sahil jagtap

@twtofsahil

chief virtualization officer @drumslabs | @fdotinc | @agentbudgetsdk | built @airstitch (acquired) | i love both the OSes (operating systems & open-source)

Mumbai/SF

Pinned
sahil jagtap
@twtofsahil
Oct 30
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,
2
sahil jagtap
@twtofsahil
Nov 14
@paperinstr's Ultramarine 1 Flash. they took an existing model, GLM-5.3 Flash, and report improving it with RL across six professional-work benchmarks, by 4.1 points on average. They used LoRA, not full-weight training That makes adapter-based RL worth considering seriously, rather than assuming we need full-weight training from day one
1
sahil jagtap
@twtofsahil
Nov 14
burnt all of $10k AWS credits that I got from @ycombinator startup school :) asked for $15k more let’s see ran a lot of RL experiments, also some goated benchmarks (a big launch) coming soon!
8
sahil jagtap
@twtofsahil
Nov 3
i've used about $3.9k in aws credits running computer-use RL experiments. thank you yc university program. these models learn to operate a computer by looking at the screen, clicking, and typing. but to train them, you also need a reliable way to tell whether they actually
1
sahil jagtap
@twtofsahil
Nov 2
happy halloween 🎃

sahil jagtap’s posts