a computer-use agent needs a computer to practice on and be better.
somewhere it can open a spreadsheet, change a setting, save a file and make mistakes without touching your actual work.
for reinforcement learning, the loop is: give it a task, let it try, check the result, then use that feedback to update the model. repeat across many attempts.
running those attempts in parallel means providing lots of separate environments. each needs the right starting state, and each result needs checking. clicking "save" isn't proof the file saved correctly.
so yes, fitting more desktops on one machine is useful. it can make practice cheaper.
but a thousand desktops isn't proof of cheaper learning.
start with the same model, tasks and rewards. compare against a properly tuned alternative. how much money and time does each setup need before the agent reliably completes new tasks it wasn't trained on?
include failed attempts, resets, checking and model compute in the bill.
if you're selling cheaper training, show what it costs to train the agent. a desktop count isn't enough.
if i were a YC founder and my posts weren’t getting enough engagement, i’d simply tweet:
“YC W27 deadline is coming up. comment what you’re building below. i’ll personally review your app, personally refer the top ones, and personally make sure this tweet does numbers.”
anyway i made a mango-pistachio affogato in an @Apple cup today.
btw there’s absolutely nothing wrong with helping people hahahaha ifykyk
lowkey cool seeing myself among the contributors of something I use daily
had the chance to chat with @FlyaKiet , one of the goated founders I’ve come across. insane execution from the entire @ParadevAI team.
absolute killers. massive W. well deserved 🫡
I think this is the best time in history to start a company alone.
Not because startups got easy.
Because the cost of finding out if an idea is worth building has collapsed.
A thread:
your coding agent gets dumber the longer the session runs.
not the model. the session.
sharp at minute 5. by hour 2 it’s re-reading files it already read and re-suggesting changes you already rejected.
the context filled with junk and buried the signal. everyone feels this. nobody measures it.