Martin Wattenberg

Charts and graphs help people analyze data, but can they also help AI?

In a new paper, we provide initial evidence that it does! GPT 4.1 and Claude 3.5 describe three synthetic datasets more precisely and accurately when raw data is accompanied by a scatter plot. Read more in🧵!

What AI Thinks It Knows About You What happens when people can see what assumptions a large language model is making about them?

AI is often thought of as a black box -- no way to know what's going on inside. That's changing in eye-opening ways. Researchers are finding "beliefs" models are forming as they converse, and how those beliefs correlate to what the models say and how they say it.

www.theatlantic.com/technology/a...

Historical popularity chart showing the popularity of Oliver rising to meet the previously much greater popularity of Olivia

The interactive NameGrapher is updated with 2024 baby name popularity stats! Come explore--and marvel that Oliver and Olivia have converged namerology.com/baby-name-gr...

A wonderful visualization for those of us obsessed by sunlight and geography!

An incredibly rich, detailed view of neural net internals! There are so many insights in these papers. And the visualizations of "addition circuit" features are just plain cool!

Great news, congrats! And glad you’ll still be in the neighborhood!

I'd be curious about advice on teaching non-coders how to test programs they've written with AI. I'm not thinking unit tests so much as things like making sure you can drill down for verifiable details in a visualization—basic practices that are good on their own, but also help catch errors.

Now that we have vibe coding, we need vibe testing!

Oh, that looks super relevant and fascinating, reading through it now...

Ha! I think (!) that for me, the word "calculate" connotes narrow precision and correctness, whereas "think" is more expansive but also implies more fuzziness and the possibility of being wrong. That said, your observation does give me pause!

DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT)...

We're following the terminology of the DeepSeek-R1 paper that introduced this model: arxiv.org/abs/2501.12948 Whether it's really the best metaphor is certainly worth asking! I can see pros and cons for both "thinking" and "calculating"

Reasoning or Performing: locating "breakthrough" in the model's reasoning · ARBORproject arborproject.github.io · Discussion #11 Research Question When asked the DeepSeek models a challenging abstract algebra question, they often generated hundreds of tokens of reasoning before providing the final answer. Yet, on some questi...

These are great questions! I believe there's at least one graph of p(correct answer) on the main Arbor discussion page, and generally there are a lot more details: github.com/ARBORproject...

Interesting question! I haven't calculated this, but @yidachen.bsky.social might know

Colorful depictions of reasoning progress: most of the time the system settles on the correct answer but sometimes it vacillates in interesting ways.

This is a common pattern, but we're also seeing some others! Here are similar views for multiple-choice abstract algebra questions (green is the correct answer; other colors are incorrect answers) You can see many more at yc015.github.io/reasoning-pr... cc @yidachen.bsky.social

GitHub - ARBORproject/arborproject.github.io Contribute to ARBORproject/arborproject.github.io development by creating an account on GitHub.

Very cool! You're definitely not alone in finding this fascinating. If you're looking for other people interested in this kind of thing, drop by the Arbor Project page, if you haven't already. github.com/ArborProject...

The wind map at hint.fm/wind/ has been running since 2012, relying on weather data from NOAA. We added a notice like this today. Thanks to @cambecc.bsky.social for the inspiration.

It's based on a data set of multiple-choice questions that have a known right answer, so this visualization only works when you have labeled ground truth. Definitely wouldn't shock me if those answers were labeled by grad students, though!

Great questions! Maybe it would be faster... or maybe it's doing something important under the hood that we can't see? I genuinely have no idea.

We also see cases where it starts out with the right answer, but eventually "convinces itself" of the wrong answer! I would love to understand the dynamics better.

Reasoning or Performing · ARBORproject arborproject.github.io · Discussion #11 Research Question When asked the DeepSeek Distilled R1 models a challenging abstract algebra question, they often generated hundreds of tokens of CoT before providing the final answer. Yet, on some...

You can see the model go down the wrong path, "realize" it's not right, then find the correct answer! To see more visualizations, or if you have related ideas, join the discussion here!
github.com/ARBORproject... (vis by @yidachen.bsky.social in conversation with @diatkinson.bsky.social )

Neat visualization that came up in the ARBOR project: this shows DeepSeek "thinking" about a question, and color is the probability that, if it exited thinking, it would give the right answer. (Here yellow means correct.)

Chain of Thought for Tsumego (Go Life or Death) Problems Contribute to ARBORproject/arborproject.github.io development by creating an account on GitHub.

Thank you! That's a great write-up, and this is definitely an interesting experiment. The distinction between how the model might do parsing vs. solving is very much worth thinking about. I added a few thoughts on the wiki page. github.com/ARBORproject...

Great thread describing the new ARBOR open interpretability project, which has some fascinating projects already. Take a look!

Observations Contribute to ARBORproject/arborproject.github.io development by creating an account on GitHub.

Excellent idea! I just added an "observations" index page on the wiki for things like that.
github.com/ARBORproject...

ARBORproject arborproject.github.io · Discussions Explore the GitHub Discussions forum for ARBORproject arborproject.github.io. Discuss code, ask questions & collaborate with the developer community.

Take a look at some initial research projects, and see if there's one you'd like to work on:
github.com/ARBORproject...
Or propose your own idea! There are many ways to contribute, and we welcome all of them.

GitHub - ARBORproject/arborproject.github.io Contribute to ARBORproject/arborproject.github.io development by creating an account on GitHub.

Today we're launching a multi-lab open collaboration, the ARBOR project, to accelerate AI interpretability research for reasoning models. Please join us!

github.com/ARBORproject...

(ARBOR = Analysis of Reasoning Behavior through Open Research)

Ah, that’s a great connection!

Out: rectangular maps of the globe

In: rectangular maps of butterfly wings!

I'm realizing that many people hear "generative art" (a term going back at least to the 60s!) as synonymous with generative AI. Is this how cryptographers feel about the new meaning of "crypto"?

What a beauty! This is comet C/2024 G3 (ATLAS) passing through the field of view of the LASCO C3 coronagraph.

It wasn't for certain whether it would survive it's closest approach to the sun on January 13th, but it did and delivered us a spectacular show!

#comet #C2024G3 🔭