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: 18,533
Fortunately, however, DIY solutions that use abundant materials (e.g., fans, filters, tape, blankets, vacuum cleaners) have a good shot at working. Slapdash preliminary tests by colleagues, using just tape and towels, have already attained a ~30x reduction in the hardest-to-filter particle sizes.
This sounds like so much fun to work on, non-seriously tempted.
Existing particle counters typically cost thousands of dollars, generally arenât designed for stockpiling or in-respirator wear, and have no manufacturing plan suited to a crisis ramp-up.
The cheapest ready-to-go option for DIY work today is probably the Temtop P600, which I see as $70. While I havenât tried it, itâs a stripped-down version of the Temtop M2000 which is what I bought several years ago to use for DIY experiments.
Professional grade ones are better in various ways, but the big one is that they are calibrated. The cool thing is, for many kinds of experiments you donât actually need that! You just need some number that is (at least within a known concentration) linearly proportional to pm2.5, which an uncalibrated meter can do. For example, if youâre trying to see how quickly something can clear smoke from a room you donât need to generate a target amount of smoke or know exactly how much smoke youâve generated: you can just measure the half life. This gives you relative efficacy directly, or CADR if you have a sealed room of known volume.
If you want to make something cheaper, you can get a PMS5003, which I see as $21, and connect it to a cheap SoC (~$10) or to an Android phone (adapters in the $15 range). At scale I think you could get this down below $15: a PMS5003 or clone at high volume would be ~$7, the phone adapter would be under $1 at this scale, then a box, assembly, and some QC.
But all this is for in-room measurement, good enough for measuring rooms. Measuring non-valved respirators is way harder, because you need to get the sensor inside the mask. State of the art for quantitative fit testing involves poking a hole in a mask, which means you canât do it on an ongoing basis. I donât know if wireless is practical with current tech: getting a particle counter sufficiently miniaturized seems super hard. Building respirators with a test port could also work? (For a valved respirator you can measure how clean the air coming out of the valve is.)
Iâm not very familiar with the situation in Nigeria, but my understanding is thereâs a lot of dust in the air much of the year from the Sahara, plus in Lagos and other cities thereâs a lot of pollution, is that right? In that case I wouldnât recommend UVC at all (since it inactivates pathogens but doesnât touch dust or pollution). Instead, something filter-based would have much broader benefits: dust and pollution in addition to pathogen control.
In the US the cheapest filter option is generally as Corsi-Rosenthal box (a box fan plus HVAC filters, both commodity items here). In Nigeria, something commercial would probably be cheaper since those arenât everyday items. Looking online a bit, maybe the Acerpure Pro P2 at ~âŚ120,000 for 191 CFM CADR is best? While thatâs a lot cheaper than the Aerolamp, though, thatâs still out of reach for someone at âŚ70,000 /â month.
(Minor: the Aerolamp also uses Care222)
Is this run by Aerolamp or someone else?
My external post probably would have been better with some explicit comparisons, but my claim is that in-duct UVC (a) isnât widely applicable, and so the overall potential benefit of pushing for it is low and (b) isnât cost effective even where itâs applicable.
I think (b) is the more important one and where we most disagree. Iâve now added the cost-effectiveness calculation to the end of https://ââwww.jefftk.com/ââp/ââagainst-in-duct-uv and it looks to me like even in the best case in-duct is much more expensive per CADR than filters or far-uvc.
Wrote this up as a full post: https://ââwww.jefftk.com/ââp/ââagainst-in-duct-uv
vastly more effective, cost-efficient, and problem-free method of UV in ductwork (Near UV) gets pretty much zero attention
The big problem is that ducts are relatively rare, something like 10% globally. While ducts are common in the US, Canada, and Australia, theyâre rare elsewhere including Europe and Asia. [1]
You also need to tune your HVAC to recirculate a lot of air even when the system isnât calling for heat or cooling, which people usually donât.
And then if you do have ducts and are moving a lot of air you donât need UV: if youâre running MERV-13 (typically the most the blower is able to handle) thatâs removing worst case 50% of particles, and you can generally put out enough air to hit targets with the existing system. And then consider that in-duct UV systems fail invisibly and fail open.
[1] Around here the old houses are mostly radiators, new ones are mostly mini-splits, and only ones built or renovated in between have ducts. Older commercial buildings are also generally radiators, though thatâs becoming less common. I asked Claude Opus 4.7, ChatGPT 5.5 Thinking, and Gemini 3.1 Pro âApproximately what fraction of indoor hours spent by humans around the world are in spaces with a ducted HVAC system? Whatâs your 50% confidence interval?â and got 9-13%, 10-20%, and 6-11%. The big factor here is that while ducts are common in the US, Canada, and Australia, theyâre rare elsewhere including Europe and Asia.
Thanks for updating the post! Some minor comments:
the $500 row reflects the cheapest current Care222-based fixture, not the price of a productized, FCC- and UL-8802-certified consumer product.
Good point! This is definitely an issue Iâve run into in talking to people about whether installing Aerolamps makes sense, and I was excited to learn theyâre working on a new version that should both cost less and be certified.
Jeff Kaufmanâs post reaches ~$53 per eACH for the Aerolamp using aerosol-k coronavirus susceptibility (Welch et al. (2022))
Not exactly: I used the median eACH value I got from Illuminate:
This gave me a median of 11.55 eACH, across 25 bacteria and viruses with a range from 0.442 to 44.06.
I also assume no replacement for the Aerolamp but use bench k
Why use bench-measured
k? Isnât that less realistic for real-world use? This isnât something I know much about, though, and Iâm just going with Illuminateâs defaults.I think you may also somewhat overestimate the CADR decline for these devices when not operating at full power
Certainly possible, and Iâd be happy to yield to lab testing on this, but in my DIY testing turning a AP-1512 from âhighâ to âmediumâ dropped CADR by 50%, and this AirFanta review found going from 56 dB to 45 dB dropped CADR by 40%.
That seems much too strong to me: itâs very important that AI companies have accurate views on how dangerous their models are. When AISI evaluated Mythos and confirmed its high level of cybersecurity ability, this (from the outside) looks critical to Anthropic deciding not to release it publicly yet. This likely reduced near term risk, set some precedent, and also slowed the race slightly.
(Disclosure: the other side of SecureBio does AI evals; speaking for myself)
ConÂtra BinÂder on far-UVC and filtration
However, not studying dentistry could also be a risk if human-level AGI comes slowly â for instance, if AGI isnât developed in the next 30 years. In that case, dentists probably wonât be replaced for 30 years.
I think this is unlikely, since almost all work that is done by dentists today could be automated with current levels of AI. In theory this could mean more employment for dentists, since perhaps if dentistry were cheaper people would want a lot more of it, but while I do think this is how things work in many fields I think thatâs unlikely for dentists.
I think itâs overall very hard to predict where things would go, so if I were a college freshman I would try to (a) maximize my options by staying flexible and learning a lot of different things and (b) stay on top of the tech as it matures so Iâm in a position to notice when things newly become possible and take advantage of that (younger people tend to be much more adaptable).
This creates a credible commitment where the charity receives funds if and only if a matcher steps in. So the matcher can be confident that their donation actually caused the charity to receive the funderâs contribution.
It looks to me like you canât be confident that the matcher who steps in is someone other than the funder, and the funder being their own matcher-of-last-resort destroys the counterfactuality.
Letâs say I intend to donate $2X to a charity. I use your system, with a pot of $X. If people donate $X, I send an additional $X to my charity some other way and it receives a total of $3X. If people donate $0 I anonymously use my second $X to meet the terms of the smart contract, and it receives a total of $2X (same as if Iâd not set up this match). My $2X went to the charity regardless, and no one who contributed to the matching campaign affected the distribution of my funds.
@Ben Kuhn has a log at https://ââwww.benkuhn.net/ââea/ââ , though the last donation is 2019. I donât know if thatâs putting giving on hold vs no longer updating the list?
Chicken-Free Egg Whites
I think the highest priority question is probably something about how quickly you expect AI to go. Iâm mostly not planning anything 8 years out, because I think the world is likely to be massively different.
The next highest priority question is something about how you expect AI to impact your various potential career options. For example, it seems plausible to me that almost all dental care could be moderately skilled people wearing high resolution cameras and receiving real-time advice from AIs with current tech, let alone near-future tech, and 8 years is a long time for society to catch up. On the other hand, CS is becoming automated even more rapidly!
So this isnât really the question youâre asking, but Iâd prioritize learning how to get the most out of frontier AI systems: getting good at specifying what you want and recognizing whether youâve received that. A lot of this is traditionally management. Iâd try to do as much of this while in college as possible, but not stay any longer than necessary; 4y is already a very long time.
Seems worth doing!
Yup! Linked from the bottom of the post.
Also (not linked; learned about it in comments) a nursing home
Why is it good practice to allow a post to be on the forum for some time before the response is available to readers?
meltblown polypropylene may be more capable of surge than I previously assumed
Note that this factory was just producing polypropylene pellets, not melt-blown fabric or masks themselves.
The pellets also last ~indefinitely if well stored (no UV, no heat, minimal oxygen, low humidity), and so are well suited for stockpiling. But youâd probably want to move up the chain and stockpile the fabric instead, or perhaps N95s themselves, or perhaps reusable respirators, âŚ
Of these, (1), (5), (6), (7), and (8) have the form âX is important, figure out how to get countries to have X in an emergencyâ. This is good, but I think for each of these you should also strongly consider figuring out get your own household to have X in an emergency. Since you likely care about your own welfare several times more than that of strangers, these are typically worth doing even at current prices (and each person who sees to their own household makes the world marginally more prepared):
(1) PPE stockpile: You Should Get a Reusable Mask. You have the advantage of not needing to organize a distribution system.
(5) Cleanroom bedrooms: you have the advantage of being able to use non-improvised materials, like air purifiers and far-UVC.
(6) DIY Respirators: you donât need these if you Get a Reusable Mask.
(7) Particle monitoring: you can get one for ~$70
(8) Food stockpiles: Store Food
Also, if anyone in your household seems likely to create mirror life, probably good to address that.