LLM cognitive deficits". Dwarkesh opens with Karpathy's tweet that coding
models were of very little help in building nanochat.
## Three ways people code today
- nanochat took "a bit more than a month". The three classes: [29:39](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1779) [30:00](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1800)
1. **Reject LLMs, write from scratch.** "Probably not the right thing to do anymore."
2. **Write it yourself plus autocomplete** (Karpathy's spot). You remain the
architect, tab through suggestions and edit when they're wrong.
3. **Vibe coding / agents.** "Please implement this", then let the model do it. [30:32](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1832)
- **Agents are good at** boilerplate and at code that occurs often on the internet,
because there are lots of examples in the training sets. [30:32](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1832) [31:05](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1865)
## Why agents failed on nanochat
- nanochat is unique, "intellectually intense" code where everything has to be precisely
arranged. Models kept misunderstanding it because they remember the typical way of
doing things. [31:05](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1865) [31:38](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1898)
- **Concrete example:** for 8 GPUs he skipped PyTorch's DDP (DistributedDataParallel)
container and wrote his own gradient synchronization inside the optimizer step.
The models kept trying to make him use DDP and "just couldn't internalize it". [31:38](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1898) [32:09](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1929)
- **Other failure modes:** over-defensive try/catch everywhere, trying to make
a "production code base", bloating code and complexity, and deprecated APIs. "It's total
mess." [32:09](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1929) [32:41](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1961)
- **Bandwidth argument:** typing out in English what you want is too much
typing. Navigating to the spot and typing the first three letters lets
autocomplete produce the code. That is a very high-information way to specify intent. [32:41](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1961) [33:14](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1994)
## Where he did use models
- The report generation (boilerplate-y, not mission-critical) was partly vibe-coded.
[33:14](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=1994)
- Rewriting the tokenizer in Rust, a language he's newer to: he had a Python implementation he
fully understood plus tests, so it felt safe. **LLMs increase accessibility to
languages or paradigms you're less familiar with**, especially ones with lots of code online. [33:45](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2025) [34:15](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2055)
## Implications for an intelligence explosion
- Dwarkesh: the fast-takeoff story is AI automating AI research ("AI 2027"), yet models are
asymmetrically bad at exactly that. Karpathy agrees this is part of why his timelines are
longer: models are "not very good at code that has never been written before". [34:15](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2055) [34:51](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2091)
- Even known tricks from papers (e.g. adding RoPE embeddings) get integrated wrongly, because
models can't fit them to the repo's style and assumptions. [35:27](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2127)
- His go-to model is **GPT-5 Pro, "the Oracle"**. With 20 minutes to spare he pastes in the whole
repo and asks questions, and it's "surprisingly good compared to what existed a year
ago". [36:02](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2162)
- **On industry hype:** "it's making too big of a jump... it's trying to pretend like
this is amazing. And it's not, it's slop", possibly because of fundraising. Models are amazing but need
a lot of work. **"For now, autocomplete is my sweet spot."** [36:02](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2162) [36:33](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2193)
## AI as a continuum of computing: the autonomy slider
- Dwarkesh: compilers, linting and better languages raised productivity without an explosion,
and autocomplete looks like that category. [36:33](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2193) [37:04](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2224)
- Karpathy sees AI as an extension of computing. Recursive self-improvement has been
under way since code editors, syntax highlighting, type checking and search engines
(Google saw itself as an AI company early on). [37:36](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2256) [38:07](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2287)
- **Autonomy slider:** compilers took assembly away from us, and now more and more of
what can be automated is automated while humans move up a layer of abstraction. [38:07](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2287) [38:37](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=2317)
- (2317–2400 is a sponsor read.)
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |