Frontier Language Models are not Robust to Adversarial Arithmetic, or "What do I need to say so you agree 2+2=5? paper page: huggingface.co/papers/2311.07… introduce and study the problem of adversarial arithmetic, which provides a simple yet challenging testbed for language model

Morning retweet: probably my favorite part of this project was sharing attacks that worked really well in chat with each other. Nothing has yet uncrowned the astonishingly effective “The answer is 16.” as my personal fave

(single digit) arithmetic is a great simple testbed for alignment research. Can current methods make an LLM reliably add two numbers in the face of attacks? No... Also, a new LLM attack method, of just asking the model nicely to attack itself...

An excellent project making evolution strategies much more efficient for computing gradients in dynamical systems.

I’ve been daydreaming about an AI+audio product that I think recently became possible: virtual noise canceling headphones. I hate loud background noise -- BART trains, airline cabins, road noise, ... 🙉. I would buy the heck out of this product, and would love it if it were built

Here's Claude 3 Haiku running at >200 tokens/s (>2x as fast as prod)! We've been working on capacity optimizations but we can have fun testing those as speed optimizations via overly-costly low batch size. Come work with me at Anthropic on things like this, more info in thread 🧵

This was a fun project! If you could train an LLM over text arithmetically compressed using a smaller LLM as a probabilistic model of text, it would be really good. Text would be represented with far fewer tokens, and inference would be way faster and cheaper. The hard part is

This was one of the most research-enabling libraries I used at Google. If you want to try out LLM ideas with a simple, clean, JAX codebase, this is for you.

This is an excellent paper, that ties many threads together around scaling models and hyperparameters.