Sentinel
Article
Sentinel is a recurring organization in the Astral Codex Ten archive, appearing 5 times across 5 issues between September 16, 2024 and February 25, 2026. The archive places it in contexts such as “Sentinel is a foresight and emergency response team seeking to detect and react fast to large-scale global risks”; “Sentinel is an organization that forecasts and responds to global catastrophes”; “Sentinel (group with superforecasters monitoring world events) predicts a 39% chance”. It most often appears alongside California, FDA, United States.
Metadata
- Category: Organizations
- Mention count: 5
- Issue count: 5
- First seen: September 16, 2024
- Last seen: February 25, 2026
Appears In
- Open Thread 347
- H5N1: Much More Than You Wanted To Know
- Links For February 2025
- ACX Grants Results 2025
- The Pentagon Threatens Anthropic
Related Pages
-
- California (3 shared issues)
-
- FDA (3 shared issues)
-
- United States (3 shared issues)
-
- Africa (2 shared issues)
-
- Anthropic (2 shared issues)
-
- Australia (2 shared issues)
-
- China (2 shared issues)
-
- COVID (2 shared issues)
-
- Dean Ball (2 shared issues)
-
- Manifold (2 shared issues)
-
- Manifold Markets (2 shared issues)
-
- Metaculus (2 shared issues)
External Links
Source Context
Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.
2: Sentinel is a foresight and emergency response team seeking to detect and react fast to large-scale global risks (eg a pandemic). The foresight team is run by some members of Samotsvety, the forecasting team that won the CSET Foretell tournaments several years in a row, and they publish some weekly minutes that tracks ongoing threats. They're seeking a larger audience for their reporting, collaborators, and funding; if you want to reach out you can do so at hello@sentinel-team.org.
H5N1 may cross to humans, but it could take a while. Superforecaster Juan Cambeiro at The Institute For Progress estimated a 4% chance of a “worse than COVID” H5N1 pandemic in “the next year”, but their estimate was made in 2023, without the benefit of the Metaculus estimates or most of our current knowledge. This feels high now - Metaculus says 5% total for H5N1 pandemic, and most pandemic flus are not worse than COVID. IFP also seem to be expecting a case fatality rate greater than 10%, which I find unlikely for the reasons mentioned above. I trust their estimate less than Metaculus’ current ones. I conclude that the most plausible estimate for the chance of an H5N1 pandemic in the next year is 5%. Interestingly, 5% is about the base rate for pandemic flus per year: five in the past century = one per twenty years = 5% chance per year. Isn’t it surprising that we’re still at the base rate when we can see a dangerous-looking flu virus spreading through the types of animals that have caused pandemic flus in the past? Part of the answer is that we’re not - in addition to the 5% chance of H5N1, we have to add the chance of some other pandemic flu. This probably isn’t 5% on its own; scientists monitor flu strains closely, and they haven’t found any others which are giving off as many red flags as H5N1. Still, something could always come out of left field. Maybe we should add a 2.5% chance of some other strain, for a total of 7.5% chance of a flu pandemic (ie beyond normal seasonal flu) next year. But still, isn’t it surprising that we’re so close to the base rate? One way to think about this: the base rate represents how concerned we should be if there was no epidemiological monitoring at all. In that case, we would estimate a probability distribution across different epidemiological landscapes, most of which contain some concerning-looking flu strains. Since we are doing the epidemiological monitoring, we can collapse that distribution into a single picture: one flu strain, H5N1, is in fact pretty concerning, and other strains mostly aren’t. This is enough to move our prior from 5% to 7.5%, but no more. The forecasters I talked to raised one other point of uncertainty: does the flu work more like a dice roll, or like a bus? Dice rolls are uncorrelated with their predecessors; even if it’s been a hundred rolls since you last rolled a 6, your chance this time is still 1/6. But buses come at fixed intervals; if the buses are hourly, and you haven’t seen a bus in the past 59 minutes, then your chance of seeing a bus in the next minute is very high. It’s been 16 years since the last flu pandemic; these pandemics come (on average) every 20 years. I don’t think anyone has a good sense of how to think about this. But it was 40 years between the Spanish and Hong Kong flus, so the twenty year number is at best a rule of thumb. The 5% number feels very low to me (and, apparently, to the average Manifold forecaster). Isn’t H5N1 spreading to cows and pigs and all sorts of other mammals? Isn’t it in the news all the time? I trust Metaculus a lot, but I agree that this is a surprising update, and I’m taking it on faith rather than feeling it in my bones. What Would The Fatality Rate Be For An H5N1 Pandemic? There are four basic stories you could tell about likely H5N1 mortality. First, maybe mortality would be 50%. The argument here is that official statistics report this mortality rate in the chicken farmers who have been infected with H5N1 so far. Several news sources and even some scientists have raised the specter of a pandemic version of H5N1 pandemic with this same death rate, which could kill a quarter to a third of the world population. THIS IS EXTREMELY FAKE. The official statistics only report fatality rate in the infections we know about. Bird flu is rare, there’s no mass testing, and we only learn that somebody had it if they’re in a hospital and the doctors are worried enough to test for rare conditions. Of Americans who got bird flu in the past year, 0 out of 61 have died. Probably this is mostly because America upped its detection game and is now finding milder cases; we also can’t rule out the virus mutating to become less virulent. Metaculus estimates the current true mortality rate as 1.25%. …but leaves a wide 90% confidence interval, from 0.5% to 7%. Second, maybe mortality would be somewhere around 1.25%. The argument here is that Metaculus uses this as its central estimate of US mortality. But Sentinel discusses some reasons to be skeptical of broad inferences from the US numbers: Scientists have been puzzled by the apparently low H5N1 case fatality rate in humans in the US. They offer a number of hypotheses: “The way in which the virus is being transmitted — along with the amount of virus exposure — is limiting the severity of disease.”
Thanks to Nuño Sempere and Sentinel for help and clarification. Sentinel is an organization that forecasts and responds to global catastrophes; you can find their updates, including on H5N1, here. As usual, any errors are mine alone.
23: Sentinel (group with superforecasters monitoring world events) predicts a 39% chance that the Trump administration ignores at least one SCOTUS decision (conditional on there being one against them), and a 72% chance of a “free and fair” election in 2028 (assuming no existential catastrophe before then). I wonder what their 28% vision of a “non free and fair” election looks like.
Inline links: predicts
Nuno Sempere, $50K, for disaster forecasting and response. Nuno runs Sentinel, a team of superforecasters which tracks incipient disasters (pandemics, wars, etc) and brainstorms pre-hoc and post-hoc responses. Their model for response are groups like VaccinateCA, a small team of Californians who noticed that the state’s COVID vaccine policy was disorganized and made a site that helped connect people with spare vaccination capacity. You can see their blog here. Nuno is an ACX Grants evaluator; due to conflict of interest, this grant is being covered in conjunction with an outside funder.
Inline links: here
Superforecaster Nuño Sempere, maybe as part of his work with Sentinel. He seems to think higher chance of supply chain risk than others, but that supply chain risk might be handled in a way that only affects DoD contracts themselves, which wouldn’t be so bad. I haven’t heard anyone else make this distinction. Tweet here, full document here. And big praise to most other AI companies, including Anthropic’s competitors, for standing up for them and for the AI industry more broadly: