tom white

the models, they just wanna converge

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

@sebkrier one-shotted just enough to win

This is wild. theaustralian.com.au/business/techn…

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)...

"pirate ship" (ImageNet class 724)

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.

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

not seeing it? don't worry - your favorite imagenet model is.

traffic sign, baseball player, pomegranate

@johnhewtt x.com/dribnet/status…

Lots of interp thought discusses the linearity of the residual stream! This blog post: the residual stream isn't linear in a way that provides formal leverage, and interp methods based on linearity should not be preferred beyond empirical utility. cs.columbia.edu/~johnhew/resid…

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.

piggy bank (ImageNet class 719)

@DlSPUTED thanks! its a chonky 18in screenprint

@dribnet love the baseball player

@hyhieu226 enjoying this slow takeoff and will genuinely miss it

Today, I finally feel the existential threat that AI is posing. When AI becomes overly good and disrupts everything, what will be left for humans to do? And it's when, not if.

@farmgeek fwiw: the email was garbage but the disclosure within the app is actually pretty good - they show at the document/file level what was accessed

“All patients who are not impacted can see that in their MMH app” That’s great, except when you can’t log in and every method (password reset, one-time pass) fails. Shitshow.

@RT_Artwork Got this too. 5 minutes before call they ask you to use the riverside client and forward you to a website clone (riverside dot name - BEWARE) with their malware installer (I stopped there). (would be up for a themed exhibit on scams showcasing all artists with this invite! 😂)

Be careful with artwork commission requests. Starts out with request for a video call. I contacted the comnpany directly and found out this request was a scam. Getting on a call could have led to some software being downloaded and wallet drained. Anyone have experience with this?

@CSProfKGD looks different here in nz 😉 (happy new year!)

@NeelNanda5 Understandable - though the distribution shift implies not all Gemma 3 concepts will be represented in these SAEs. Were there any others such as language filtering? Might be worth updating the technical paper which seems to be misleading on this point.

@dribnet Text only, sorry! Images are annoying infra wise

I'm excited to release Gemma Scope 2: a comprehensive set of interpretability tools on Gemma 3. SAEs & transcoders on every layer of every model! Gemma 3 27B shows lots of rich safety-relevant behaviour I want to enable deep dives into what's really going on Check out our demo!

@GaryMarcus @Ted_Underwood always felt you were wrong about this, but ChatGPT in fact completely agrees with you and helped me to better understand and appreciate your point here 👍

@dribnet @Ted_Underwood one of key examples in 2001 was object permanence; we gave examples of an object permanent failure today, two decades later, even though GPT has a trillion times more data and compute. grammaticality has improved, via pastische, but conceptual understanding has not.