Dmytro Hrybov

ml research engineer | prev google, straumann, tcl
computer vision, evals, world models

[Dmytro Hrybov @dimentary]

@26pablo7 that one is maybe less tractable...

[Dmytro Hrybov @dimentary]

makes me thinking that giving LLMs sense of visual beauty is more tractable than it seems

[Surya @kickingkeys]

You can just RL a coding model to paint with javascript btw

[Dmytro Hrybov @dimentary]

great article, congrats!

interesting that their physical prompting can to be done both with egocentric and teleop approaches. curious how they compare in terms of resulting quality @SkildAI

new kind of data engineering emerging.

[Skild AI @SkildAI]

Introducing S1, our new foundation model that learns from one example.

It can be taught 10-minute long tasks that it has never seen before, from one video prompt without any fine-tuning.

Watch S1 operate in real-time via in-context learning:

[Dmytro Hrybov @dimentary]

@26pablo7 @thekithe amazing!

[Dmytro Hrybov @dimentary]

damn

[Eren Chen @ErenChenAI]

1.3 meter long Dexterous hand.

And it’s for sale.

[Dmytro Hrybov @dimentary]

@datapointai @_akhaliq this is big for the image gen community, congrats!

[Dmytro Hrybov @dimentary]

@Hongyu_Lii @bowenwen_me @zhu_xinghao @du_yilun @YunzhuLiYZ @BirchfieldStan @SohaPouya @80gg_overmind @Dr_YanChang amazing! which use-cases had the biggest gap between predictions and ground truth from your experience?

[Dmytro Hrybov @dimentary]

@chris_j_paxton came there today from that hill too :)

[Dmytro Hrybov @dimentary]

@jabstec @onblueroses amazing read

[Dmytro Hrybov @dimentary]

@irvinxyz @dee_hw @livinoffwater @skylark_mmm @Karim_RC @SarkaryShahvir finally real deployments

[Dmytro Hrybov @dimentary]

@acheong08 cracked

[Dmytro Hrybov @dimentary]

@meowbooksj maybe anthropic is entering robotics game

[Dmytro Hrybov @dimentary]

@arena @izziebuilds they made RL for RL

[Dmytro Hrybov @dimentary]

@tunguz looks like a new ai startup logo

[Dmytro Hrybov @dimentary]

@philhie @ScottWu46 damn

[Dmytro Hrybov @dimentary]

@philhie goated

[Dmytro Hrybov @dimentary]

@fh7ww @benchmark we need evening basketball sessions

[Dmytro Hrybov @dimentary]

@izziebuilds let’s build robot school

[Dmytro Hrybov @dimentary]

@assylunicorn show them what you got

[Dmytro Hrybov @dimentary]

egocentric video is turning into robotics' pretraining corpus - what web text was for llms. dyna-2's own ablations already treat action-labeled hours as the scarce axis: 50k, against 1M hours of video.

[Dmytro Hrybov @dimentary]

most of the work worth capturing for robotics (factories, labs, kitchens) happens in gloves. so i put some gloves on egocentric training data using gemini omni video to see if hand extraction survives it.

- nitrile gloves: tracking looks perfect, but the extracted wrist rotation (the action a robot would learn to copy) is 3.5x more wrong.
- black work gloves: the hand simply isn't found in roughly 1/5 of frames.
- leather work gloves: tracking misses half the time.

Dyna-2 training data pipeline is based on a similar extraction, so the quality of the robotics action models depends on it.

[Dmytro Hrybov @dimentary]

how: egodex clips with various tasks where wilor (the detector extraction stacks build on) tracks ~100% of frames. a video edit adds the gloves; nothing else changes - same hands, same task, same camera. regenerating clips unchanged costs 0.1% coverage and 3.3 degrees of wrist rotation, so it's the gloves, not the edit.

n is small (10 videos, one detector) - so that's an early signal and needs replication at scale.