Alessio Fanelli: prediction

Transcript context

…I think one thing that's interesting here, one obviously is that in AI, you kind of get paid to promise things and then you, to deliver them, you know, people have a lot of patience. I think that patience has come down over time. One example here is Devin, right this year, where a lot of promise in March and then, and then it took nine months to get to GA. Uh, but I think people are still coming around now and Devin, Devin's [00:55:00] product has improved a little bit, hasn't he? Even you're going to be a paying customer. So I think something Devon like will work. I don't know if it's Devon itself. The Auto GPT has an interesting second layer in terms of what I think is the dynamics going on here, which is a very AI specific layer. Over promising under delivering applies to any startup, but for AI specifically, there's this promise of generality that I can do anything, right? So Auto GPT's initial problem was making money, like increase my net worth. And I think. That means that there's a lot of broad interest from a lot of different people who are trying to do all different things on this one project. So that's why this concentrates a lot of stars. And then obviously, because it does too much, maybe, or it's not focused enough, then it fails to deploy. So that would be my explanation for why the interest to usage ratio is so low. And the second one is obviously pure execution, like the team needs to have a vision and execute, like half the core team left right after AI Engineer Summit last year. [00:56:00] That will be my explanation as to why, like this promise of generality works basically only for ChatGPT and maybe for this year's Notebook LM. Like, sticking anything in there, it'll mostly be direct. And then for basically everyone else, it's like, you know, we will help you complete code, we will help you with your PR reviews. Like, small things. Alright, code interpreting, we talked about a bunch of times. We soft announced the E2B fundraising on this podcast. Code sandbox got acquired by Together AI. Last week, um, which are now also going to offer as an API. So, uh, more and more activity, which is great. Yeah. And then, uh, in the last step, two episodes ago with Bolt, we talked about the web container stuff that we've been working on. I think like there's maybe the spectrum of code interpreting, which is like, You know, dedicated SDK. There's like, yeah, the models of the world, which is like, Hey, we got a sandbox. Now you just kind of run the commands and orchestrate all of that. [00:57:00] I think this is one of the, I mean, it'd be screwed. That's just been crazy just because, I mean. Everybody needs to run code, right? And I think now all the products and the everybody's graduating to like, okay, it's not enough to just do chat. So perplexity, which is a easy to be customers, they do all these nice charts for like finance and all these different things. It's like the products are maturing and I think this is becoming more and more of kind of like a hair on fire. problem, so to speak. So yeah, excited to see more. And this was one that really wasn't on the radar when we first wrote the four wars. Yeah, I think mostly because I was trying to limit it to Ragnops. But I think now that the frontier has expanded in terms of the core set of tools, core set of tools would include Code interpreting, like, like tools that every agent needs, right? And Graham in his state of agents talk had this as well, which is kind of interesting for me. Cause like everyone finds the same set of things. So it's basically like someone, [00:58:00] everyone needs web browsing. Everyone needs. Code interpreting, and then everyone needs some kind of memory or planning or whatever that is. We'll discover this more over time, but I think this is what we've discovered so far. I will also call out Morphlabs for launching a time travel VM. I think that basically the statefulness of these things needs to be locked down. A lot. Basically, you can't just spin up a VM, run code on it, and then kill it. It's because sometimes you might need to time travel back, like unwind, or fork, to explore different paths for sort of like a tree search approach to your agent development. I would call out the newer ones, the new implementations as The emerging frontier in terms of like what people kind of are going to need for agents to do very fan out approaches to all this sort of code execution. And then I'll also call out that I think chat2bt canvas with what they launched in the 12 days of shipmas that they announced has surprisingly superseded Code Interpreter. Like [00:59:00] Code Interpreter was last year's thing. And now canvas can also write code and also run code. And do more than Code Interpreter used to do. So right now it has not killed it. So there's, there's a toggle box for Canvas and for Code Interpreter when you create a new custom GPTs. You know, my, my old thesis that custom GPTs is your roadmap for investing because it's, it's what everyone needs. So now there's a new box called Canvas that everyone has access to, but basically there's no reason why you should use Code Interpreter over Canvas. Like Canvas has incorporated the diff mode that both Anthropic and OpenAI and Fireworks has now shipped that I is going to be the norm for next year. Uh, that everyone needs some kind of diff mode code interpreter thing. Like Aitor was also very early to this. Like the Aitor benchmarks were also all based on diffs and Coursera as well.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.