My Little AI Factory | Hacker News Reader

My Little AI Factory

dominis.blog

14 pointsdominis10 comments

All in all, this feels well thought out. Glad to see something beyond another singleprompt, vibecoded house of cards. I buy the software factory idea. The pain is real... you and your team get the harness working, the models change, and suddenly you’re back fixing things that worked last week. Feels like we’re working harder than ever, but spending 70% of the time just trying to get back to last week’s efficiency. Meanwhile, we’re shipping fixes to our own tooling instead of shipping value to the customers paying our salaries. Hard to scale that. Anyway, /shouting@clouds=0. I’d love something that makes us less dependent on whatever Dario, Sam, and Sundar(just kidding ;)) feel like tuning that morning. Anyone actually running something like this in production? I’ve been trying Dex's Humanlayer feels like what is 80% done (like all most of my AI projects...), promising, but not prod ready on my end. Also I'm on the Warp Factory waitlist...

This is what I mostly miss in all the AI-related stuff churned out. Any tool will only be ever useful as well we understand it. Like, how many hours must have been put into this project, insane! And the results look tangible. I hope other project authors take note b/c this seems to me to only way to diverge from the path of overwhelming information that Stanislaw Lem has warned about.

What’s your experience with Lex? Does it actually fulfil its promise? My experience is that Claude is quite handy at circumventing most guardrails, you give it a task and it executes, if that means storing an API key in the session or extracting it from a vault it will do so. How can Lex handle this vs. others in the market?

on individual level for basic things I found hooks pretty good, you can block tool executions easily.

I think here the question will be how to scale rules on team level.

Obviously sandboxing will be needed if we want proper gates.

"I’ve also learned that many people don’t use English to interact with agents"

WTF... that's weird. English is not my first language but I still ask every question in English. Does the LLM write code comments in other languages then?

Can anyone actually attribute revenue to token spend, or is '$1 in, $100 out' the new 'digital transformation'?

I wouldn’t be surprised if in most cases it’s $100 in $1 out.

not yet, but you know its probably the most exciting aspect of Ai to the bean counters

Very interesting article!

Great journey and nice story!

Reply on news.ycombinator.com