david yan @dzyan01

A striking difference between language and vision is that training on LLM outputs works really well. While pixel outputs of video models are still useful (e.g. as policies with an IDM or for evals), they seem to be difficult to directly train on. So, whither generated pixels?

1 0 6 1K 0

View Details

david yan @dzyan01

a week ago

Beginning to start taking the idea that I should take ideas seriously, seriously. Based on my forecasts, I'll take my first idea seriously in ~ten years.

2 1 10 886 0

View Details

Tom Silver @tomssilver

2 weeks ago

New blog post: tomsilver.github.io/blog/2026/now-… This is the second in a "series" of posts about how to define and work on research problems. (The first post was 3 years ago...) This one addresses the agentic elephant in the room.

0 15 124 6K 90

View Details

david yan @dzyan01

2 weeks ago

@astridwilde1 prior to ~2025 most synthetic stereo depth datasets didn't randomize the camera baseline (and it's really important for generalization, turns out)😆

1 0 3 160 0

View Details

david yan @dzyan01

2 weeks ago

read this cool work from @ShmuelBerman!

NaN Xue @NanXue7

2 weeks ago

anthropic.com/research/claud… Claude plays robotics A must-read!!!

3 63 393 93K 541

0 0 1 441 0

View Details

david yan @dzyan01

2 weeks ago

speaking of perceptual similarity, somebody should totally train a RGB depth map LPIPS model

0 0 0 68 0

View Details

david yan @dzyan01

2 weeks ago

Every vision task eventually hits a critical threshold where qualitative visualizations get good enough and all the new papers start spamming them. Unfortunately, soon after this threshold, human perceptual similarity becomes saturated.

Massimiliano Viola @massiviola01

3 weeks ago

Image foundation models never stop surprising!😮 And this time, it's not the usual big tech players. A few days ago, @robbyant_brain released LingBot-Vision, a new family of vision encoders built natively for dense spatial perception.

2 43 302 35K 222

2 0 4 778 0

View Details

david yan @dzyan01

2 weeks ago

@SALTWATERSOPH just gotta read principles a couple more times

1 0 0 43 0

View Details

david yan @dzyan01

3 weeks ago

@mean_field_zane @alz_zyd_ Agreed, I've looked at Princeton's historical first-destination data before and I don't believe there's a significant downward trend (even for CS).

1 0 1 139 0

View Details

david yan @dzyan01

3 weeks ago

@DimitrisPapail

0 0 0 33 0

View Details

david yan @dzyan01

3 weeks ago

@andersonbcdefg @stevenxiao_ Moloch, whose love is Lincoln-Douglas debate circuits!

0 0 0 23 0

View Details

david yan @dzyan01

3 weeks ago

@flawedaxioms @AaronBergman18 @slatestarcodex For a "posthuman" society, the diff wouldn't mean much. Bostrom's original napkin math was that a planet-sized computer might have around 10^41 FLOPS, and that simulating all minds in all of human history is likely around 10^36 FLOPS.