Tarun Khajuria

We underestimate how much "abstract" thought is just repurposed sensorimotor control circuitry. A lot of reasoning is essentially about moving through idea-space the way we move through physical space.

🚨Transformers don't learn Newton's laws? They learn Kepler's laws! Like us, transformers don't predict a flying ball via a differential equation, but by fitting a curve. Moreover, reducing context length steers a transformer from Keplerian to Newtonian. Compression in play.

1/ I've been thinking about the "non-linear features" objection to interpretability work. But it can be hard to reason about, since the space of possible non-linear features is so large. Here's an attempt to untangle it: livgorton.com/non-linear-fea…

@gialdegheri @jaaanaru Thank you! The thorough reviews were a great help in improving the paper.

@jaaanaru @tarunkhajuria Congrats, I really enjoyed this paper! (full disclosure: I was one of the annoying reviewers 😅)

Under challenging conditions, human vision turns into iterative problem solving. Typical deep learning algorithms don't capture this behavior. Our paper out now in Plos Computational Biology Led by @tarunkhajuria journals.plos.org/ploscompbiol/a…

@JaipurDialogues Yes!

Justice Gavai mocked Bhagwan Vishnu : No Action Ajeet Bharti mocked Justice Gavai : Picked up for Questioning by Noida Police Is Contempt of Judges Higher Than Contempt of God?

I just pulled the numbers on vision-language benchmarks for Llama-3.2-11B (vision). Surprisingly, the open-source community at large isn't behind in the lightweight model class! Pixtral, Qwen2-VL, Molmo, and InternVL2 all stand strong. OSS AI models have never been stronger. The last 3 lines are API-only frontier models. Gemini-flash and GPT-4o (likely in heavier-weight class) are still the reigning champions. But never bet against OSS. Never underestimate the combined firepower of so many talents distributed all over the world.

What aspects of human knowledge are vision models missing, and can we align them with human knowledge to improve their performance and robustness on cognitive and ML tasks? Excited to share this new work led by @lukas_mut! 1/10

Super excited to finally share what I have been working on at OpenAI! o1 is a model that thinks before giving the final answer. In my own words, here are the biggest updates to the field of AI (see the blog post for more details): 1. Don’t do chain of thought purely via prompting, train models to do better chain of thought using RL. 2. In the history of deep learning we have always tried to scale training compute, but chain of thought is a form of adaptive compute that can also be scaled at inference time. 3. Results on AIME and GPQA are really strong, but that doesn’t necessarily translate to something that a user can feel. Even as someone working in science, it’s not easy to find the slice of prompts where GPT-4o fails, o1 does well, and I can grade the answer. But when you do find such prompts, o1 feels totally magical. We all need to find harder prompts. 4. AI models chain of thought using human language is great in so many ways. The model does a lot of human-like things, like breaking down tricky steps into simpler ones, recognizing and correcting mistakes, and trying different approaches. Would highly encourage everyone to look at the chain of thought examples in the blog post. The game has been totally redefined.

Could mimicking psychedelic effects in virtual reality bring along therapeutic effects associated with psychedelics?🤔 In our new work, we developed a psychedelic VR experience - Psyrreal - and showed its potential for reducing depressive symptoms. 1/n frontiersin.org/articles/10.33…

In science we value citations but it is quite refreshing when someone does a video about your work instead. 🙏@tipado VR experience is not the same as psychedelics but it's worth studying the potential of virtual psychedelics. psyarxiv.com/uh9kf youtu.be/Aqxk9SQ41Js?t=…

Did you know that MIT offers an entire course on Deep Learning for art and creativity? Slides and schedule: ali-design.github.io/deepcreativity/ Youtube: youtube.com/watch?v=MABLFo…

Academic Hell

“It is Okay to Not Be Okay (ArtEmis 2.0)”@ CVPR22, congrats Youssef! artemisdataset-v2.org #cvpr2022

Is it possible that adversarially-trained DNNs are already more robust than the biological neural networks of primate visual cortex? Here is a short thread for our #ICML2022 paper arxiv.org/pdf/2206.11228…. 1/8

Can you guess what's hidden in these images? How did you arrive at the solution? Can we build AI algorithms that solve it in a similar way? We developed a new dataset to investigate the iterative generation and refinement of perceptual hypotheses.🧵 openaccess.thecvf.com/content/CVPR20…

Please share 🙏 A fully funded Ph.D. position on "Injecting background knowledge for Explainable AI" is available at the University of Tartu. The project revolves around extracting knowledge from text to help interpreting AI models for biomedicine. reaalteadused.ut.ee/en/content/doc…

@timesofindia Don't we see a problem in this arrangement?

India avoids condemning #Putin to get weapons for China fight Read: toi.in/spTrKZ/a24gk #RussianUkrainianWar

@prxgyx @Vasundhara_19 @pxthxk +1 :))

@Vasundhara_19 @pxthxk +1 :)

🥂