New Directions

Oct 26 2022

I’m excited to finally share some news: I’ve resigned my position on the NYU faculty and started working full time as Vice President of Information Design at Nomic, a startup helping people explore, visualize, and interact with massive vector datasets in their browser.

This will be a big shift. I’ve spent my whole career up to this point in academic institutions; but right now, Nomic is the best possible place to tackle the most important and interesting questions that I’ve spent years thinking about. How do we interact with huge collections of texts, images, and information? How do we interpret, critique, and improve the implicit knowledge bases that institutions rely on? Today that means being able to give shape to digital text and images and to build new tools for machine learning interpretability.

Almost two years ago I wrote a blog post about the web and the future of data programming. I scratched from the early drafts a few paragraphs about Halt and Catch Fire, a top-10 all-time TV show, about the joys and frustrations of knowing that something important is amassing on the horizon and not being sure if you’ll be able to take part. For three years, I’ve been watching as representation learning models (e.g. BERT, GPT-3, CLIP, and DALL-E), multi-language binary serialization formats (e.g. Apache Arrow), and tools for scalable data visualization and analytics in the browser (WebGL and WebGPU), have all simultaneously experienced massive technical inflections, directing them towards a common destination.

I want to be as close to that impact site as possible, and for me it won’t be in a history department. While historical datasets present some of the most compelling playgrounds for work applying these technologies, the academic habit of treating building as play makes it hard to fully realize the potential of these shifts. Actually developing the tools and frameworks necessary for this visualization has been a spare-time hobby compared to teaching, administration, and research. Even academic centers in data science and CS (which obviously produce incredible work in the AI field) are well behind industry in thinking through the systems and engineering required to bring these tools to the world.

Knowing this, I’ve been talking to a lot of people in these fields recently. Out of all them, Brandon and Andriy at Nomic, and their vision for making AI more transparent while making datasets more visible via AI models, are the people that most trip my Halt and Catch Fire test. Something interesting is happening right now as AI models get bigger, as dimensionality reduction algorithms proliferate, and as web standards emerge that make the browser a compelling computing environment.

Over the past few months I’ve been watching Brandon and Andriy improve their models and create rich interfaces for exploring, filtering, and even editing embedding spaces. I’ve been incredibly impressed by their progress and am convinced that, given the extremely specific interests I’ve developed over the past few years, Nomic is the best place to be doing the kind of work I’m really interested in doing.

A map I’m excited to share soon

If you pay only glancing attention to “artificial intelligence,”embedding spaces might seem like an arcane detail to be so excited about. But they’re critical–not just for machine learning pipelines, but for the whole cultural apparatus we inhabit today. When you listen to new music from your streaming subscription, it’s chosen based on embedding vectors for the songs and an embedding vector for you. Unified spaces for representing image and text embeddings have unleashed a dizzying cascade of innovations in generative AI over the last six months through models like DALL-E and Stable Diffusion. Search engines, recommendation systems, translation algorithms–anywhere there is an AI model, there is an embedding space underpinning it. And understanding and navigating these multidimensional spaces has been a key concern of data visualization for longer than most people know. For years I’ve assigned my classes–to the bemusement and amusement of students–an absolutely amazing Stanford Linear Accelerator video featuring the legendary statistician John Tukey manipulating a nine-dimensional scatterplot with a custom-made array of knobs. Nowadays we all use UMAP, T-SNE, and newer methods for trying to disentangle spaces like this, but the concerns and goals are real and satisfy a need that’s been around since the earliest days of exploratory data analysis.

I’ve worked on a lot of different projects in this general area over the years, but one that’s especially important here is Deepscatter, my personal typescript/WebGL library for visualizing arbitrarily large collections of points in the browser. For the last two years I’ve been captivated by the possibilities here, even though they haven’t fit into any of the work I’ve been doing at NYU. While I’ll have to set a lot of my other projects aside, at Nomic I’ll get to spend a lot more time expanding the possibilities for defining and exploring large embedding spaces. I met Brandon and Andriy through their contributions to deepscatter, and providing pointers as they build a fork into their new product, Atlas. As part of this new position I’ll get to spend more time working building out features I’ve long had in mind for Deepscatter but haven’t had the bandwidth or support to pursue, and sharing some new and exciting maps. This should be good news for everyone I know using Deepscatter now, both because I’ll be able to implement these features, and because Nomic’s internal fork enables some very exciting possibilities including search, selection, and filtering.

From now on this improved library will live at github.com/nomic-ai/deepscatter repo under a CC-BY-NC-SA license, where NC means research and personal use is encouraged, but any commercial applications require a license from Nomic. If you have any questions about using Deepscatter for something, join .

But you can also start making maps more easily and robustly by using Atlas. If you have a large collection of text, embeddings, or something else, do reach out! Atlas is invite-only right now, and you can join the waitlist here. I’m excited to start showing off some of what we’ve been working on–helping set up full-text search has been revelatory about what kinds of data interactions are now possible.

I’ve written and discussed a lot over the years about the humanities, the university, the sciences, and all the rest, so leaving at this moment feels a bit more fraught for me than it would for most. Some of our redoubts are dealing with a slight fire and brimstone problem–I’m sure I’ll take some chances to look back on those bigger questions soon. But not too soon–don’t want to turn into a pillar of salt.

I do want to thank and note some people at NYU as I go, though. In the past three years many students and faculty have made great strides in digital humanities, and it has been exciting to help introduce many students to digital humanities work and to create spaces that encourage new and interesting work. In my role as director of digital humanities I launched, alongside Zach Coble in Digital Scholarship, a new seed grant program that has funded sixteen DH projects: several have already earned major external grants, and I’m sure you’ll be hearing more from some of them in the future. . I also managed to cobble together funding for a new series of summer fellowships starting in 2021: running this summer class with Jojo Karlin and others at the libraries has been extremely rewarding. (–and I should say it’s a delight to be able to link to the new website that we built last spring and which Marii Nyrop superintended in just one of their irreplaceable contributions to DH community life at NYU.) I co-directed, with Ellen Noonan and Sibylle Fischer, the Asylum Lab, what was my intellectual lodestar at the university taking an interdisciplinary approach to understanding the life stories migrant records from the last hundred years with a group of graduate students and an undergraduate class. And teaching, talking to, and working with students from all levels and fields at NYU was uniformly a joy.

But while it’s hard to walk away, like so many people during this pandemic I realized that there’s no time to waste. And I’m excited to see what’s next.