Dileep George @dileeplearning

I'll be at Cosyne workshops next week (16th and 17th). We are hiring at Astera NeuroAI. If you are at Cosyne workshops next week and want to learn more about this exciting research team we are building at Astera, please DM me or email me at dileep@astera.org

🧵1/12 What if the DMN, limbic system, hippocampus, neural oscillations, gradients, dementia syndromes, mixed pathology, and aphantasia all fall out of the same generative brain computation? 🤯#endalz Introducing #SLOD (preprint): a new #NeuroAI framework w/ @drbreaky.bsky.social

6/12 The clone-structured cognitive graph (CSCG) framework pioneered by @dileeplearning.bsky.social offers a natural computational mapping for how latent diffusion could be implemented biologically along hippocampal circuitry.

My latest on Substack -- a write-up of the talk I gave at NeurIPS in December. aiguide.substack.com/p/on-evaluat...

4/4 “Language is a mechanism to control other people’s mental simulations.” Dileep George, @dileeplearning, of @GoogleDeepMind at the Simons Institute workshop on The Future of Language Models and Transformers. Video: simons.berkeley.edu/talks/dileep...

1/4 Do LLMs understand? "They understand in a way that’s very different from how humans understand," Dileep George, @dileeplearning.bsky.social, of Google DeepMind at the Simons Institute workshop on The Future of Language Models and Transformers. Video: simons.berkeley.edu/talks/dileep...

the one where #AGIComics figures out what academic training is all about … www.agicomics.net/c/artificial...

Check out my blog on the risks of *not* building AGI! blog.dileeplearning.com/p/if-no-one-...

Can AIs be conscious? Should we consider them as persons? Here are my current thoughts..... blog.dileeplearning.com/p/ai-conscio...

blog.dileeplearning.com/p/quick-note... TLDR: It was fun and the process felt 'magical' at times. If you have lots of small project ideas you want to prototype, vibe-coding is a fun way to do that as long as you are willing to settle for 'good enough'.

Those who think there's an AI bubble are unaware of a recent breakthrough.... www.agicomics.net/c/ag-breakth...

New and improved and 10000% vibe-coded! Check out www.agicomics.net

1/4) I’m excited to announce that I have joined the Paradigms of Intelligence team at Google (github.com/paradigms-of...)! Our team, led by @blaiseaguera.bsky.social, is bringing forward the next stage of AI by pushing on some of the assumptions that underpin current ML. #MLSky #AI #neuroscience

1/ 🚨 New preprint! 🚨 Excited and proud (& a little nervous 😅) to share our latest work on the importance of #theta-timescale spiking during #locomotion in #learning. If you care about how organisms learn, buckle up. 🧵👇 📄 www.biorxiv.org/content/10.1... 💻 code + data 🔗 below 🤩 #neuroskyence

#AGIComics now has a website! And it is 100% vibe coded! Check out agicomics.net

12 leading neuroscientists tackle a big question: Will we ever understand the brain? Their reflections span philosophy, complexity, and the limits of scientific explanation. www.sainsburywellcome.org/web/blog/wil... Illustration by @gilcosta.bsky.social & @joanagcc.bsky.social

🎯

𝗛𝗼𝘄 𝘀𝗵𝗼𝘂𝗹𝗱 𝘄𝗲 𝗱𝗲𝗳𝗶𝗻𝗲 𝗮𝗻𝗱 𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗲 𝗮 𝗯𝗿𝗮𝗶𝗻 𝗿𝗲𝗴𝗶𝗼𝗻'𝘀 𝗶𝗺𝗽𝗼𝗿𝘁𝗮𝗻𝗰𝗲? We introduce the idea of "importance" in terms of the extent to which a region's signals steer/contribute to brain dynamics as a function of brain state. Work by @codejoydo.bsky.social elifesciences.org/reviewed-pre...

It's kinda obvious. #AGIComics has already figured out which brain region is the most important. 😇

ohh...yes...this is exactly what I think after reading some of the "deep research" reports. ....written by a committee

jumping on the Gemini 2.5 bandwagon... it's an incredible model. really feels like an(other) inflection point. talking to Claude 3.7 feels like talking to a competent colleague who knows about everything, but makes mistakes. Gemini 2.5 feels like talking to a world-class expert with A+ intuitions

Give me 10 billion dollars and I’ll do it. 1 billion for developing the hardware and 9 billion to pay for my opportunity cost 😇

Nope. It is an engineering problem. Give me an algorithm you think is not being scaled because of a hardware mismatch, I can make the hardware (chip + interconnect + datacenter) given enough money. Purely an engineering problem.

ok...in that case which other existing model would you make a bet on scaling up? Pick one. I'd be happy to raise money for it.

This is great....this is a hypothesis. What other such hypotheses are there? Is there a space of such hypotheses, or is this just one? (Also, how does this match with Konrad's @kordinglab.bsky.social argument about being hypothesis-free?)

But why would you stop there? Why aren't you convinced this is how it works? It's SOTA on all the tasks.... You are not convinced because it doesn't match some brain data?

Here's my Q to @kordinglab.bsky.social who argues brain is too complex to understand.... How do you know that current systems AI are NOT how the brain works?

Wait... so you think brain is NOT a large transformer? How do you know that? The current empirical studies strongly support the idea that brain is a large transformer.....

TLDR: We need large amounts of data to finally show that brain is a large transformer that we cannot understand. 😇😝

The bitter lesson is that neuroscience seems to be out of ideas... 😇

Although Konrad is being brief here, the idea that “ANN are now often part of normative models and they are definitely SOTA” is something I see often. Including a component that is sota elsewhere doesn’t make a model any better. It is very possible to build sucky models with ANN.

If it doesn’t work, it’s called a normative model. Otherwise it’s a just a model. 😇

This research is "nonsense". The reviewers and editors that let this pass should be "ashamed".

This remains important today. As industry places huge financial bets on large language models, the research community needs to investigate other approaches that may achieve similar or better results via entirely different methods. end/