Laura Burkhauser: belief

Transcript context

…So as you try to push the frontier on this, I can kind of imagine a couple different strategic directions that you might go in terms of how to get the best performance out of the available models with the various constraints that they have. And maybe you're even doing like multiple of these. But one angle would be to say, okay, well, Claude or GPT or maybe Gemini is going to be probably the best reasoning and tool use agent. So what we really need to do is set whichever one of those we're using up for success. How do we do that? Well, we need to give it a richer understanding of what it's working with. But since it can't natively detect these like awkward moments, for example, maybe we need like an awkward moment detector that we can run and then kind of feed in to the model to flag when these things are happening so that it knows to reason appropriately about that. But then you can imagine a different version where you're like, I've been hearing very good things about GLM 5.1, and I think the weights are out there for this. Maybe we want to try to do something deeper, you know, where we actually teach the core model to understand some of these inputs. And I'm adding video as a modality to a GLM 5.1 doesn't sound easy at all, but you could do some sort of late fusion, you know, cross-training, what have you. And I guess this probably This sort of decision probably depends a lot on what resources you have. Like, do you feel like you can hire the team to do Frontier work at that level? Or is it just so hard to compete with the Frontier Labs for that kind of talent that that's out of range? And it might also depend on like, do we think that open source models, it definitely would also depend on, do we think open source models are going to continue to be competitive Or do we feel like Claude 5 is going to run away from, for compute reasons or, whatever, constitutional reasons, whatever else, if it runs away from the open source bases, then like we can't really keep up even if we do get good at, doing more advanced stuff on open source bases. So I guess to bottom line all that, what's the model strategy? How do you think about where you want to, like what trends you want to bet on carrying you forward? I think that we are betting on the main bet that we've made with our agent is trying to build a very generalized harness and to give the agent access to a bunch of low-level tools, assuming that generalized intelligence is going to get better and better, and that we'll use that we'll use like probably a handful of like whatever model Anthropic or OpenAI or Gemini comes out with that, you know, when, yeah, that when the next cloud model drops, we'll have it evaled within 15 minutes and in the product, and that building for that, at least for the short to medium term, is the right bet to make, rather than investing a lot of time and money and research trying to keep up with the labs. So then it's about like, how do we build an agent harness that's going to be able to instantly take advantage of leaps in general intelligence? How do we not get bitter lessened into not being able to immediately take advantage of those of those leaps. And so that's how we've tried to build our agent. Just give it like a ton of context about the Descript model and also about video editing and how to think about user requests and give it access to our low-level tools. Yeah. And then we do some stuff. We have various experiments that we're doing to sort of make it better through personalization over time. Okay. I mean, I'm very interested in the personalization. Does that boil down to basically saying, though, that you want Underlord to be essentially in the same position as your human users? I mean, obviously, it can't see as well, and it can't hear in the same native way. But subject to the constraints of like some of these things having to be arm's length tool calls to do the sort of sensing, it sounds like aside from that, you're sort of like building one harnessed kind of for both humans and for AIs at the same time. Is that a reasonable way to think about it?…

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