Jayoo Hwang

Independent researcher @ml_collective | Agents, Computer use

[Jayoo Hwang @JayooHwang]

https://t.co/E6lpFxlJiF

[Jayoo Hwang @JayooHwang]

@JFPuget @GalKaplun Timely stuff, thanks for sharing.

[Jayoo Hwang @JayooHwang]

@gabriel1 If 5.6 sol/opus 5 level computer-use became fast and cheap...

[Jayoo Hwang @JayooHwang]

@giffmana I wonder how far it could get in NetHack.

[Jayoo Hwang @JayooHwang]

@a1zhang @lateinteraction Cool work! Really like these example traces from the blog.

[Jayoo Hwang @JayooHwang]

@dillon_mulroy @uwukko Currently working on an ultra token-efficient browser tool. Basically it splits the page into sections and summarizes with a cheap model so that the main agent can choose specific page regions to look at.

https://t.co/xIjMiTPj1G

[Jayoo Hwang @JayooHwang]

@sethkarten I feel like I've gotten 10x better at skim reading since I started working on agents

[Jayoo Hwang @JayooHwang]

@lefthanddraft Might get solved by models optimized for video. I'd be interested to see an update from these guys
https://t.co/jvEgqUFjL4

[Standard Intelligence @si_pbc]

Computer use models shouldn't learn from screenshots.

We built a new foundation model that learns from video like humans do. FDM-1 can construct a gear in Blender, find software bugs, and even drive a real car through San Francisco using arrow keys.

[Jayoo Hwang @JayooHwang]

@tanishqkumar07 Guess it’s hard to beat the density of programming/math

[Jayoo Hwang @JayooHwang]

@_xjdr Just casually serving up 300B frontier-level tokens on the local setup.

[Jayoo Hwang @JayooHwang]

@_xjdr Thanks again. Those were some pretty nice tokens.

[Jayoo Hwang @JayooHwang]

@dwarkesh_sp This has been bugging me for a while too.

I've written some more about why computer use is so clunky and explored a few solutions recently: https://t.co/Gdttd3OGdi

[Jayoo Hwang @JayooHwang]

https://t.co/E6lpFxlJiF

[Jayoo Hwang @JayooHwang]

@j_golebiowski Cool work. Are there any patterns regarding queries where the small model tends to be overconfident?

[Jayoo Hwang @JayooHwang]

@tli104 Nice work! Seems sensible to have this by default. I find agents perform much better in my use when I do manual compaction at good moments.

Another thing I do quite often is rewind + summarize from here. Would be interesting to also allow the LLM to compact the past N turns.

[Jayoo Hwang @JayooHwang]

@Dorialexander So all you need is a bunch of cracked Tsinghua researchers designing agentic workflows for creating RL envs?

[Jayoo Hwang @JayooHwang]

@gabriel1 League skill ceiling is actually insane. Top teams train like 14h, 7days a week.

[Jayoo Hwang @JayooHwang]

@fa1zvn Cool project. I think you might also like our recent paper where we create a harness that makes browser control much easier for smaller models.
https://t.co/Ih5O07RcOA

[Jayoo Hwang @JayooHwang]

@_xjdr Doing the Lord's work

[Jayoo Hwang @JayooHwang]

@jasondeanlee Agreed, feels like current models should be overkill for this but they still get tripped up sometimes. I'm more optimistic about smaller models with specialized harnesses handling these kinds of tasks https://t.co/Ih5O07RcOA