mean we kind of look at large language models as a front end an inner face looking for a back end but I think if you're not you know working with technology every day it's easy to think that it's this complete solution that can do all these things that can't do but but obviously there's it lowers the barriers between humans and machines but but there's a huge missing component to it it's just such a small piece of of what automation entails and and I think you know we talk a bit about in the book the need to just iterate and get going right away especially with with just invisible machine type designs where you're creating these automations like you know find a safe space I think in the hbr article you mentioned mass general Brigham kind of taking that approach where where you know you find the people on the team who understand what you're trying to automate and have them working with the technology to kind of create those automations and and chat GPT certainly it might make that easier it certainly I think helps people see that as a path to communicating with machines and kind of shortens the on ramp a bit but but it is there's like so much collateral damage and I guess it's probably different with robotics too because you can't really like test as willy nilly as you might with something more abstract one

of the things I worry about and and I've seen you know because it's such a low barrier to entry to plug in a new GPT driven chat bot in your software you see a lot of companies I think using it in ways that could be damaging because you can't really predict where it's going to make errors right it could just have factual inaccuracies in totally random places and you don't really know how it came up with that or how it could be fixed in the future and if you give it the same prompt again it might not have that inaccuracy where if you're dealing with so for example if we're asking a student to do a task we have a we have some idea of the high likelihood errors they would make in the low likelihood errors they would make and it would make it easier for us to fact check that document and edit it or whatever the task is so I do worry in particularly tasks where quality control needs to be high and and like you say precision needs to be high that the the task of editing and quality checking the output from these type of probabilistic models it's really challenging and I think it's going to take a different type of fact checking mind than is currently present in some industries so I'm encouraged in part by the experimentation I know there's been a lot of kind of the marketing and sales world around drafting emails or finding you know customer audiences where you know if you make a mistake it's probably not the end of the world but the the idea that the the technology just is never the plan isn't to make it completely robust because that's not how it's designed that limits the amount of use cases that that it can can operate on its own and I think it also makes it so that there's a clearly a skill shift where certain skills like fact checking editing supervising that type of algorithm are really the demand for those skills we imagine could

go