Field Building

Let’s wander through the fields for a bit.

I’ll document what field building appears like to the person experiencing the building. I’ll start with basic beliefs about communities, then get specific about the experience, talk problems, and end on solutions.

Communities

Communities are defined by their knowledge. Knowledge communities, those that create new knowledge are especially sensitive to who decides what is considered knowledge and what is not. New scientific knowledge communities always start off quite rebellious. That’s the spirit of science. It’s about telling truth to power. It’s about contesting authority. Sure, you may believe something, but here are reasons to the contrary.

And then infrastructure shows up and frames, shapes, and sometimes, strangles that community. Journal editors and university administration are the usual piñatas, but I reckon it has more to do with a system of incentives. Reputations don’t just defend themselves after all, and sometimes stakes in the mirror are smaller than they appear.

I’ve joined established communities with enormous infrastructure. And I’ve been to a few inceptions.

Years ago, a good friend on a Standards Council Canada committee told me about AIGS (Artificial Intelligence Governance & Safety Canada) and suggested that I check them out. I attended a meeting, joined their slack, and lurked. Years laters, somebody posted that Trajectory Labs was hosting a meeting on technical AI Safety. And that seemed pretty neat. I went.

And I loved it. So much so that I returned. These are amongst the best events in the city. Georgia is among the most organized of the organizers, the talks are high quality and the people who self-select there are smart and passionate. It’s comparable to an AI meetup in SF. A rare feat in Toronto.

Mario, the lab founder, and Georgia will promote adjacent events with fantastic one liners. There’s a hackathon on some problem coming up. There’s a hackathon on some other area that you’ve never heard of that’s going on next month. One caught my interest. Trajectory Labs housed the hackers for the weekend. I met good people. I learned a lot. There’s good signal. I tried another hackathon and learned more. Georgia hosts a book club. She’s just brilliant at socially sorting people by their p(doom) numbers. It’s a lot of fun.

At a regular Thursday technical meetup Georgia announces that ML4Good is hosting a one week intensive in June. “You should apply,” she says. 

“They’re good people?”

“Yeah!” She says without hesitation.

That’s signal. And I have some intuition that there are severe open problems in AI Safety. That’s intriguing. I am intrigued. Can I help?

I apply. I’m interviewed. I’m accepted.

It takes around 25 hours to do the pre-work in May. A third of it is standard AI safety reading. Two thirds are technical review and coding. It’s all quality material. I re-integrate some of the AI safety language with comparatively fresher product management experience and some of the older road safety policy knowledge. It’s all great. I attend the ML4Good AI Safety Technical Bootcamp (Canada26, June 1 to June 9) out in the woods outside of Montreal with 19 other students, 1 professor, 3 teaching assistants and Georgia.

The days are optimized for absorption. I learn new perspectives about the material I know. The material I’m familiar with is extended. The new material goes in. The second order new material causes a bit of choking but it’s flowing down the gullet. By the third order of new material I’m leaning on the LLM to translate and integrate with my home field language. I experience the rare stackoverflow.

And it’s glorious.

ML4Good curated people from across the continent. An astrophysist from Colorado. A lawyer from Montreal. A data scientist from Connecticut. Quite a few software engineers from a range of enterprises, and two entrepreneurs among many. There were quite a few similarities amongst the group. Visually, if everybody were silent, you wouldn’t be able to tell that you were at an AI Safety event. They don’t wear a uniform like many professions. There are patterns in their perception space and how they talk.

How can you quickly assess the health of a community? Listen to the sound it makes when it’s physically present. Laughter is usually the best signal. Loud chatter is second. There’s a lot of both.

I enjoyed the social hacking games in particular. By the end of the last game I was convinced that my incentives were getting shaped in ways I didn’t understand.

There was a key moment, around 11pm, during a particularly fun slice of social hacking which involved a game theoretical technique described as “diabolical”, I had been seen. And I knew. A tourist no more.

I was a part of it now.

Over the next few walks I listened to the underlying opportunity space a lot more. Where was the pain felt. Where it was unconscious.

I saw good people experience dozens of problems. I watched some struggle. Some of them appear to emerge from incentives endemic to the field. Some appear wicked. Some are deeply interesting. Some of them appear to have potentially interesting solutions. I got what I came for. And I’m thankful for the entire experience chain.

Problems

I see problems, unsatisfactory gaps in the status quo, with fresh eyes. Sometimes that perspective is useless. Sometimes it’s used less. And sometimes it might be useful. It’s difficult to know.

The unifying stance appears to be: the perception of uncertainty regarding AI technology outcomes.

Members express their uncertainty with error bars and outcomes with stories supported by scalar values. Knowledge of the stories aren’t themselves contested. Variance is expressed as uncertainty and backed up with proportional action. If you believe a story is likely, it follows that you invest attention in reducing its probability.

Members experience genuine urgency. Communities advance their science at different rates. Sometimes they slow down because they fail at fieldwork. Sometimes the infrastructure itself causes the slowdown. There’s some awareness that given the distribution of risk over time, one should prioritize moving quickly against what matters.

Risk communication is difficult skill and easy to get wrong. The public has a hard time with understanding it because it usually involves several logical chains. X has N properties that make Z likely with probability P with consequence Y. And as a direct result it can be easy to misunderstand the claims. The assumption that there is baseline effort to understand a claim can be problematic. Relevance only goes so far in connecting an incentive to an action.

Advancing the knowledge frontier is difficult and can be slow. Becoming aware of the frontier itself is a challenge given just how much tacit knowledge remains buried in minds and on slack. Once you’re there you can’t advance the science at a typical peer reviewed pace. You got to automate the test bench, create evidence, shift theory, and advance the state of the art.

I heard stories of burnout.

Solution Building

The solution space isn’t obvious.

The wrong infrastructure will certainly strangle progress. So what is the right infrastructure?

A lack of urgency will reduce the rate that the knowledge frontier expands. But extreme urgency causes burnout, which destroys social capital and the ability for the frontier to advance.

Risk communication is both an art and a science, because every generation wants to be messaged differently, and because every generation overestimates its uniqueness.

Decision making is the process of taking personal responsibility for a proposed course of action.

Field building may be thought of the system of helping people arrive at a proposal for a course of action. There’s a clue here that a solution has been discovered with field building.

Thanks to Georgia, Mario, Trajectory Labs, and ML4Good in particular for fielding me along on the journey.