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/10-future-of-education-eureka.md
View on GitHubhttps://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=6980
# 10: Future of education: Eureka, ramps to knowledge, how to teach and learn (1:56:20–2:25:14)
Chapter "Future of education", about Karpathy's education project Eureka.
## Why education and not another AI lab
- He feels "some amount of determinism" in what the AI labs are doing. He could help, but not
uniquely. **His big fear is AI happening "on the side of humanity"** with humans
disempowered. He cares about what happens to humans, not just the Dyson spheres, and fears a
*WALL-E* or *Idiocracy* future. Education is where he can add unique value. [1:56:37](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=6997) [1:57:08](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7028)
- **Eureka is "trying to build the Starfleet Academy"**: an elite, up-to-date institution
for frontier technical knowledge. [1:57:39](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7059)
## The bar: a great human tutor
- Today's obvious move (an LLM you ask questions) helps but "still feels... a bit slop".
The capability for a real tutor isn't there yet. [1:58:10](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7090) [1:58:40](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7120)
- **Korean-learning story:** he went from self-study online, to a ~10-person class in Korea, to a one-on-one
tutor. From a short conversation she understood what he knew, probed his world model,
and always served material at the right difficulty, never too hard or too trivial. **"No LLM will do that for you
100% right now, not even close."** With her, "I was the only constraint to learning." [1:59:11](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7151) [1:59:41](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7181) [2:00:00](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7200)
- **So it's not the right time to build the full AI tutor**, though people will build
useful versions and ChatGPT is already "an extremely valuable educational product". In
computer-vision consulting his value-add was often **telling companies not to use AI**. [2:00:31](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7231) [2:01:01](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7261)
## What Eureka is building now
- First, a state-of-the-art course on AI, the field he knows. **nanochat is the capstone
project of LLM101n**, the class he's building. He still needs the intermediate material and a small
team of TAs. [2:01:32](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7292) [2:02:03](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7323)
- **Education is "the very difficult technical process of building ramps to
knowledge."** The target metric is **"Eurekas per second"** (understanding per second):
learners should never be stuck and never bored. [2:02:33](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7353) [2:03:04](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7384)
- **AI's role today is collaborator, not author.** Asking ChatGPT to "teach me AI" gives slop, and "AI is
never going to write nanochat right now". Compared with building CS231n at Stanford (which he
believes was its first deep learning class), LLM101n goes much faster with LLMs doing the boring
parts, but he still does the creative content. Keep calibrating to what exists. [2:04:05](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7445) [2:04:35](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7475) [2:05:05](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7505)
- **How it evolves:** now, faculty working hand-in-hand with AI. Later, some TAs become
AIs answering basic questions from the course materials, while faculty keep the course
architecture. Eventually AI may design better than he does. He will hire faculty for domains outside his
expertise. [2:05:05](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7505) [2:05:36](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7536) [2:06:06](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7566)
- Shape: a **physical institution** (full-time, working through material start to end) as the top tier, and a
**digital tier** below it (internet material plus an LLM assistant, "a bit more gimmicky") that reaches
8 billion people. Re-education matters as jobs change, and many people want to upskill in AI. [2:06:37](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7597) [2:07:07](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7627) [2:07:38](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7658)
## Pre-AGI vs. post-AGI education
- **"Pre-AGI education is useful, post-AGI education is fun"**, like the gym. We don't need
physical strength to move heavy things, yet people lift because it's fun, healthy and
attractive. Learning is hard today and people bounce off material. A perfect tutor is a technical,
tractable problem, and it would make learning anything trivial and desirable. [2:07:38](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7658) [2:08:09](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7689) [2:08:39](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7719) [2:09:09](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7749)
- Long term he calls it "a bit of a losing game", but on a longer horizon than most of the industry expects. People can go
much further than they do now: "anyone speaks five languages", and anyone knows the undergrad curriculum. Dwarkesh: the gym
analogy, where benching two or three plates is now common. Karpathy bets "on some of the timelessness of
human nature" (aristocrats, ancient Greece). [2:09:41](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7781) [2:10:00](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7800) [2:10:30](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7830) [2:11:01](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7861)
- Humans stay in the loop during a transitional period if they understand things. Long term that likely
goes away, and knowledge may become a sport: "what is power lifting in a cognitive era?"
Today's geniuses "are barely scratching the surface of what a human mind can do". [2:12:04](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7924) [2:12:34](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=7954)
- Online courses haven't made everyone know everything because they are "motivation laden": there are
no obvious on-ramps and it's easy to get stuck, and bouncing off material feels like negative reward.
For a while it will be AI plus human collaboration, then maybe just AI. [2:13:37](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8017) [2:14:09](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8049)
## How to teach (advice for educators)
- **Physics "boots up the brain".** Everyone should learn it early: building models and abstractions,
first-, second- and third-order terms, abstracting a noisy system to its fundamental
frequencies. The "spherical cow" is "brilliant thinking". He recommends the book *Scale*
(a physicist on biology), which charts animal scaling laws such as heat dissipation growing with surface area
(squared) while heat generation grows with volume (cubed). [2:14:41](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8081) [2:15:14](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8114) [2:15:45](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8145) [2:16:16](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8176)
- **Find the first-order term and serve it on a platter.** **micrograd** is 100 lines of
Python that do forward and backward passes over a graph of simple ops (plus, times): the
recursive chain rule. "Everything else is just efficiency" (tensors, strides, kernels,
memory movement). [2:16:46](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8206) [2:17:17](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8237) [2:17:47](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8267)
- **Untangle knowledge into a ramp where everything depends only on what came before.**
Dwarkesh: his transformer tutorial starts from a bigram lookup table. Karpathy: **present the pain before the
solution**, so every piece is motivated. [2:18:20](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8300) [2:18:50](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8330) [2:19:22](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8362)
- **Let the student guess first.** Showing the solution before they've tried is "a dick move". Trying
first shows them the action space and the objective, which maximizes "knowledge per new fact". [2:19:22](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8362) [2:19:52](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8392) [2:20:00](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8400)
- **The curse of knowledge** affects him too. Trick: he asked ChatGPT his "dumb questions" about a biology
paper with the paper in context, then shared the thread with the author. He wants people to share
their dumb ChatGPT conversations about his material. [2:20:00](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8400) [2:20:34](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8434) [2:21:07](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8467)
- **The lunch explanation beats the paper.** How someone explains their work over lunch is more
understandable and more accurate than the jargon-filled write-up that "clears its throat for four
paragraphs". At conferences, authors summarize a paper in three sentences over beers: "why isn't
that the abstract?" [2:21:07](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8467) [2:21:39](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8499) [2:22:12](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8532) [2:22:42](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8562)
## How to learn (advice for students)
- No unique tricks: "it's kind of a painful process". **Alternate depth-wise, on-demand learning**
(a project you'll get a reward from) with **breadth-wise learning** (school's "trust me, you'll
need this later"). He loves the on-demand kind. [2:23:13](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8593) [2:23:45](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8625)
- **Explain things to others.** "If I don't really understand something I can't explain it."
Explaining exposes gaps and forces you to reconcile them. [2:24:15](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8655) [2:24:47](https://api.substack.com/feed/podcast/176425744/3127b612a1604474d2b28ab4e9b658f4.mp3#t=8687)
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: 323k | 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: 323k | 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: 323k | Updated | Claude CodeCodex | Verified | Install |