The models are trained on vast data but don't actually know how to do specific work like medical billing workflows—that knowledge lives in office managers and historical practice data, not on the internet.
Frederick explains that despite their scale, AI models don't inherently know how to execute domain-specific workflows like insurance billing, which Lassie had to learn by doing the work itself and analyzing historical data from practice management systems. ✦ AI generated
Frederick · a16z Podcast · 2026-07-30 · original ↗
starts at this moment · 46:49
“what are the hardest problems to solve like what is it when we talk about like software that does the job of labor what cannot be done right now”
The models are trained on so much data and they're so large and yet they actually don't really know how to do any of this work. Like they don't have the workflows encoded in any way... there's also a big amount of um just like human knowledge that is encoded in say these office managers and they just like know how to do this work uh that's weirdly not that accessible on the internet. Um I think we have a big advantage there because we have um all of this like historical data out of their ERPs that we can look at and kind of infer you know some of these workflows from.
verbatim transcript · starts at 46:49
46:30technology and what it enables at large where does work still need to be done >> where it's just not quite good enough um and then what do you think the curve of that looks like when so it's not you know it's like a question of AGI for small business like you know what do you need and where are we on that curve if you had to estimate
46:49>> I think one thing that's interesting is that the models are trained on so much data and they're they're so large and yet they actually don't really know how to do any of this work. Like they don't have the uh workflows encoded in any way. Um so for example, we're working on a product now where we we have to like collect all of these like basically SOPs
47:11on like and documents about like how are you supposed to bill insurance claims to certain payers and all this kind of stuff. uh which uh to some extent humans would do the same but there's also a big amount of um just like human knowledge that is encoded in say these office managers and they just like know how to do this work uh that's weirdly not that accessible on the internet um
47:34>> I think we have a big advantage there because we have um all of this like historical data out of their ERPs that we can look at and kind of infer you know some of these workflows from um but that's something we notice a lot I think actually when we started using some of the, you know, later reasoning models, uh, we kind of assumed like, oh, they probably just
47:52know how to do this work because like why would they not, right? Like they're trained they're trained on on all of this data. Um, but it it it turns out that they they don't know all the intricacies of most of these workflows. I think um, yeah, this is like less specific to Lassie, but I'm I'm personally kind of excited about uh smaller models that uh, you know, learn
48:12faster and can learn on less data. I think uh that will be uh that will be really cool to see because I you know I think over time um we'll have intelligence kind of I think disseminated everywhere uh and uh you know like to the point where like you know like the the Pixar lamp that has its own personality like why why not I I I feel like I I want my intelligence to