This is a great review of the language network. This network is distinct from systems of thought. This distinction might be the basis of more human-like and powerful artificial intelligence models!

Agree that LLMs face "a quartet of cognitive challenges: reasoning, planning, persistent memory, and understanding the physical world." Addressing these challenges requires developing separate modules for them and training those modules on data from other sensors (like vision)!

Here's my conversation with Edward Gibson (Ted Gibson, Language Lab MIT), a linguist and psychologist at MIT, heading the MIT Language Lab. We talk all about the human language: syntax, grammar, structure, theories of language, evolution of language, how it reflects culture, and of course LLMs,

Research should adopt the principle shown by this video: when you have a new claim/model, try hard and thoroughly to break it so that (if it's indeed not broken) it's clear your new thing is better!

✨🎓 I defended my dissertation “The Relationship between Linguistic Representations in Biological and Artificial Neural Networks” on Tuesday! 🎓✨ Incredibly grateful for my amazing PhD advisor Ev (like in 'evidence', not Eve) Fedorenko 🇺🇦 and a wonderful journey at MIT Brain and Cognitive Sciences! 🧠🤖

We should be smarter than just scaling! We should create data-efficient algs. Humans are great at this, algs should learn from humans. This is what I have been working on (t.ly/KB818). This is also what our BabyLM is about (babylm.github.io)!

Have laws gotten simpler over time? Analyzing all laws by congress 1951-2022, we (w/ Frank Mollica, Ted Gibson, Language Lab MIT) find laws remain laden with complex structures vs baseline, suggesting efforts to simplify have largely failed Out now American Psychological Association JEP-General! psycnet.apa.org/record/2024-76… 1/

Exciting news! The Murty Lab Georgia Tech, has received its first significant funding from the NEI. We are now actively recruiting up to two postdocs in experimental and computational neuroscience in human vision NeuroAI. 1/n

1/ Our work on unified principles for Topographic Deep Artificial Neural Networks is finally out in Neuron! 7 years in the making. tinyurl.com/2dtkh9fc

I am at the Mexico City attending NAACL 2024 from now to June 22nd. The first paper will be presented at tomorrow's first oral session at Don Alberto 1. If you are interested in the work, please come to the talk! If you want to chat, just DM me! Looking forward to the conference!

The presentation will start very soon! Come to the Don Alberto 1 for the talk!

🎉 Thrilled to share that our paper (the first paper here) has won the Best Paper Award at NAACL 2024! A huge thank you to my mentors, Ev (like in 'evidence', not Eve) Fedorenko 🇺🇦 and Jacob Andreas, and everyone who supported this work. 🙌 Check out the details and our findings here: aclanthology.org/2024.naacl-lon…

Awesome win for open-source and neurosymbolic AI :) Combining CoT planning in free-form text w/ interleaved program generation, execution, and repair, small Numina 7b model wins AIMO progress prize, solving challenging competition-level math problems. huggingface.co/blog/winning-a…

Sad to see you go. Working under your leadership has been an incredible experience. The AI4Science community is lucky to have your energy and expertise. Wishing you every success ahead!

Having compositionality in representations doesn’t mean having compositional models. E2E learning is strong in generalizing progress in one subject into others. Reasoning is the core part behind this generalization and likely not further modularizable.

I mostly agree—but what's more important than being 'general' is that an AGI benefits all of humanity. To truly achieve this, an AGI must naturally interact with humans, requiring robust vision, speech, and motor-control capabilities.

Human-AI interfaces must surpass human conversation. We are not there yet, as we type and use mouse. Speech outruns typing, while gestures like circling outperform mouse dragging. ChatGPT’s AVMs are built towards this goal. The future should real-time all-modality conversations.

Glad to see that my first publication w Daniel Yamins was continued by Aran Nayebi ! AI has so much potential to contribute to other fields of science, particularly neuroscience, considering how much is unknown and the fascinating parallels—and differences—between AI and the brain.

Excited about this update from our team, esp Pingchuan Ma, Honglin Chen, and Damian Mrowca! Sounds more natural and human like, with better translation support. Enjoy the chats while we further improve the model, and let us know what could be even better!