Charlie Snell

I'm a third year CS PhD student in Berkeley EECS advised by Dan Klein and Sergey Levine. Previously, I was a UC Berkeley undergrad, where I had the great opportunity to work with and learn from a number of fantastic AI researchers, such as Sergey Levine, Ruiqi Zhong, Dan Klein, Jacob Steinhardt, and Jason Eisner. I was also previously a Student Researcher at Google DeepMind.

Email  /  Google Scholar  /  Twitter  /  Github

Can we predict emergent capabilities in GPT-N+1 🌌 using GPT-N, which has random performance on the task? To do this, we use information about how pre-emergence model checkpoints behave under the influence of task-specific finetuning to obtain predictive power about the point of emergence in the few-shot setting.

On difficult problems, humans tend to think longer to improve their decisions. Can we instill a similar capability into LLMs? And how well can it perform? We find that by optimally scaling test-time compute we can outperform much larger models in a FLOPs matched evaluation.

Recent systems – like Koala, Vicuna, and Alpaca – finetune a weaker language model to imitate the outputs of a stronger model, like ChatGPT or GPT-4. In this work, we critically analyze the shortcomings of this approach.

Language models significantly benefit from context tokens, such as prompts or scratchpads. We propose to apply context distillation so that a language model can improve itself by internalizing these gains.

We propose an effective and easy-to-use offline RL motivated method for steering language models towards successfully completing language tasks, such as goal directed dialogue, controled generation, and word games.

We introduce APEL, a new framework that enables non-programmers to indirectly annotate natural language utterances with executable meaning representations, such as SQL programs.

How do two distributions of text differ? We propose a method for automatically summarizing the differences by "learning a natural language hypothesis".

We extend techniques from learning-based control, such as task relabeling, to derive a simple and effective method to finetune language models in a goal-aware way.

Why do models often attend to salient words, and how does this evolve throughout training?

We augment the original Omniglot dataset with a new dataset of children's handwritten characters. We then study the properties of a Bayesian Program Learning model trained on this new data.

A tour through the wonderful AI art scene that emerged when CLIP was released in January 2021.

A technical and philosophical discussion of how DALL-E works, why it is so effective at generating images from a text prompt, and its theoretical limitations.

How do vector quantized variational autoencoders (VQ-VAEs) work? And what role do they play in modern generative models, such as DALL-E and Jukebox?

Built on top of HuggingFace's Transformers library, JaxSeq enables training very large language models in Jax with model and data parallelism across both multi-device and multi-node clusters.

Re-create the dramatic train/test curves from the original paper; experiment with the grokking phenomenon yourself.

Harness the power of deep music representations to generate playlists and visualize your music preferences in an interactive web app.

A deep learning framework implemented from scratch in C++/OpenCL. Implements GPU kernels that can run on a 2013 Macbook Air GPU (and other Apple computers). Implements LSTM training/inference for music lyric generation.

A humorous sound-box app.

Scroll through an infinite 2D block-world consisting of rugged terrain, endless caves, fluffy clouds, and extreme biomes all synthesized by PRNGs and Perlin Noise.

Website design from Jon Barron