Andrej Karpathy's talk on the future of the industry • Hacker News Board

Re: Andrej Karpathy's talk on the future of the industry

#191

Post by msgodel » Thu, Jun 19, 2025, 11:48 PM UTC

alganet wrote: post #190

msgodel wrote: post #189

Earlier quoted context omitted.

Are they focusing on beginners? I think everyone in tech has always been excited about bringing more people in but I don't feel like there's more of that here than anywhere else. Have you tried building things with this stuff? I have, it's exciting, there's all kinds of new things to build. I haven't been as excited to build things for a while before this.

> Are they focusing on beginners? That's what their PR material seems to imply. > Have you tried building things with this stuff? Yes. > it's exciting, there's all kinds of new things to build Yes, and no. I don't want to pay to build, or be dependent on an API or model delivery. It means I want to build the juicy stuff, the models themselves. Generate synthetic data, experiment with different approaches for training…

Yeah I think the hype is over inflating prices but all the expensive stuff is exactly that: hype.

I've self hosted everything except training, I've been using colab for that, not because it's faster (it isn't) but because the heat is awful.

Re: Andrej Karpathy's talk on the future of the industry

#192

Post by alganet » Fri, Jun 20, 2025, 8:58 AM UTC

msgodel wrote: post #191

alganet wrote: post #190

Earlier quoted context omitted.

> Are they focusing on beginners? That's what their PR material seems to imply. > Have you tried building things with this stuff? Yes. > it's exciting, there's all kinds of new things to build Yes, and no. I don't want to pay to build, or be dependent on an API or model delivery. It means I want to build the juicy stuff, the models themselves. Generate synthetic data, experiment with different approaches for training…

Yeah I think the hype is over inflating prices but all the expensive stuff is exactly that: hype. I've self hosted everything except training, I've been using colab for that, not because it's faster (it isn't) but because the heat is awful.

I think AI companies overdid it, and they're struggling to get high-value collaborators to their ecosystem.

Before, Google would launch a new Android version. Or Facebook would launch React, and developers would flock en masse to the ecosystem and the new features. These are very low-barrier entry ecosystems. All you need is a cheap computer or phone, which you probably already have for other reasons.

I get the impression that this is not happening with AI tech. Whoever is flocking to these is not contributing much to improve the ecosystem. There's a lot of users, but they're not doing anything interesting, and it's not getting the traction they expected.

I could be wrong, but I think I'm on the right track on this one. High-value collaborators know that they can jump in at any time, there's nothing really special about it that requires the early investment, and currently there is no clear reward for being an early adopter.

Re: Andrej Karpathy's talk on the future of the industry

#193

Post by Vegenoid » Tue, Jun 24, 2025, 2:38 PM UTC

ethbr1 wrote: post #167

Vegenoid wrote: post #132

Earlier quoted context omitted.

> people interface with LLMs directly rather than software at all What does this mean? An LLM is used via a software interface. I don’t understand how “take software out of the loop” makes any sense when we are using reprogrammable computers.

The strongman for it would be: Our current computing paradigm is built on APIs stacked on APIs. Those APIs exist to standardize communication between entities. LLMs are pretty good at communicating between entities. Why not replace APIs with some form of LLM? The rebuttal would be around determinism and maintainability, but I don't think the strongman argument is weak enough to dismiss out of hand. Granted: these wou…

Maybe this is a failure of my imagination, but I still don’t understand how “replace an API with an LLM” makes sense. LLMs just generate text, the only way they have of interacting with software is by generating text that the software can parse - a.k.a. an API call.

Maybe I have some misconception here. I think seeing a program or system that is doing this “replace APIs with LLMs” thing would help me understand.