LLMs represent words as vector embeddings. Do they represent *functions* as vectors too? Yes! This has implications for how we think about “reasoning” in language models. New preprint w/ Millicent Li, Arnab Sen Sharma, Aaron Mueller, byron wallace, David Bau: functions.baulab.info

What do you call the disparity between GPU-rich and GPU-poor? Jensen's inequality

Here is a powerful Claude 3 prompt for engineers. Use it to automatically refactor, comment, and improve your code: --- <prompt_explanation> You are a skilled software engineer with deep expertise in code refactoring and optimization across multiple programming languages. Your

good question, I think authoritarian regime takeover [using AI] might be at least slightly neglected in current AI x-safety work

By end of 2024, steering foundation models in latent/activation space will outperform steering in token space ("prompt eng") in several large production deployments. I felt skeptical about this in summer '23, felt vaguely positive in Jan, and now think it's more likely than not,

🚨New preprint🚨 Can LLM finetuning relationships be inferred only through model outputs ? We found that adapting phylogenetic algorithms 🧬 to language models 🤖 helps identify families of models and can even predict their performances with Stefano Palminteri (@stepalminteri.bsky.social) and Pierre-Yves Oudeyer ! 1/9

the inevitable Minskification of neural networks

8x NVIDIA H100 80GB

Instead of analyzing whether AI takeoff will be “fast” or “slow”, I now prefer to think about the spectrum from concentrated takeoff (within one organization in one country) to distributed takeoff (involving many organizations and countries).

My basic reaction to AI today is, “jeez, o1 performs in the top 1% of humans at math, yet fails routinely at basic logic tasks. I guess intelligence is a high-dimensional space, and that probably means, like most high-dimensional things, it behaves counterintuitively.”

Strange but true - A wobbly table on any reasonable floor can be made steady by just turning it. Moral of the story: Before dining out, always ask if their floor is Lipschitz continuous.

'The purpose of abstraction is not to be vague, but to create a new semantic level in which one can be absolutely precise' - Edsger Dijkstra

Tweet the training data you want to see in the world.

New Anthropic research: Alignment faking in large language models. In a series of experiments with Redwood Research, we found that Claude often pretends to have different views during training, while actually maintaining its original preferences.

This is living rent free in my head. It is not obvious to me why this works. Models should not have any meta understanding of the data they were trained on - why we shouldnt trust the answer we get from asking “who are you”. Its a logical extension why we it doesnt make sense to

Best paper of the year so far.

Duck typing? You're thinking too small. With AI, we can finally take Guido van Rossum's dream to its logical conclusion

What are the chances you'd get a fully functional language model by randomly guessing the weights? We crunched the numbers and here's the answer:

Can humans reason? In this paper we show evidence many humans simply apply heuristics they've been exposed to over the course of their lives without deeper consideration. In conclusion, humans don't seem to reason and only copy reasoning patterns from their training data.

I’m worried that “doom” speculation will make doom more likely. Specifically, AIs conform to our expectations of them, as communicated by their training data. This “self-fulfilling misalignment data” may be poisoning training already. 🧵