tom white
@dribnet

creations with code and networks

 Wellington, New Zealand
drib.net
Joined June 2011
1,511 Photos and videos
enjoying the new global workspace anthropic paper - especially using the Jacobian lens to provide readout interpretations of prints shown at last month CVPR
3
1
28
2,053
more replies
here's the same effect with a different print: on the 48 layer gemma3-12b, at layer 47 you get bird, woodpecker, ornothology but upstream at layer 37 its again just "bird birds birds"
1
2
285
just some early results - slightly cherry-picked, but things seem to improve with scale which is a good sign. kudos to @wesg52 @sofroniewn @Jack_W_Lindsey and team for the refrence implmentation which made it easy to explore!
Anthropic
@AnthropicAI
Jul 6
New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.
1
2
303
pleased to be showing new screenprints at @CVPRConf this weekend. these use mechanistic interpretability techniques to highlight features shared across a wide family of vision models
Shelfwild

ALT Shelfwild

Vibe of the Tiger

ALT Vibe of the Tiger

1
3
20
1,756
more replies
Each print's unique backlighting element is provided by a TIPS model. This shows spatial patch level activations highlighting visually where that model attends to when activating its analogous feature.
Shelfwild activation heatmap

ALT Shelfwild activation heatmap

Vibe of the Tiger activation heatmap

ALT Vibe of the Tiger activation heatmap

1
4
511
If you are at CVPR drop by - six pieces are on display today and tomorrow just behind the Google booth in the #CVPR2026 art show. I'll be there for daily artist talks on the official schedule or DM if you'd like to meet up and chat.
1
2
14
1,599
saxophone print + top-150 SigLIP image probe though mknn model agreement (a la platonic representation hypothesis) is not part of the test time compute process, it climbs naturally as the as the optimization evolves
2
13
841
the models, they just wanna converge
8
483
unsure how AI interprets this print? treating the image as a linear probe on your favorite vision model and scraping a diverse dataset for maximum activations provides a coherent suggestion.
1
2
11
1,567
"pirate ship" (ImageNet class 724)
1
5
476
45% mutual kNN between CLIP and SigLIP — not bad for two model families trained on different data with different objectives when probed with this print. revisiting ImageNet so I can build a toolbox for navigating more uncharted waters without class labels (stay tuned)...
7
311
traffic sign, baseball player, pomegranate
2
1
16
929
not seeing it? don't worry - your favorite imagenet model is.
1
7
692
or query your favorite vision model for semantic nearest neighbors - here's OpenAI-CLIP's top hits across CC3M using the baseball_player print as a probe
2
237
piggy bank (ImageNet class 719)
1
2
379
Weapons-grade piggy bankness: One drawing. No training. Subtract the style, get a direction in SigLIP space. Sort 50K ImageNet images by cosine similarity: 41 of the top 50 are piggy banks (P@50 = 82%). The drawing is the classifier.
1
1,305
see the skull?
2
8
848
a hard "yes" from Gemini 3 ✔️
4
647
Load more