Exactly

Good point

Yup exactly

True. It is the case for most things in life that what one says about a thing says more about them than the thing.

i guess i'm a karpathy-level coder then. had no idea. thanks.

Truth. LLMs are the coding cheat code for us mere mortals. Floor raised, big time

The problem is those people are more like construction workers of code. They are building the same code over and over again, because of inefficiencies in the software industry that are taking care of themselves. They didn't exist before the internet and they shouldn't exist now. AI is accelerating their demise but they would disappear anyway.

It’s not missing. It’s just unsaid because that’s obvious

This does not cut. I am an average coder, nothing exceptional. Last weekend, I had a superb experience with Codex, producing excellent deployed code in few hours which otherwise would have taken days or weeks. But fast forward today, my whole weekend is ruined because Gemini cannot make 'head and tail' of the "--add-cloudsql-instances" flag for the "gcloud run deploy" command. This thing is so stupid at times, worse than an entry-level programmer. But the thing is it costs you precious time! (and endless advils...). Very, very uneven performance. I will be very careful here... Just imagining how much distance to cover before it starts getting used in Medicine. Or even in warfare.

True, LLMs are good enough to build CRUD apps, which is what the majority of apps making money are. You might not be able to vibe code an app that scales to millions of users, but very few apps need that kind of code quality. You can get really far even if the code has some slop.

mmmmmmm not the right take. In order to be a good coder you have to aspire to be a Karpathy level coder not trust LLM's to lift your floor. AI will uplift the willing to learn type individual vs the I need uplifted so lift me individuals.

I'm sorry but being better at writing programs in programming languages than at writing them in natural language isn't something reserved to the top 1/50000 programmers, it's more like the criteria for being slightly above average. Natural language sucks donkey balls, LLM agents are not useful because they use natural language as input. They are useful because they contain a fuckton of compressed solutions to a wide range of problems and it's likely that it covers a significant part of your daily work, because not everything you do needs max brain power, a lot of it is trivial but tedious work.l, and LLMs are awesome at trivial and tedious. The fuzzyness of natural language works *only* because someone at some point did the hard work of extracting precise semantics and convert those to code, code that the LLM has read. If any part of your problem is out of distribution you *will* have to extract those semantics and write them in a formal language anyway, once this is done and you have put *both* in the context, the LLM probably can write the code... But you already did ! The hard part has never been "typing code"...

This interview showed that Karpathy doesn't spend much time among people who are not extremely driven or average-smart. From that perspective it was very naive.

Vibe coding works for simple problems that are very similar to code found on the public internet. Complex, novel problems are another story

What Karpathy is suggesting there isn't limited to Karpathy-level coders: Any decent coder would exercise such due diligence. He's arguing for being able to judge the quality and validate the feasibility of any generated code, especially if money, business or lives are at stake

Not really

Yes he is god level, but that’s not the point. The point is writing software that is in the LLM’s distribution makes you 100x productive. If not you are better off writing code on your own with autocomplete.

But it also gives them the ability to generate endless amounts of garbage code which ultimately is not helpful. I suppose it remains isolated to their codebase so it shouldn't bother me, you or Karpathy. But I also see his point about long term security issues.

I started trying to build a game with 0 game Dev skills (I can't even code) and after 4 months I agree with Andrej. I need and want to baby sit it. Cos it will always make assumptions about something, delete comments that are useful, make inconsistent architectural choices etc

I’m not at his level and yet I can experience what he said. Agents constantly add if and try catch blocks for things no sane developer would do. It over complicates simple algorithms into a mess and so on.

I'm not at Karpathy's level, but I agree with him. I find LLMs most effective when collaborating, asking questions, or requesting feedback incrementally while working on a problem. Ultimately, I'm responsible for the output.

🎯 It’s all a matter of perspective

Yes it’s AGI to the average folks 😘

He also consistently says that he uses tab complete all the time, so he's not chiseling out of the screen from scratch (even though he could).

And they are pretty good at translating between languages. For example if you have some relatively simple web API written in Python, you can just tell them to write it in Go and it just works. A task that could have taken several hours in the past might take minutes.

AI Agents democratize coding by raising the productivity floor for 99% of developers and making computing accessible through natural language

Even for someone who knows a lot about coding, LLMs can help them move much faster

Coding is highly domain-specific. Andrej is super smart to become for example a top developer for new CRM systems, but it would take time and be a waste of his talent.

Excellent point...and therein lies the marvel especially if you are a god-level coder.

"Debt crisis" (2025, 80 x 100 cm) This painting is about the debt crisis in the US and the coming devaluation of the dollar. People all over the world, both poor and rich, will lose their savings. This is the inevitable consequence of the US leftist economic policy. When the king runs out of money, he takes it from his vassals.