David Luan: recommendation

Transcript context

…So like, you know, you're a portfolio allocator. You have people know about Persimmon, people know about Fuyu and Fuyu Heavy. Can you take us through like how you think about that evolution of that and what people should think about what that means for adepts and sort of research directions? Kind of take us through the stuff you shipped recently and how people should think about the trajectory of what you're doing. The critical path for adepts is we want to build agents that can do a higher and higher level abstraction things over time, all while keeping an insanely high reliability standard. Because that's what turns us from research into something that customers want. And if you build agents with really high reliability standard, but are continuing pushing a level of abstraction, you then learn from your users how to get that next level of abstraction faster. So that's how you actually build the data flow. That's the critical path for the company. Everything we do is in service of that. So if you go zoom way, way back to Act One days, right? Like the core thing behind Act One is can we teach large model basically how to even actuate your computer? And I think we're one of the first places to have solved that and shown it and shown the generalization that you get when you give it various different workflows and texts. But I think from there on out, we really realized was that in order to get reliability, companies just do things in various different ways. You actually want these models to be able to get a lot better at having some specification of some guardrails for what it actually should be doing. And I think in conjunction with that, a giant thing that was really necessary is really fast multimodal models that are really good at understanding knowledge work and really good at understanding screens. And that is needs to kind of be the base for some of these agents. Back then we had to do a ton of research basically on how do we actually make that possible? Well, first off, like back in forgot exactly one month to 23, like there were no multimodal models really that you could use for things like this. And so we pushed really hard on stuff like the Fuyu architecture. I think one big hangover primarily academic focus for multimodal models is most multimodal models are primarily trained on like natural images, cat and dog photos, stuff that's come out of the camera. Coco. Yeah, right. And the Coco is awesome. Like I love Coco. I love TY. Like it's really helped the field. Right. But like that's the build one thing. I actually think it's really clear today. Multimodal models are the default foundation model, right? It's just going to supplant LLMs. Like you just train a giant multimodal model. And so for that though, like where are they going to be the most useful? They're going to be most useful in knowledge work tasks. That's where the majority of economic value is going to be. It's not in cat and dogs. Right. And so if that's what it is, what do you need to train? I need to train on like charts, graphs, tables, invoices, PDFs, receipts, unstructured data, UIs. That's just a totally different pre-training corpus. And so a depth spent a lot of time building that. And so the public for use and stuff aren't trained on our actual corpus, it's trained on some other stuff. st a totally different pre-training corpus. And so a depth spent a lot of time building that. And so the public for use and stuff aren't trained on our actual corpus, it's trained on some other stuff. But you take a lot of that data and then you make it really fast and make it really good at things like dense OCR on screens. And then now you have the right like raw putty to go make a good agent. So that's kind of like some of the modeling side, we've kind of only announced some of that stuff. We haven't really announced much of the agent's work, but that if you put those together with the correct product form factor, and I think the product form factor also really matters. I think we're seeing, and you guys probably see this a little bit more than I do, but we're seeing like a little bit of a pushback against the tyranny of chatbots as form factor. And I think that the reason why the form factor matters is the form factor changes what data you collect in the human feedback loop. And so I think we've spent a lot of time doing full vertical integration of all these bits in order to get to where we are.…

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