Anti-Hype Guide, Part 2: Karpathy

In partnership with

LEADERS

Anti-Hype Guide, Part 2: Andrej Karpathy’s View.

Last week, we saw how Rich Sutton, one of the fathers of Reinforcement Learning (RL), Turing Award in 2025 (the equivalent of the Nobel Prize in Computer Science), bashed the current state of AI, arguing that LLMs were not on the direct course to AGI.

And soon after, Andrej Karpathy, one of the most respected thinkers in the industry (founder of OpenAI, Director of AI at Tesla for several years, and the person who gave the name to things like “hallucinations” and “vibe coding”), had an even harsher take, even reaching the point of saying “RL is terrible,” or that code from AIs is “slop” (jargon for ‘very bad’).

This newsletter will cover things about AI you didn’t even realize you had to be asking yourself. Andrej’s views will make you think about biology, evolution, how dreams could inspire new training methods, and even ghosts.

Finally, he also shared his vision of what might be the defining form factor for AGI, the “cognitive core”. And you’ll be shocked to see how little it has to do with current models.

Let’s dive in.

To say Karpathy is a legend is an understatement.

He is, without a doubt, the most influential figure in the space right now, at least when it comes to shaping how to think about the technology itself.

And while most people expected an optimistic Andrej when he went live on the Dwarkesh Podcast, what we got was a man who had “suddenly” become very pessimistic about a lot of what is going on these days.

The Year of Agents

Since the early innings of 2025, many researchers shared their excitement about 2025 being the ‘year of the agents’.

Probably the clearest example of an ‘AI agent’ in 2025 is coding agents (probably the only one). They are capable of literally building entire apps on their own and have been reported to work for 30 continuous hours.

But does Karpathy use them? No.

In his case, he still relies primarily on autocomplete. That is, he writes the code himself and uses Cursor’s autocomplete mechanism that finishes the coding sentences. No vibe-coding, no nothing.

That’s quite possibly the worst response he could have ever said on this matter. What I will say is that, personally, I do use them a lot in their agent form (e.g., “build this” or “refactor that”), and what’s probably going on here is two things:

I’m nowhere near his quality as a programmer, as one of the best AI engineers on the planet

He’s probably working on really out-of-distribution tasks.

The critical point is the second one, and it explains a lot of the other issues Andrej sees that I’ll outline next.

LLMs are still extremely dependent, or bound, by what they’ve seen during training.

We’ll later give Andrej’s reasons for this, but it's hard to disagree with it. The reason this happens to them is that they are still primarily trained through imitation…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.