Boston-based, Director of Detection at SecureBio, GWWC board member, parent, musician. Switched from earning to give to direct work in pandemic mitigation. Married to Julia Wise. Speaking for myself unless I say otherwise. Full list of EA posts: jefftk.com/ânews/âea
Jeff Kaufman đ¸
Karma: 19,156
State of PanÂdemic Early Warning
IniÂtial DIY CleanÂroom Experimentation
Adds additional context to https://âârevive.blog/ââb/ââworldoptimization on Manifoldâs new Blog Revival Project, with Austin pledging towards reviving his then-anonymous coworkerâs blog.
(Just found this interesting and mildly funny, not saying the pledging or project is bad. Though I agree with others that handling finances at a two-person org is very clearly a bad role for someone convicted of financial fraud, and âI felt a keen debt to the FTX Future Fundâ should point the other way.)
It would be possible, though I do think we get quite a lot of value from being a more general biosecurity organization, including cross-pollination of ideas and joint policy work.
Thoughts on TakÂing OpenAI FounÂdaÂtion Funding
Thatâs a great question, and Iâm sorry it has taken so long to get an answer for you! We think we can achieve a three day end-to-end time by mid-2027. This will be close to the limit of what is possible to do with current technology, when you count time for shipping samples (~3/â4 day), preparing them to go on the sequencer (~1/â2 day), running the sequencer (1 day), and processing the output (a few hours).
I do think itâs worth it to continue looking for ways to improve from there, and I expect improvements in sequencer technology to allow us to eventually cut off another ~half day. On the other hand, getting down to a day or lower is less important than you might think. Say you have a detection system with sensitivity S and end-to-end time E, and youâre trying to flag a pathogen that doubles every Td days. Unless S is so high that you can flag something when the very first people are infected, this system will have approximately the same real-world performance as one with sensitivity 2S that runs in E + Td days. Now, doubling sensitivity can be an expensive proposition: after youâve picked the low-hanging fruit itâs approximately doubling the cost of your detection system. But it means that unless youâre targeting a sensitivity high enough to identify outbreaks when only a handful of people have been infected (which would be fantastic of course, but isnât economically feasible with current technology) you can get to a given sensitivity target without pushing end to end time to near zero.
I think thereâs another category before 9
Edited the post to add this.
I think most cuts should be something that we could replicate without significant price difference.
Sorry, I was a little fuzzy in my wording. My claim is (a) meat eaters have preferences + variety seeking even at the level of differences between cuts, (b) meat replacements will likely get very close to animal meat but not as close as some of these differences, and so (c) humans without a moral objection to eating animals will still eat some animal meat for variety.
For example, people still eat rutabaga, even though for any given rutabaga use other options (usually potato) are tastier and cheaper. People still eat red delicious and other kinds of apples that I think are basically obsolete. Iceberg lettuce is still widely eaten even though there are multiple options with simultaneously better flavor and nutrition. And I donât think thereâs much of a case for lima beans. There are some obsolete vegetables that weâve stopped eating, but they basically need to be significantly worse, either in the ânot worth eatingâ sense or âtoo expensive for what you get senseâ. In my model a food keeps some market as it still wins on some axis: cost, convenience, storage, flavor, texture, maybe tradition. And to disappear it needs to be beaten on every axis at once (or close enough that itâs not worth the extra cost).
Also probably cultivated gives you more freedom to explore conventionally infeasible options.
I think thatâs right, but I think that doesnât get factory farming to zero: it gives people more options, but a drive for variety keeps all the options in play.
I think it goes well beyond species: meat eaters have strong preferences over specific portions (âcutsâ). They have very different tastes and textures that I think will be very challenging to replicate. Instead, my expectation is that weâll get something more similar to the introduction of vegetable shortening: products that replace many animal uses, but not all.
Assuming we end factory farming before 2100, which factor will have contributed the most?
I expect these will not be exact substitutes, and so absent moral changes people would eat factory farmed meat for variety,
A bunch of objections!
-
My largest is that I donât think we should rely heavily on the outside view. The whole EA project is trying to figure out where we can have the largest impact, and this involves thinking carefully about the effects of actions. What happened when people tried things in the past is some evidence, but there are also important differences between what weâre trying and what people have done before (we can learn!), and between previous and current situations (the world changes).
-
I donât think this analogy fits: Iâm donating to people in both US parties who look like theyâd do good work to reduce risk from AI and bio. I donât really see a way describing these folks as âmy Groupâ makes sense.
-
It also doesnât look to me like this cuts cleanly even if we do accept the analogy: thereâs also historical precedent for âtry to help poor people effectivelyâ going poorly.
-
I initially thought you were saying this was hard in the sense of it being hard for an evaluator, but then I noticed that your comparison was to âfollowing GiveWell recommendationsâ and not âbeing a GiveWell evaluatorâ. How much are you thinking about each?
If itâs recommendations, what I do personally is follow the ones from @Eric Neyman âs working group.
If itâs evaluation, I agree itâs hard, though I also think a GiveWell evaluator has a hard job. In fact, I think hardness is just the norm here. For example, consider my decision whether to continue working at SecureBio. How likely is it that someone might engineer a pathogen? How much earlier in expectation is one flagged due to our efforts? How much harm is averted via earlier notification? What is my marginal contribution to our efforts? How much is people thinking SB owns this problem crowding out other work in the field?
(I am very happy with my role and not considering leaving; this is just for illustration!)
My impression of how to make progress with identifying the right places to make political donations is similar to what youâd do when assessing other donation opportunities: have people give it their full attention. My impression is they talk to existing people in the space to understand what has worked in past campaigns and how this is changing, look at the research to the extent itâs any good, and talk to candidates and evaluate their public statements.
Expanded this and my earlier comment into a post: https://ââwww.jefftk.com/ââp/ââhigh-dynamic-range-diy-air-testing
Measuring air from in a positive pressure respirator or clean room that is more than three logs cleaner than the room air is doable if youâre thoughtful about experimental setup. Fundamentally, each sensor gives you about three logs of range, you donât have to use the sensor raw. Hereâs an example with four sensors:
You have your outer room, where youâre going to have a very high particle concentration, and you put a sensor here. This sensor will read out of range (too high) for most of the experiment.
In your outer room, you put a fan with a MERV-16 filter, blowing air into a bax (plenum) at very very slight positive pressure, where you put another sensor. This will read within range for your whole experiment if youâre lucky, or perhaps just a second 2â3 of the experiment.
Do the same as (2) with a HEPA filter. This will read within range for your whole experiment if youâre lucky, or perhaps just the first 2â3 of the experiment.
The final sensor goes in the place youâre trying to measure. It will read out of range (too low) at the end of the experiment.
Also put a small air purifier in your outer room, so that the particle count will decrease over time in a smooth way.
Before starting, calibrate your sensors by putting them all the same room and seeing how they handle the same input.
To begin the experiment, you put a huge amount of particles into the air of the outer room, perhaps dried salt. You continuously monitor your four sensors to do math on afterwards. Youâre going to want to do it on count bins, not mass, because that should be more stable over time and the experiment isnât able to measure everything simultaneously.
You should be able to calculate ratios between all of these sensors, either directly or by chaining. What you want to know is the ratio between your initial room (1) and your clean room (4), which you can get chained from (2) and (3).
If you use salt for your particles, the most you can get in the air is probably 100-500 mg/âm3 and the sensors can do ~1-1000 ug/âm3, this gives us 5-6 logs of dynamic range.
You can also get additional sensitivity by running longer, but
I think sensitivity increases with the square root of time, so 10x sensitivity means 100x longer, which is pretty annoying. [EDIT: actually I think itâs linear, so this is a good way to extend your range.]
ComÂing Around To PoliÂtiÂcal Donations
SeÂcureBio DeÂtecÂtion is Hiring SoftÂware Engineers
On alt proteins, if we ever substantially beat price parity (say by 50%), itâs just hard for me to see how we wouldnât get mass consumer adoption.
For an exact substitute like precision fermented egg whites I think I agree. But alt proteins are usually more like the difference between eating different animals (turkey bacon instead of pig bacon) and people often pay >>2x for preferred animals. Even if this gets down to only as different as cuts of meat within a given animal people often pay >>2x for specific cuts. And then people really love variety, so while I see a path to replacing say 2â3 of a typical personâs meat consumption the remaining portion is far harder.
It seems to me very unlikely to be that these big, thoughtful orgs are directionally wrong
That doesnât seem so unlikely to me. There are many patterns that push towards doing things over research: donors prefer it, volunteers prefer it, hard to justify research when that means doing nothing about atrocity today, things that looked good on BOTEC often donât get more investigation as reliance increase, etc. Add on top of this the poor epidemics of the animal welfare movement and I really wouldnât be surprised at all.
Iâve been thinking more about #5, Develop rigorously tested DIY protocols for converting bedrooms into cleanrooms, and Adinâs It May Be Possible to Improvise A High Grade Bioshelter, which both primarily address environment-to-human risks (E2H). It seems to me that the impact case for both of these depends on (a) a limited duration of the threat and (b) electricity.
On (a), my impression is that for E2H we wouldnât expect the pathogen to just go away on its own very quickly, and so weâd need a many people with good equipment working to develop countermeasures. Is that also something you see as a critical component? Is humanity in a good place on this? Are you looking to fund work here?
On (b), it looks like the assumption is that the power grid continues functioning, and so DIY methods built on standard electric fans keep working. I donât know if the thought is that the power grid is so critical that if it goes down we lose anyway? But it seems to me that even if it mostly still works it might be much less reliable in this situation, and any power outage means loss of positive pressure and filtration. Perhaps some combination of fossil fuel generators, portable power supplies, electric cars, bicycle attachments, and repurposing rooftop solar could be worth exploring?
This line of thought also suggests a project (11) which looks like:Figure out what prep it is cost-effective for people to do ahead of time as individuals.
Produce resources that make this as easy as possible (what to buy, what to test ahead of time, what to have locally on paper)
Clearly and persuasively make this case publicly.
Benefit is a combination of some people being directly protected by having prepped better, plus higher capacity supply chains for scaling this up quickly in an emergency.
Weâre in a bit of a pitchforks mood, but I donât think whether Austin and Caroline would have transferred this particular bit of money is a good target.