🗜️Compress is (or at least demonstrates) intelligence🧠. Starting with a picture of me and prompting Gemini with "Create an exact replica of this image, don't change a thing." yields this quite bizarre progression!. Does this seem like 🧠 or 😂 to you? This was a fun experiment!

Interpreting the video 2/2:.- Middle row shows foveated information, gates (also new!), and predictions, plus attention (both in model space and reprojected to euclidean space. - Bottom row shows the neural dynamics of randomly chosen neurons.

Interpreting the video 1/2:.- The original image is top left. It has an overlayed polar grid that denotes foveated sampling. - The middle is what the CTM sees (i.e., that which goes into the backbone).- The right is the foveated perspective reprojected into euclidian space.

CNN filters:.CNNs are typically large models that have to handle a very large number of resolutions where objects can appear close or distant. Enabling the CTM to effectively "choose its own resolution" might let us use far smaller and efficient backbones.

Inspired by nature:.Foveated vision is ubiquitous in nature. The CTM gets a small "fovea" (focus) and a compressed, lower-res view of the "periphery". This allows it to dynamically choose between examining details and understanding context.

Emergent and simple:.This is all done with no extra auxiliary losses and without RL. The objective is simply supervised cross entropy loss.

Action and feedback: .The CTM chooses where to look and how "zoomed in" its effectively lens is. Internal and ongoing feedback enables some fascinating behaviour.

Learning where to look, learning what to see. Work on Continuous Thought Machines continues to be exciting! As a teaser, take a look at our approach to "Foveated Active Vision". Check it out 👇. (details in the thread).

RT @SakanaAILabs: When AI Is Designed Like A Biological Brain 🧠. Full Video:

RT @konstmish: I'm surprised this post didn't get more attention. Most papers arguing we are far from AGI just show LLMs failing on tasks t….

RT @nhk_kanazawa: 北國銀行がＡＩ活用しサービスを向上 開発会社と基本合意 #石川NEWSWEB

If you’re interested in learning about Continuous Thought Machines (, we made interactive notebook tutorials so you can hack around with CTMs. ImageNet:. MNIST Tutorial:. Let me know if you have any feedback!

RT @SakanaAILabs: Introducing The Darwin Gödel Machine: AI that improves itself by rewriting its own code. The Dar….

RT @SakanaAILabs: Announcing a Multiyear Partnership between Sakana AI and MUFG Bank. We’re excited to announce th….

Open science is truly wonderful. After releasing out work on Continuous Thought Machines (, the community discussions and suggestions have been so helpful, forthright, and gratifying.

RT @risi1979: Super excited about the CTM and feeling lucky to have been part of the amazing team w/ @LearningLukeD @ciaran_regan_ @jeffrey….

RT @hardmaru: New Paper: Continuous Thought Machines 🧠. Neurons in brains use timing and synchronization in the way that they compute, but….

RT @YesThisIsLion: Why not try something different? Instead of using the activations in a layer of a neural network as the representation,….

RT @SakanaAILabs: Introducing Continuous Thought Machines. New Blog: Modern AI is powerful, but it’s still distinc….