back
3 comments
Specifically it looks like he's planning to extend the ideas from https://github.com/karpathy/autoresearch into a larger effort towards recursive training improvement [1]:

> Excited to welcome Andrej to the Pretraining team! He'll be building a team focused on using Claude to accelerate pretraining research itself. I can’t think of anyone better suited to do it — looking forward to what we build together!

[1] https://x.com/nickevanjoseph/status/2056760504949842219

So he's working on the singularity
Am I the only one who wasn’t particularly impressed by AutoResearch? If you looked at what the agent was actually doing, it was just tuning parameters mostly, not really trying different novel approaches.

I couldn’t help myself but consider this mostly a very inefficient variant of hyperparameter optimization, but someone correct me if I’m wrong, I may be looking at this too pessimistic.

I didn't dig into what the actual repository was doing, but personally, I took some inspiration from the idea after reading about it and realizing that I might have been underestimating the ability of LLMs. I put a bit more work into a performance harness I was using locally and just set some agents to brainstorming and they did seem to find some great stuff. So I don't really have a stance one way or another on this specific repo, but the general idea seems like a really good one.
Could you elaborate in specifics how you had been underestimating models? Ypu mean just using more tighter harnessing to make them work in structured agentic eay or something else?
The specific code I was working on, I had a general idea of the sort of performance improvement that would be possible. I just thought that it would be too hard for the models to figure out without a lot of hand-holding.

But it ended up being not "too hard ever", but more like, in 1 out of every 5 tries, the model did in fact manage to get a large refactoring to the point where it improved performance. So once I set it up to try something, use the perf test, see if it worked, if not, throw it away, repeat. Then it started, slowly, finding some useful things.

Just remember that the will do clever but useless things to improve. Like changing the random seed as per autoresearch's hero image. lol! imo, out of the box thinking is needed.
Ever since AlphaEvolve - the idea that if you build a harness which can evaluate solutions and give LLMs a database where they can keep storing their work and then sample from it - they do find non-trivial solutions over time leaning from their own past ideas.

It is the ultimate manifestation of test-time scaling. I think karpathy just popularised it.

Karpathy embedded within an organization is way more impressive than him out on his own with hot takes and little projects. I hope he does great things for Anthropic.
Absolutely, I wasn’t saying that him being at Anthropic wasn’t going to be effective, I just think his little projects wouldn’t be very interesting if his name wasn’t attached to them.
I was impressed that I was able to take the same basic idea and apply it to anything that a Claude could construct a metric for. It's nice being able to just run /autoresearch and speed up your test suites, and shave time off your builds etc.

It's a decent tool to have in the toolbox.

I was trying to look options outside the box (everything is more context or RAG) and been using this approach for about a month with good results. https://github.com/VDP89/fscars

    > Am I the only one who wasn’t particularly impressed by AutoResearch?
isn't it just a nerfed AlphaEvolve? https://arxiv.org/abs/2506.13131
Inefficient variants with $100m+ worth of compute will still probably outperform the best team of researchers
That's not the question. The question is how much you need to give the best team of researchers to beat $100m+ worth of compute. $1m of compute? $10m? Clearly giving the best team $100m is going to beat out giving an efficient group $100m. It does in fact matter who you throw your money at...
I guess we must expect it at this point. But funny that has model written tokens like ’ instead of '
More like he'll blog and tweet about using Claude and get gullible software engineers to buy Claude subscriptions and work on their own obsolescence while paying for it.

Many people are still deluded and think he is the same person who wrote the informal AI tutorials in plain html. He isn't, he is selling stuff now.

I'm as jaded as can be but I think Anthropic is now beyond the point where they'd place much value on farming Karpathy's name recognition. I'm sure they considered it an extra plus in his hiring package but they wouldn't do the level of comp package he'd want if they didn't believe the odds were decent that he'll contribute serious value.

Sure, it can always not work out but that's no more a risk with him than any high-profile hire who doesn't really need the money and will always have other options.

What is he selling? How is this time different compared to when he was at OpenAI or at Tesla? You could say he was shilling those products too. I don't see any shift. He's still posted free in depth YouTube videos recently.
FYI, Karpathy has 2.5M followers on twitter, Anthropic has 1.3M (OpenAI has 4.8M, for comparison). I'm sure Karpathy will be doing mostly research and will make real contributions, but I also think it would be naive to ignore the weight of his voice. It's not negligible, nor is it the only thing he brings.
> What is he selling?

Is that a serious question? He already promoted vibe coding and AI hype. Now he is literally there to promote Anthropic and its IPO price.

When he was at OpenAI it wasn't overtly commercial yet. At Tesla he had a way lower profile. Now he is the vibe coding Jesus for deluded software engineers. The impact is much larger.

> At Tesla he had a way lower profile.

?

He was literally rolled out in front of camera as Tesla's AI prodigy at multiple streamed events designed to appeal to techy consumers and dev recruitment. He's definitely been one of AI's public personas for a long time now, and his employers have regularly aided/directed/utilized him accordingly.

I think he's just genuinely excited about the capabilities.

(I do understand that for Anthropic it's a brand boost as well, just like signing other prominent researchers, as it was with LeCun and Meta etc).

This is good branding move for Anthropic. Karpathy is well respected among ML crowd.
Speaking of, how did he not lose credibility at “full self driving next year, better buy it now”-Tesla?

It might be Elon who went and said that and said they don’t need lidar, but as director of AI and auto vision Karpathy bears the responsibility for those features.

>Speaking of, how did he not lose credibility at “full self driving next year, better buy it now”-Tesla?

That I also want to know. He bailed out of Tesla right when the limitations of his "LIDAR-less cameras only self driving" system were becoming obvious, and nobody asked him about the hindsight of this obvious fuckup.

>but as director of AI and auto vision Karpathy bears the responsibility for those features.

Exactly. You lead the R&D, so it's on you. If your boss makes stupid decisions in public overriding your best judgement, the leave and go somewhere where your decisions be respected. The ML market was red hot for people like him back then so it's not like he didn't have alternatives.

Although I doubt Elon forced that idea on him, since he's the one who was confidently claiming that vision only is better since Lidar pollutes the sensor fusion data.

Guess his boyish looks and his videos educating outsiders and students about AI contributed..

Elon makes it so easy to hate him as much as to admire. No comparison.

>Guess his boyish looks and his videos educating outsiders and students about AI contributed..

IDK but if I judge a head of ML, I'd care more about whether he has delivered technically, not if he's good at PR.

The ZIRP era lead to SV ML scene to be full of false prophets who were just good at PR and advertising for their employer because they were popular with the industry hype, than at actually getting shit that works delivered out the door.

It's good that the money and hype bubbles bursts every now and then as only then are the scammers, false gods, and naked emperors revealed and can the house get cleaned of them.

>Although I doubt Elon forced that idea on him, since he's the one who was confidently claiming that vision only is better since Lidar pollutes the sensor fusion data.

How do you get a labeled dataset for your ML model? Obviously you use lidar, except the engineers had to do it in secret because Elon has fired engineers that used lidar even during the development process.

> vision only is better since Lidar pollutes the sensor fusion data.

Did he never experienced optical illusions? I don't get it.

With a cursory glance at Tesla's hardware the rest of the self driving car industry quickly surmised that it was at the time nowhere near sufficient to to deliver L4 autonomy, and that's before sensor modalities entered the equation. Karpathy was either BSing for money, or he actually believed the hype. Either way it was a bad look.

  > it was at the time nowhere near sufficient to to deliver L4 autonomy
It still isn't. Show me a video based only vision system that's even L3. Current systems like Waymo are using video, radar, and lidar. Even lidar isn't enough. Each of them provides a different benefit. I mean radar can even do something humans can't, see in fog, why would you not want radar? And I mean at that time Tesla was using 1.2MP cameras...

Truth is what they did was impressive, but given their constraints, but the constraints were self-imposed and unnecessary. So it's more impressive in the way that it is impressive when someone that was shot crawled their way to the hospital. Still impressive even if they shot themselves, but it sure does make you feel a little different...

The original Autopilot team had some great talent. Elon fired them and then sued them because they wouldn't shut up about Lidar. Elon couldn't say "Every new Tesla is FSD ready" if every new Tesla needed Lidar for that to be true, that was the constraint.

The Woz was on TV in an interview saying he bought 2 Model 3s, one for himself and one he could put to work after the robotaxi update. Regular everyday suckers too were taking on crippling debt to buy Teslas because their cars were gonna pay for themselves after the magical update.

>The original Autopilot team had some great talent. Elon fired them and then sued them because they wouldn't shut up about Lidar.

OK, so then why didn't Karpathy also leave in protest? Unless he also supported Elon's decision in which case he's complicit and just as responsible as ELon.

Minor celebrity fwiw - deserved though.
He's a celebrity all right - promoting bunk just like the rest of that lot, along with his Lord and Savior Elon Musk.
Why do they need this when they have the next gen mythos? Surely that can manage everything?
You don't understand: no ones ever reading more than 1% of the training material; so they need someone who has reduced that to 0.1%. The less you know, the more you know!