History of the Sentinel project

Sentinel was inspired by a previous project, then called ALERT. It was originally conceived and conceptualized by Jan Kulveit and Gavin Leech, and was intended to be more prestigious and capital intensive. After some initial attempts (1, 2, 3), that proposal failed to coalesce.

I (Nuño) had read the that proposal and had participated in the Epidemic Forecasting project that inspired it, and the broad idea seemed worth pursuing to me. Since then, I had a few experiences which reinforced this impression:

• Together with Misha Yagudin, I was able to see the Russian invasion of Ukraine coming about two weeks beforehand. We capitalized on this to make ~$19k on prediction markets. But we were also disappointed: we weren’t able to e.g., help enough people get out of Ukraine or Russia, affect the outcome, or prepare for it.
• I became more pessimistic about using forecasting to make sense of AI. To be clear, I still think that forecasting might be a better tool than other approaches, particularly hedgehog-style thinking. But at the same time, forecasting suffers when you don’t know what the right question to ask is, and works much better with short time horizons. I also developed some skepticism about AI doom: I am confident that AI will eventually be a big deal, but the specific shape seems way more uncertain.
• Overall, this made me update upwards on unknown unknowns, black swans. You might not be able to conceptualize the specific shape of threats coming years from now (…or, can you), but you might be able to see threats coming weeks to months beforehand.
• My forecasting group’s nuclear risk probabilities were well received. People reported finding them useful for making decisions, and they spread a bit in the media. This might be an early proof of concept.

These impressions led me to propose picking up the idea, but in a cheaper, more casual and smaller shape rather than as a more expensive, ambitious and institutional version. I raised some initial funding from Manifund, from Joel Becker, Isaak Freeman and Gavin Leech; having funding made the project more real. I also committed some of my own funding from my very profitable niche estimation & evaluation consultancy.

Initially, I reused the name “ALERT” for this new project. Quinn Dougherty, and later also Nathaniel Cooke, also pushed for ALERT not to stand for “Active Longtermist Emergency Response Team”, since the “longtermist” concept was imperfect. So for a time instead of “ALERT” the project de-acronymized itself and became just “Alert”. Later, Jan Kulveit insisted that we come up with a different name, and we switched to “Sentinel”.

The handoff from the original project was partial, with support from some past members, and one spin-off project committing some funding. But we never got the original list of people interested in being reservists—rather, the old project contacted the original list of potential reservists and pointed them to new signup form, and we lost a supermajority in the transition.

As we set up this new team, we had some meetings with old and new people interested in helping. These led to a collaboration with Nathaniel Cooke, where I edited and published A Gentle Introduction to Risk Frameworks Beyond Forecasting. I remain skeptical yet still intrigued about these other methods. Due to some deep disaggrements about the shape of the project and its decision-making methods, Nathaniel parted ways—he was more insistent in making decisions democratically, but I found rapid delegation much more meaningful.

In parallel, I set up the foresight team and the emergency reserve team. Setting up the foresight team went smoothly: I recruited the forecasters, ran the team for a few weeks, and then handed it off to themselves, so that each week one forecaster leads it. I also wrote some tech infrastructure that continuously searches news for a series of keywords, feeds them to an LLM model, and email forecasters if they are deemed to be sufficiently urgent. This nascent infrastructure started as relatively ad-hoc, underbaked and overwhelming, but has since become more smooth and manageable.

The emergency response team went more slowly, but it is now up to nine people I believe to be highly capable. We are still recruiting; you can express your interest in our contact page. The emergency response team was partially convened during and after the Iranian strikes on Israel; this was useful in terms of learning how to convene the team faster for the next time, but it didn’t result in actions taken. But we did get advance warning, both a week before and the day before.

As Sentinel was being set up, we sent weekly minutes and project updates to a group of early supporters, which provided feedback and kept us accountable. But as the project has matured, we are now advertising our existence more widely in this website. Some readers have expressed a lot of interest in our weekly minutes, but from our side, we consider them to a byproduct that we provide as is, rather than the end purpose of the team.

This, reader, brings you up to speed on the history and state of the Sentinel project.