Sam Schillace: evaluation

Transcript context

…very vulnerable to that stochastic noise, which is like, I totally want to put this on a bumper sticker that like metacognition is susceptible to stochasticity would be like the great bumper sticker. So what, these things are very vulnerable to feedback loops when they're trying to do autonomy, and they're very vulnerable to getting lost. So we've had these, like, multi agent Autonomous agent things get kind of stuck on like complimenting each other, or they'll get stuck on being quote unquote frustrated and they'll go on strike. Like there's all kinds of weird like feedback loops you get into. So what we've learned to answer your question of how you put all this stuff together is You have to, the model's good at thinking, but it's not good at planning. So you do planning in code. So you have to describe the larger process of what you're doing in code somehow. So semantic intent or whatever. And then you let the model kind of fill in the pieces. I'll give a less abstract like example. It's a little bit of an old example. I did this like last year, but at one point I wanted to see if I could generate textbooks. And so I wrote this thing called the textbook factory. And it's, it's tiny. It's like a Jupyter notebook with like. You know, 200 lines of Python and like six very short prompts, but what you basically give it a sentence. And it like pulls out the topic and the level of, of, from that sentence, so you, like, I would like fifth grade reading. I would like eighth grade English. His English ninth grade, US history, whatever. That by the way, all, all by itself, like would've been an almost impossible job like three years ago. Isn't, it's like totally amazing like that by itself. Just parsing an arbitrary natural language sentence to get these two pieces of information out is like almost trivial now. Which is amazing. So it takes that and it just like makes like a thousand calls to the API and it goes and builds a full year textbook, like decides what the curriculum is with one of the prompts. It breaks it into chapters. It writes all the lessons and lesson plans and like builds a teacher's guide with all the answers to all the questions. It builds a table of contents, like all that stuff. It's super reliable. You always get a textbook. It's super brittle. You never get a cookbook or a novel like but like you could kind of define that domain pretty care, like I can describe. The metacognition, the high level plan for how do you write a textbook, right? You like decide the curriculum and then you write all the chapters and you write the teacher's guide and you write the table content, like you can, you can describe that out pretty well. And so having that like code exoskeleton wrapped around the model is really helpful, like it keeps the model from drifting off and then you don't have as many of these vulnerabilities around memory that you would normally have. So like, that's kind of, I think where the syntax and semantics comes together right now. fting off and then you don't have as many of these vulnerabilities around memory that you would normally have. So like, that's kind of, I think where the syntax and semantics comes together right now. And then I think the question for all of us is. How do you get more leverage out of that? Right? So one of the things that I don't love about virtually everything anyone's built for the last year and a half is people are holding the hands of the model on everything. Like the leverage is very low, right? You can't turn. These things loose to do anything really interesting for very long. You can kind of, and the places where people are getting more work out per unit of work in are usually where somebody has done exactly what I just described. They've kind of figured out what the pattern of the problem is in enough of a way that they can write some code for it. And then that that like, so I've seen like sales support stuff. I've seen like code base tuning stuff of like, there's lots of things that people are doing where like, you can get a lot of value in some relatively well defined domain using a little bit of the model's ability to think for you and a little, and a little bit of code. And then I think the next wave is like, okay, do we do stuff like domain specific languages to like make the planning capabilities better? Do we like start to build? More sophisticated primitives. We're starting to think about and talk about like power automate and a bunch of stuff inside of Microsoft that we're going to wrap in these like building blocks. So the models have these chunks of reliable functionality that they can invoke as part of these plans, right? Because you don't want like, if you're going to ask the model to go do something and the output's going to be a hundred thousand lines of code, if it's got to generate that code every time, the randomness, the stochasticity is like going to make that basically not reliable. You want it to generate it like a 10 or 20 line high level semantic plan for this thing that gets handed to some markup executor that runs it and that invokes that API, that 100, 000 lines of code behind it, API call. And like, that's a really nice robust system for now. And then as the models get smarter as new models emerge, then we get better plans, we get more sophistication. In terms of what they can choose, things like that. Right. So I think like that feels like that's probably the path forward for a little while, at least, like there was, there was a lot there. I, sorry, like I've been thinking, you can tell I've been thinking about it a lot. Like this is kind of all I think about is like, how do you build. Really high value stuff out of this. And where do we go? Yeah. The, the role where we are. Yeah. The intermixing of code and, and LMS is, is a lot of the role of the AI engineer. And I, I, I think in a very real way, you were one of the first to, because obviously you had early access. Honestly, I'm surprised. How are you so hands on? How do you choose to, to dedicate your time? How do you advise other tech leaders? Right. You know, you, you are. You have people working for you, you could not be hands on, but you seem to be hands on. What's the allocation that people should have, especially if they're senior tech…

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