University of Michigan has an open assistant professor position in Complex Systems. Amazing opportunity to do join an interdisciplinary community of faculty and students. Looking for discipline thinkers who transcend disciplines.

🔔 New paper out in @PNASNews 🔔 “Large Language Models based on historical text could offer informative tools for behavioral science” W/ Michael Varnum, Nicolas Baumard, & @kurtjgray

Large study shows humans can learn from AI feedback but access to AI also amplifies existing inequalities by increasing the skill gap and reduces intellectual diversity: everyone learns to specialize in the same areas https://t.co/igSIj33n8r

Haha we've all been there. I stumbled by this tweet earlier today and tried to write a little utility that auto-generates git commit message based on the git diff of staged changes. Gist: https://t.co/CNqxyUXcqM So just typing `gcm` (short for git commit -m) auto-generates a

Our project for experimentally testing planetary regulation (Lovelock & Margulis) on the test tube, using a synthetic microbial Gaian system (+1 PhD), has been funded by @AgEInves. A great adventure that started at @sfiscience 2 yrs ago with @VictorVmaull & @jordiplam

Super excited to share our new work (and the first of my PhD) : "Evolving Self-Assembling Neural Networks: From Spontaneous Activity to Experience-Dependent Learning" We propose Lifelong Neural Developmental Programs for continually self-organizing artificial neural networks !

Deadline shifted to June 15!

I've gotten quite a few cold emails from students who found this long review I wrote useful and I'd like to do...something with it. It's way too long to submit as a stand-alone paper to, say, Entrop tho. Does anyone publish big tutorial reviews like this? https://t.co/9TKcNQ3Qmm

Now out in Trends in Cognitive Sciences!

Anyone else find themselves estimating the "GPT grade" of things you hear/read? When something is poorly written or generic, it's "GPT-2 grade" content. When something is lit, you can complement it as being "GPT-7 grade" etc. This reminds me of a fun side project I had saved for

Honesty and helpfulness are two central goals of LLMs. But what happens when they are in conflict with one another? 😳 We investigate trade-offs LLMs make, which values they prioritize, and how RLHF and Chain-of-Thought influence these trade-offs: https://t.co/wXJ9hFs3Vc [1/3]

In our new paper in @Ecology_Letters we apply modern coexistence theory to higher order species interactions where we compute niche and fitness differences as changin over time Thanks to Agnieszka Majer, Anna Skoracha (not on twitter) and @L_Kuczynski https://t.co/kAs6ygg29b 1/7

the divide on perceived ideology of Target but not Walmart is pretty striking

How can we model the collective behavior of complex systems with many dimensions? Is it possible 2 find a model reduction that captures the key components? Check this great paper in @PhysRevLett & how to use it in many different contexts (networks+Ising) https://t.co/hhUasBIuXu

And just like that... :)

It’s amazing how many empirical features of interactions and collective response can emerge from a simple imperative: to minimize surprise. Out today in PNAS @PNASNews - with @conorheins @richardpmann Karl Friston and colleagues! @CBehav @maxplanckpress https://t.co/gZcyDuG285

Highly amusing update, ~18 hours later: llm.c is now down to 26.2ms/iteration, exactly matching PyTorch (tf32 forward pass). We discovered a bug where we incorrectly called cuBLAS in fp32 mathmode 🤦‍♂️. And ademeure contributed a more optimized softmax kernel for very long rows

THE REVENGE OF PYTORCH just kidding :) @cHHillee (from PyTorch team) was kindly able to help improve the PyTorch baseline, done by 1) upgrading to nightly, 2) using the "compound" F.sdpa (scaled dot product attention) layer directly, and turning on a torch compile flag:

Have you ever wanted to train LLMs in pure C without 245MB of PyTorch and 107MB of cPython? No? Well now you can! With llm.c: https://t.co/w2wkY0Ho5m To start, implements GPT-2 training on CPU/fp32 in only ~1,000 lines of clean code. It compiles and runs instantly, and exactly

Interesting paper argues that scale is all you need to explain humanity’s unique levels of intelligence. “We propose that global, genetic differences in learning and memory are sufficient to account for uniquely human reasoning across domains,” https://t.co/Kiq88lKLDb