Better AI Models, Better Startups

Every time there's an OpenAI product release now, it feels like there's a bunch of startups waiting with bated breath to see whether OpenAI is going to kill their startup.

This is actually a really, uh, crazy moment for all startups. Adding more types of modalities and more capabilities, uh, per model, the, the better off every startup is.

You have to be on top of these announcements and be... kind of know what you're going to build in anticipation of them before someone else does, versus being worried about OpenAI or Google being the ones to build them.

Welcome back to another episode of The Light Cone. I'm Gary. This is Jared, Harj, and Diana, and we're some of the group partners at YC who have funded companies that have gone on to be worth hundreds of billions of dollars in aggregate. And today, we are at an interesting moment, uh, in the innovation of large language models in that we've seen a lot of really new tech come out just in the last few weeks, whether it's GPT-4o, it's, uh, Gemini 1.5. Harj, how are you thinking about, you know, what does it mean for these models to be so much better?

Anytime I see a new announcement from one of the big AI companies with the release of a new model, the first thing I think about is, what does this mean for the startups, and in particular YC startups? And when I was watching the OpenAI demos, it was pretty clear to to me that they are really targeting consumer. Like, all of the demos were cool consumer use cases and applications, which makes sense. That's kind of what ChatGPT was, was a consumer app that went really viral. I just wonder what it means for the consumer companies that we're funding and, in particular, like, how will they compete with OpenAI for these users? What did you think? Like, what, wh- even if we take it back, like, how do consumer products win from, like, first principles? Like, is it more about the product or the distribution, and how do you compete with OpenAI on either of those things?

Yeah, that's a great question. I mean, I think ultimately it's both. And then, uh, how I want it to be is that the best product wins. Uh, how it actually is, is whoever has the best distribution and a sufficiently good product seems to win. Either way, I actually think we're at, uh, sort of, uh, in this moment where the better the model becomes, if you're already using four and suddenly four, you know, you can, uh, change one line of code and suddenly be using 4o, uh, you basically just get smarter by default every generation. And that's really, really powerful. It means that, you, I think we're entering this moment where the i- the IQ of these things is still, you know, four is arguably around 85. It's not that high. (laughs) And then if the next generation, if Cloud Three really is at 100 or, you know, the next few models end up being closer to, you know, 110, 120, 130, this is actually a really, uh, crazy moment for all startups. And, uh, the most interesting thing is, like, uh, adding new capabilities, so having the same model be great at coding, for instance. Uh, that means that, you know, you might have a breakthrough in reasoning, not through just the model reasoning itself, but you could have the model actually write code and have the code (laughs) do better. And even right now, it seems like there's, um, a lot of evidence that if, instead of trying to prompt the model to do the work itself, you have it write code and you execute the code, it can actually do things that reasoning alone could not do. So adding more types of modalities and more capabilities, uh, per model, the, the better off every startup is.