AI for small-, mid-caps and local government. Currently in a start-up bringing AI into compliance and legal. Only AI-news, where possible from original sources.
@emilylambert@consumervijay 8/n Any process where you have well defined criteria, and where it's about triaging unstructured data in those criteria: prime candidates. custom builds however: different story: either models can make the decisions for you and you have no moat, or you'll make a lot of decisions
@emilylambert@consumervijay 8/n if I say buy houses cheap, renovate them, sell them: then you can run 24/7 screening and have the best prospects delivered to you 24/7, better, faster and cheaper than any human could bring that to you...
@emilylambert@consumervijay 7/n the models are needed because you need text analysis, but the process is a loop where new information is fed in and triaged... anything that fits that bill in any way shape or form could indeed be overnight employees that will these days perform better than most humans
@emilylambert@consumervijay 6/n there's exaptions: big companies. Say I'm consultant for say strategic marketing and I offer a service to guard possible predators of a market for a portfolio of clients: sure you can run that overnight 24/7 and have reports spit out:
@emilylambert@consumervijay 5/n anyone who ever had a really specific and defined idea about what the outcome should be will tell you a whole other story than "I have a shitload of agents running and my head doesn't hurt from all decisions that were involved to keep that going 24/7"
@emilylambert@consumervijay 4/n and all those things are still on the edge of where model harnesses are today... so pretty sure: most people saying "I have 7 different projects non-stop running" is just the equivalent of letting an AI generate an image and to confuse that with "that's exactly what I needed"
@emilylambert@consumervijay 4/n you front run all your decisions, you overcame the models tendency ask you stuff, to keep working, even when it does have to ask you stuff, if you conveyed what it should decide on itself and if you have a process that keeps all that in check and survives context compactions
@emilylambert@consumervijay 3/n those 2 hours of decision making bought me an overnight run of 1.5 days. So what this bought me is that I traded 2.5 hours for 8 hours of faster result. Those things are a reality and you can do that with any well-defined project, if:
@emilylambert@consumervijay 2/n so I'm building a data-store that unifies all that data cleaning. I have specific needs and I can probably express those needs in about 4 or 5 hours of decision making, but in practise things ain't perfect, so let's say I manage to convey 70% of that...
@emilylambert@consumervijay 1/n I can give you my example: I daily run analysis on data. I get that data from various sources. These are all deterministic processes, no models needed. But what I noticed is that I use that data in various analysis and that I tend to do the same data-cleaning ad-hoc
@consumervijay@emilylambert 6/n so it is possible, but hype as well: managing builds through the night, will demand decision making for the bulk of my day and not the easy ones you can outsource or have decided for you... and this will fry your brain, cause you'll be constantly on the edge of your abilities
@consumervijay@emilylambert 5/n and you'll going to have to have a process between "decisions I need and want to make" and "decisions that the model can decide better than I can" and you'll have to have a process in place to fight the urge of models to come ask you stuff they can decide better
@consumervijay@emilylambert 4/n to have overnight projects build exactly what you need, you have to front-run all your non-obvious decisions. Otherwise it's like generating a random image.
For the 1.5 day run, I had a session of probably 2.5 hours non-stop decision making. So the night ain't 'free'
@consumervijay@emilylambert 3/n You can ask an AI 'generate me an image', let it work all night, come back and say "wow, great image!'. What I see most people do when they say "it works overnight": they take a model with Einstein intelligence and they use it to create a digital typewriter nobody needs...
@consumervijay@emilylambert 2/n in order to do that I had to setup a 'work on this project'-structure that front-runs most of the decisions it needs and then stays in line with those decisions. So it's not "it magically works over night"... it can, but there is a very big distinction to be made in that:
@consumervijay@emilylambert 1/n I can 100% no-hype say that this is a false assumption: I had two project running this night, one did finish about 2 hours after I went to bed. The other worked all night, even for a big part through the day, before I had to add decisions, but where the hype does some in:
i took a day off. my PRs kept moving.
129 PRs in a few weeks. 118 fix requests to my agents. 24 nudges to reviewers.
and the loop running it has no AI in it.
part 1 of my agent setup, and how to build yours: https://t.co/4ZiwdagQTM