Sweta Karlekar

Hello!

We will be presenting Estimating the Hallucination Rate of Generative AI at NeurIPS. Come if you'd like to chat about epistemic uncertainty for In-Context Learning, or uncertainty more generally. :)

Location: East Exhibit Hall A-C #2703
Time: Friday @ 4:30
Paper: arxiv.org/abs/2406.07457

fun @bleilab.bsky.social x oatml collab

come chat with Nicolas , @swetakar.bsky.social , Quentin , Jannik , and i today

Check out our new paper from the Blei Lab on probabilistic predictions with conditional diffusions and gradient boosted trees! #Neurips2024

Check out our new paper about hypothesis testing the circuit hypothesis in LLMs! This work previously won a top paper award at the ICML mechanistic interpretability workshop, and we’re excited to share it at #Neurips2024

For anyone interested in fine-tuning or aligning LLMs, I’m running this free and open course called smol course. It’s not a big deal, it’s just smol.

🧵>>

Very happy to share some recent work by my colleagues @velezbeltran.bsky.social, @aagrande.bsky.social and @anazaret.bsky.social! Check out their work on tree-based diffusion models (especially the website—it’s quite superb 😊)!

GitHub - andrewyng/aisuite: Simple, unified interface to multiple Generative AI providers Simple, unified interface to multiple Generative AI providers - GitHub - andrewyng/aisuite: Simple, unified interface to multiple Generative AI providers

Just learned about @andrewyng.bsky.social's new tool, aisuite (github.com/andrewyng/ai...) and wanted to share! It's a standardized wrapper around chat completions that lets you easily switch between querying different LLM providers, including OpenAI, Anthropic, Mistral, HuggingFace, Ollama, etc.

Test of Time Paper Awards are out! 2014 was a wonderful year with lots of amazing papers. That's why, we decided to highlight two papers: GANs (@ian-goodfellow.bsky.social et al.) and Seq2Seq (Sutskever et al.). Both papers will be presented in person 😍

Link: blog.neurips.cc/2024/11/27/a...

Sorry John, that isn’t my area of expertise!

This is very interesting! Do you have any intuition as to whether or not this phenomenon happens only with very simple “reasoning” steps? Does relying on retrieval increase as you progress from simple math to more advanced prompts like GSM8K or adversarially designed prompts (like adding noise)?

Many circles of different sizes, representing a visualization of inequality

The Gini coefficient is the standard way to measure inequality, but what does it mean, concretely? I made a little visualization to build intuition:
www.bewitched.com/demo/gini

Interested in machine learning in science?

Timo and I recently published a book, and even if you are not a scientist, you'll find useful overviews of topics like causality and robustness.

The best part is that you can read it for free: ml-science-book.com

Why Can GPT Learn In-Context? Language Models Implicitly Perform Gradient Descent as Meta-Optimizers Large pretrained language models have shown surprising in-context learning (ICL) ability. With a few demonstration input-label pairs, they can predict the label for an unseen input without parameter u...

Learning doesn’t have to mean explicit weight changes; ICL can be viewed as temporary implicit finetuning (arxiv.org/abs/2212.10559) or like a “state” change to the model instead of a weight change, akin to how learning happens in fast RL vs slow RL (www.cell.com/trends/cogni...).

A statistical approach to model evaluations A research paper from Anthropic on how to apply statistics to improve language model evaluations

new paper from Anthropic on LLM evaluation recommendations

www.anthropic.com/research/sta...

Book outline

Over the past decade, embeddings — numerical representations of machine learning features used as input to deep learning models — have become a foundational data structure in industrial machine learning systems. TF-IDF, PCA, and one-hot encoding have always been key tools in machine learning systems as ways to compress and make sense of large amounts of textual data. However, traditional approaches were limited in the amount of context they could reason about with increasing amounts of data. As the volume, velocity, and variety of data captured by modern applications has exploded, creating approaches specifically tailored to scale has become increasingly important. Google’s Word2Vec paper made an important step in moving from simple statistical representations to semantic meaning of words. The subsequent rise of the Transformer architecture and transfer learning, as well as the latest surge in generative methods has enabled the growth of embeddings as a foundational machine learning data structure. This survey paper aims to provide a deep dive into what embeddings are, their history, and usage patterns in industry.

Cover image

Just realized BlueSky allows sharing valuable stuff cause it doesn't punish links. 🤩

Let's start with "What are embeddings" by @vickiboykis.com

The book is a great summary of embeddings, from history to modern approaches.

The best part: it's free.

Link: vickiboykis.com/what_are_emb...

(Shameless) plug for David Blei's lab at Columbia University! People in the lab currently work on a variety of topics, including probabilistic machine learning, Bayesian stats, mechanistic interpretability, causal inference and NLP.

Please give us a follow! @bleilab.bsky.social

Hi! Our lab does Bayesian stuff :) Could you add Dave Blei's lab to this pack as well if it's not already full? @bleilab.bsky.social

Could you add Dave Blei's lab to this pack as well if it's not already full? @bleilab.bsky.social

Could you add Dave Blei's lab to this pack as well if it's not already full? @bleilab.bsky.social

Could you add Dave blei's lab to this pack as well if it's not already full! @bleilab.bsky.social

We created an account for the Blei Lab! Please drop a follow 😊

@bleilab.bsky.social

GitHub - thu-ml/SageAttention: Quantized Attention that achieves speedups of 2.1x and 2.7x compared to FlashAttention2 and xformers, respectively, without lossing end-to-end metrics across various mod... Quantized Attention that achieves speedups of 2.1x and 2.7x compared to FlashAttention2 and xformers, respectively, without lossing end-to-end metrics across various models. - thu-ml/SageAttention

Almost 3x faster than FlashAttentiion2 github.com/thu-ml/SageA...

Oh, I’ve been meaning to check out that YouTube series—thanks! Also sadly, there's no class website, but I can share the "super quick intro to mech interp" presentation I made. It’s somewhat rough, but hopefully, it gets the main points across! sweta.dev/files/intro_...

Mailing list contact information Information to be added to the post-Bayes mailing list.

📢 Post-Bayesian online seminar series coming!📢
To stay posted, sign up at
tinyurl.com/postBayes
We'll discuss cutting-edge methods for posteriors that no longer rely on Bayes Theorem.
(e.g., PAC-Bayes, generalised Bayes, Martingale posteriors, ...)
Pls circulate widely!

An Extremely Opinionated Annotated List of My Favourite Mechanistic Interpretability Papers v2 — AI Alignment Forum This post represents my personal hot takes, not the opinions of my team or employer. This is a massively updated version of a similar list I made two…

Adding Neel Nanda's favorite paper list as well: www.alignmentforum.org/posts/NfFST5...
(6/n)

Can’t believe I forgot about this paper, thanks so much!!

I haven’t read the first one but it looks very informative, thank you!! We also had a separate unit on transformer architecture; I’m going to add this to that paper list as well!

Starter packs I found:
AI (*about* AI, not *for* an AI) go.bsky.app/SipA7it
Spoken Language Processing bsky.app/starter-pack...
Diversify Tech's pack bsky.app/starter-pack...
Women in Tech bsky.app/starter-pack...
Great UK Commentators bsky.app/starter-pack...
Linguistics bsky.app/starter-pack...

Ooo this is a really cool resource; I didn’t know about it before, thank you for sharing the link!

Would love to be added!