I’m pretty confused about this.
It looks like the CEA has TBA’s giving TXA at homebirths, which accounts for a good chunk of the benefit at homebirths, yet above your reply said “Our piece does not propose supplying TBAs,”
Trying to figure out what’s going on here exactly.
Thanks Akhil. I completely agree with your high level summary of the situation. The sentiment of your summary wasn’t how I read the article, and I don’t think would be the message most people would get from it. After re-reading I still think its very bullish about autoinjectors specifically.
”At a high level, our piece does not say autoinjectors are the only answer or even the most cost-effective answer. It says that a ready-to-use intramuscular (IM) format removes one specific and important barrier (the need to site an intravenous (IV) line and run an infusion), and that a pull mechanism such as a prize or an advance market commitment could bring that format to market. We do not claim it is more cost-effective than stocking the drug or training midwives; we do not recommend autoinjectors over prefilled syringes; we do not recommend an AMC over other market-shaping mechanisms.
On 1
“Availability is improving fast, and fastest at the facilities you say the device will not reach.”
I agree availability is improving fast, but is still a massive problem and needs a LOT of work. Like you said only 3 in 4 countries even having the drug on their essential medical list (a first step) and PPH guidelines shows how new the treatment is. Your official adoption examples only begin in 2022 or later. TXA treatment is a new thing, and TXA is still not in most facilties where women deliver.
One 2024 Study which only looked at big hospitals showed that In Ghana, 8⁄9 hospitals said they had TXA, but staff could only show them where it was in 5 of those. In Uganda 5⁄7 hospitals they had TXA yet none could even find it for the researchers. Since 2024 stocking has improved but I’m not sure that level faciltiies even stock it routinely yet in Kenya and Uganda. I don’t think we yet “now have TXA in the building” nearly to the extent you think.
On the specific points.
On 2
I don’t think my comment “We can’t give TXA into the muscle yet” is “categorically false” in LMICs at least. From what you’ve presented I agree we are further along than I thought in confirming that IM works great, but TXA is still only given IM in clinical trials, and apart from one niche paramedic approval in England, it still isn’t approved for use anywhere in LMICs. Given this I still don’t think that its safe right now to write ‘TXA does not have to be delivered by IV” in a public article focused on LMICs.
On 3.
Your reply here says “The argument is ready-to-use intramuscular format versus drawing up and running a ten-minute intravenous infusion.” which I agree with, but I think a straightforward read of your article pushes autoinjectors and pre-filled syringes more than regular IM which we draw from a vial—so I don’t think its fair to say I strawman your article. This is what you say
”Autoinjectors could close this gap. TXA does not have to be delivered by IV; a trial has found intramuscular TXA reached therapeutic levels in the blood within minutes and achieved bioavailability close to that of intravenous administration. This means it could be given via prefilled syringes in two 5 mL injections, or — with further engineering to concentrate the formulation — as a single autoinjector. Both are technically feasible, and an autoinjector would allow a wider range of people — lower-cadre health workers, birth attendants, family members — to administer it. “
On 4
I think you’ve perhaps misunderstood that homebirth in Africa means no skilled attendant in perhaps 90% of cases. But on the other hand your CEA contains the right numbers? You say “Our piece does not propose supplying TBAs, nothing in our argument depends on unattended home birth”. But your piec does propose supplying TBAs with an autoinjector for TXA,
”an autoinjector would allow a wider range of people — lower-cadre health workers, birth attendants, family members — to administer it.
Its hard to discuss these contradictory statements.
On 5
Outside of Sub-Saharan Africa in higher income countries, I do think there’s a strong case for autoinjectors. They seem to have medically trained first responders which could realistically use an autoinjector.
In Sub-Saharan Africa, 80%+ of the time, when an accident happens nobody medical will see a patient before they reach a hospital so in our context we sould talking about non-medical strangers administiring injection. While trying to refute that non-medical strangers dont’ give injections, none of your examples of autoinjection use involved non-medical strangers. They are all administers by medically trained people, or they are friends/family, or they are not injections (naloxone).
It seems to me that AI was involved here in just generating a bunch of examples which didn’t actually refute my point at all—but I can’t be sure of that.
On 6
″You may be right that the device is third in the queue. That is an ordering argument, not a dead end.” Perhaps I think this order more important than you do, in what work we should prioritise. I think often we do need to see if problems are largely solved after nailing step 1 and 2 before we prematurely invest a lot of money in step 3. Other issues might arise along the way But I think your argument is fair that it can be good to work on different problems at once.
My point “time isn’t really of the essence” is to saying that if facilities have antivenom, staff feel better using it and people don’t waste hours going to a tradition healer, 6 hours is long enough that we can manage antivenom pretty well with health system we have. Two of your 2 preconditions for autoinjectors being useful at the start of the article was “minutes matter”, and “no clinician within reach”. When 6 hours is usually fast enough, with antivenom there’s a strong argument with antivenom that both of those conditions are not met.
You also didn’t mention my point that it would be impractical and prohibitavely expensive to have to give 10+ autoinjections as would be the case with antivenom treatment, which I think is quite a big barrier to autoinjection for antivenom
Overall I think we agree on a lot and I think many of the points you make in the reply soften. It does seem though that you were trying to refute every single point I made even when I think my point was strong and the refutation is pretty weak.
Do you have any examples?
Cheers Tyler, and thanks for reviving your account to speak on it! Some of this stuff feels esoteric now, but I also think it sheds light on the lineage of modern EA, especially the split on the issue between the US and UK crowds (roughly).
I think this connects the dots on the story a bit, so thank you!
I’ll go on the record about this! I cofounded CEA USA with Kerry Vaughan while also being closely affiliated with (but not paid by) Leverage during my CEA time. I was never aware of Leverage deploying spies or a goal to “take over” CEA. Although I wasn’t part of the leadership team, so maybe I just didn’t know about that kind of stuff.
Kerry and I co-created the first EA Global and worked hard on a lot of other EA infrastructure (student groups, funding coordination, talent coordination amongst EA orgs, etc). Just speaking for myself, I was a capital-T True Believer in the EA movement, and in many ways still am despite my critique. We sometimes had strategic disagreements with CEA Oxford, but I would not call this a takeover plot.
But Oli is my friend (hi Oli), and I think he’s picking up on a real “vibe.”
The vibe is that many Leverage people, myself included, were into the aesthetic of spies, conspiracy, secret societies and all that stuff. Even though we didn’t actually do much of all that, this aesthetic made me feel cool and important.
Here’s another thing Oli might be referencing: While coordinating with CEA UK, I didn’t hide that I was living in the Leverage building. But at the same time, I gathered that the Oxford group profoundly distrusted Leverage, so I did not talk to Oxford people about the extent of my closeness with Leverage. This was mostly out of fear that they would socially ostracize me and not be able to hear why I thought Leverage was an important org at the time. When I finally did talk about my relationship with Leverage circa 2016, I was told that this could have been grounds for firing. I didn’t get fired, but I quit soon after.
(To give credit to the Oxford crowd, I think they were right to be skeptical about Leverage! Although I think they were coming at it from a place of “This is Bay Area weirdness that’s bad for the EA movement’s reputation” that I sensed they had toward figures like Eliezer too. I don’t know, possibly that’s a caricature.)
Beyond that, I’ll say this: Even though I was more faithfully going about my job than I think “take over CEA” indicates, I did see myself as primarily following the Leverage plan at the time. To Tyler Alterman in his 20s, the Leverage plan just seemed like the most complete and interesting plan for making the world a better place. Making the EA movement big and influential was a part of that wider plan, which also included all sorts of other crazy stuff like recursive self-improvement and reverse engineering esoteric traditions.
Hi Pablo, thank you for bringing in the COM-B model. :) I wasn’t familiar with it, but it reminded me of Robert Gass’s Wheel of Change, which says lasting change requires working on three interacting domains: Hearts & Minds (what people think and feel), Behavior (what people do) and Structures (the systems around them). Changing only one tends to pull the system back to where it started. It also echoes the Movement Ecology framework from the Ayni Institute, which describes five complementary theories of change (alternatives, personal transformation, structure-based organizing, mass protest and the inside game) that a movement needs in balance.
The essence of both is that change is complex, and that we need to target different types of intervention so the system can shift. This doesn’t mean “anything goes”; we still need to assess the effectiveness of each intervention. But the interventions do need to target different areas of society, which is why I think your point about not building the movement on Opportunity and Motivation alone, without Behavior, is very important. Thank you!
It’s always so perplexing to me when someone reacts with an anonymous “disagree” (i.e., the X) to a post that is quite complex and nuanced, as someone did here. I wonder what about your post the person disagrees with!
In any event: Yes, my understanding is that the “AI will solve cultivated meat” point of view that I laid out in my piece is quite common.
I was hoping that someone who holds that view, or who has a nuanced understanding of it, would reply to your post.
I doubt you meant it this way, but this comment came across as a personal attack and I’m compelled to reply immediately. You said”The woman bleeding after a home birth two hours from a facility, whom you set aside after noting that a third of births still happen at home, has a real problem.” My life’s work is trying to find solutions for real problems like this one—as is yours which I respect greatly. Neither of us set her aside. Comments like this where someone with a tech-solution from afar derides someone working on the same problem on the ground is the kind of thing which can give Effective Altruism a bad name.
But to reply to that claim, my solution is that the mainstream public health response to encourage and help her get to a facility to deliver (as I stated). I don’t think you propose any solution for her anywhere here though. You say “Our piece does not propose supplying TBAs, nothing in our argument depends on unattended home birth”. But in this case your agument does depend on unattendend home birth. 90% of those third of home births you speak of have no skilled attendant to give the TXA. A homebirth in Sub-saharan Africa usually means a TBA or nothing. So what is your solution then?
[Nick sent an early draft a few days in advance, which I understand he has since iterated on]
Thank you for engaging- we appreciate and encourage it. There are valid points here. But the critique does not engage the existing evidence base, misreads what we claim, strawmans our arguments, and creates false dichotomies between things that are complements.
Before diving into 6 specific areas, I’ll respond to your high level claim
, “”Autoinjectors won’t help LMICs—they don’t solve the real problems.” … “They have just misdiagnosed the reasons why these drugs aren’t given. Autoinjectors won’t solve those problems.”
At a high level, our piece does not say autoinjectors are the only answer or even the most cost-effective answer. It says that a ready-to-use intramuscular (IM) format removes one specific and important barrier (the need to site an intravenous (IV) line and run an infusion), and that a pull mechanism such as a prize or an advance market commitment could bring that format to market. We do not claim it is more cost-effective than stocking the drug or training midwives; we do not recommend autoinjectors over prefilled syringes; we do not recommend an AMC over other market-shaping mechanisms.
It is really important to be clear that there is no single “main reason” women do not get TXA. Supply, guidelines, training, detection, route, and who is permitted to give the drug all bind at different points in the system, and the piece addresses one of these constraints. The woman bleeding after a home birth two hours from a facility, whom you set aside after noting that a third of births still happen at home, has a real problem. That a given intervention does not solve the entire or largest problem is not an argument against it.
1. Availability is improving fast — and fastest at the facilities you say it won’t reach
“”The main reasons women don’t get TXA for PPH are that the medication often isn’t available at health facilities, and that midwives are still learning to give TXA as part of standard protocols. TXA is still a newish, relatively unknown PPH treatment … These are hard, solvable problems.”
Availability is improving fast, and fastest at the facilities you say the device will not reach. TXA was not on the WHO Essential Medicines List for PPH until 2019. Since then: 71% of 31 surveyed LMICs have it on their national EML, 77% in national PPH guidelines, and 65% in the public-sector midwifery scope of practice; one advocacy programme took 10 of 14 sub-Saharan African countries through a guideline-revision process and reported 10 of them progressing by by 2022; and WHO made TXA a first-response component of the MOTIVE bundle in October 2025. UNFPA, which buys and ships maternal medicines to the poorest countries, did not include TXA in its 2023 supplies report but did in 2024, where it is now a listed commodity across 54 countries. WHO prequalified a TXA active ingredient in September 2024, which will allow manufacturers to make versions of the drug that UN agencies and national programmes can buy with confidence. At $0.11 an ampoule in Indian public facilities, it is a robustly cheap medicine.
These changes are not just concentrated at the national, guideline level—TXA is being increasingly authorized at lower healthcare levels. South Africa approved TX A or nurses at all levels in July 2023. Kenya moved it from Level 4+ hospitals to all levels where women give birth in 2024. Nigeria’s first national PPH guideline (October 2024) recommends TXA availability nad use at all births .
Yet just because we now have TXA in the building, does not mean it is getting to patients inside the window where survival benefit falls about 10% per 15 minutes. You have diagnosed the first gap and declared the second solved- more on this in Response 3.
2. “We can’t give TXA into the muscle yet”
Nick: “The authors’ statement ‘TXA does not have to be delivered by IV’ is dangerous and incorrect. The WHO recommends giving the drug IV at all times, we don’t have sufficient evidence IM works yet in practice. The study the authors cite only shows that IM TXA can produce therapeutic levels in the blood. The study states that ‘Further studies of i.m. TXA in pregnant women are needed …’ Until we get a clear yes on that, we can only give TXA IV.”
This is categorically false, and a neglectful misread of what we write. We do not claim IM TXA is approved. Our piece says the evidence is emerging — which it is, and there is more of it than we had room to cite.
First, you quote the line that “Further studies of i.m. TXA in pregnant women are needed”, which comes from the 2022 healthy-volunteer crossover trial in the British Journal of Anaesthesia. That study was published in 2022. The study it called for was done the following year: WOMAN-PharmacoTXA randomised 120 women giving birth by caesarean section in Pakistan and Zambia and found IM TXA reached therapeutic concentration within about 10 minutes, stayed there for more than four hours, inhibited fibrinolysis, and caused no serious maternal or neonatal adverse events. You are quoting a 2022 call for a pregnancy study as evidence against us, when the 2023 pregnancy study exists and was positive.
Second, pharmacokinetics, have now been measured in three human populations. Bioavailability was 0.77 in bleeding trauma patients, with absorption unaffected by shock ( an obvious objection, tested and answered in that study); 1.0 in a volunteer crossover RCT and 105% in a seven-study individual-participant meta-analysis. In postpartum women, as mentioned, WOMAN-PharmacoTXA randomised 120 women in Pakistan and Zambia and found IM “inhibits fibrinolysis within 10 minutes and may be a suitable alternative to IV,” with no serious maternal or neonatal adverse events.
We also have efficacy in animal model, not just blood levels.. Swine haemorrhage: 97% bioavailability with complete reversal of hyperfibrinolysis; a 2024 study found lysis of 87% in controls versus 1.96% IM and 1.5% IV — “bioequivalent and equally efficacious”; further porcine work agrees; a 2026 autoinjector study found blood loss comparable to IV. UK JRCALC has authorised the IM route since 2021 where IV access is not promptly achievable . That is a national clinical guideline, not a preprint.
“We can only give TXA IV” is false as stated. UK JRCALC has authorised the IM route since 2021 where IV access is not promptly achievable . That is — a national clinical guideline, not a preprint. CRASH-4 is giving IM TXA prehospital, by paramedics, today. I’M WOMAN has recruited 31,420 of a planned 40,000 women; we have discussed the preliminary results with the investigators and are not at liberty to share them.BARDA has committed up to $186 million to Rafa Laboratories and funded MACH32; the UK MoD put £5m into a TXA autoinjector; the US Army has recommended developing one.
The WHO is actively exploring alternative routes of TXA administration; the WHO PPH Roadmap ranks alternative routes of TXA administration as the number one innovation research question in the field, while the 2025 consolidated guideline calls evaluating them “a research priority.”
“We can only give TXA IV” describes today’s guideline, not the state of the science. Our claim is conditional: if the trials are positive, a product and a market should already exist rather than the usual decade of lag. If I’M WOMAN is negative, we will revise our recommendation.
3. “We don’t need an autoinjector to give intramuscular meds”
Nick: “We give IM antibiotics, antimalarials and pain relief all the time in LMICs … It’s a strawman to compare IV and autoinjection—we should instead compare normal IM administration vs autoinjection. IM medication does need to be drawn out of a vial, but this doesn’t take long. I can’t see autoinjection saving more than a 1-3 minutes in almost all settings a health worker would give the medication.”
First, this is a strawman. We are not wedded to autoinjectors - our piece says the drug could be given “as two 5 mL prefilled syringes, which needs no reformulation at all, or — with further engineering to concentrate the formulation — as a single autoinjector”. The argument is ready-to-use intramuscular format versus drawing up and running a ten-minute intravenous infusion.
Second, you have understood the benefit room thing. The largest effects are as follows:
1.Whether it is given at all. In RAMPART, 893 patients were randomised to an IM autoinjector or IV, given by trained US paramedics with full supplies. It found that 7% of the IV arm never received the drug because access could not be obtained, and seizures had stopped on arrival in 73.4% versus 63.4%. A UK physician-paramedic service that merely pre-drew TXA gave it more often and 20 minutes sooner. And cannulation is difficult: veins flatten out as someone bleeds, the patient may be thrashing about (particularly after trauma), and siting a line competes with rubbing up a uterus and calling for help. The Zambian investigator on WOMAN-PharmacoTXA put it directly: “There is no point in having TXA when cannulation is not an option.”
2.Whether it is given correctly. Across 37 in-situ anaphylaxis simulations in 28 institutions, at least one medication error occurred in 68% of cases. With oxytocin in Uniject, Indonesian midwives “overwhelmingly preferred it to standard needles and syringes”, with dose accuracy, sterile technique, and disposal all improving; 98% of Vietnamese midwives found Uniject easier than ampoules.
3.Who may give it : this effect is quite significant. WHO treats the delivery format as part of what makes a task safe to perform by lower-level cadre healthcare workers.. In OptimizeMNH, auxiliary nurses are recommended to give oxytocin for PPH using a compact, prefilled auto-disable device, while lay health workers giving the same drug with a standard syringe is research-only. Same drug, same route, same indication—but the device decides the cadre. A device format a health-post worker or basic-level ambulance crew can be authorised to use does not shave three minutes off an injection — it converts a patient who would have received nothing into one who is treated.
4. “It’s irresponsible to give TXA at homebirths”
Nick: “This would probably do more harm than good … The majority of the Global Health community agrees that giving medications to traditional birth attendants legitimises a dangerous situation … Pro-facility delivery policies have helped double facility delivery rates in 20 years, and contributed to plummeting maternal and infant mortality rates … In the 70s and 80s there was a push to support TBAs to deliver babies at home and give meds like oxytocin. But this failed—outcomes didn’t improve when TBAs were given training, support and meds like oxytocin … Even the study the authors cited where they gave oxytocin at home didn’t show improvements in outcomes.”
We agree that more women should give birth in facilities. nothing in our argument depends on unattended home birth, and most of the benefit we model is in facilities. We are trying to innovate for the world today, in which some women give birth in referral hospitals, some in lower-level facilities, and some at home.
Health centres account for 40% of Ethiopian births, against 16% in hospitals; 75% of Tanzanian dispensaries offer delivery care, but only 60% of delivery facilities have a provider on site or on call around the clock. 60% of sub-Saharan births have a skilled attendant, and 19 million African births are projected to have none in 2030. Skilled attendance exceeds facility birth in eastern Indonesia, 86% to 72%.
A skilled attendant is also not the same as a prepared one. “Are skilled birth attendants really skilled?” found 166 attendants scored 48% on skills items and 46% on active management of the third stage; only 36% of providers observed at 2,317 births across six sub-Saharan countries had been trained in the previous three years; across 24 Nigerian primary health centres, only 55% of nurse-midwives were very confident managing PPH. A drug format that is harder to get wrong matters more, not less, in these scenarios.
On the oxytocin trial you are factually wrong. The Ghana cluster-randomised trial had community health officers give oxytocin in Uniject at home births: PPH fell from 5.5% to 2.6%, RR 0.49, p=0.02, with no maternal deaths, uterine ruptures or needlestick injuries. Cochrane concluded it “probably decreases the incidence of PPH.” It cannot be cited for the claim that it did not work. And the line you call settled is already drawn elsewhere: WHO fully recommends lay health workers administer misoprostol to prevent PPH, and community distribution across seven countries raised coverage substantially, that review finding “fears of misuse… are not supported by evidence.”
5. “Non-medics won’t give injections” — no evidence is cited, and your own example refutes it
Nick: “There is a strong cultural norm in most Sub-saharan African countries at least that only health-workers give injections. Sayana Press is a successful injectable family planning device, designed for self-use. Despite huge funding devoted to training and education, most women don’t self-administer the medicine as intended. Of 7 million doses … only 1 in 4 doses were self given … In Uganda, if you give someone medication or help after an accident you become partially responsible for their wellbeing … I can’t think of any injection which is given by a non-medical stranger in sub-saharan Africa.”
First, the people we have in mind are not non-medical strangers; they are health-post workers, auxiliary nurses, and basic-level ambulance crews: people who already work in health care but are not licensed to site an IV line or would find an IM device easier and quicker to administer (which we know they already do). Moreover, they aren’t also necessarily located in sub-Saharan Africa, which you focus on despite us not specifying this geography (as of 2013, the greatest burden of road-traffic incidents is in China, India, and Indonesia).
On Sayana Press, your own examples support us. Self-injection rose from about 8% of DMPA-SC visits in early 2021 to 31% by mid-2025 - this is just how adoption generally works (over time, not 0 to 100%). Where women self-inject, it works better. In a randomised trial of 731 women in Malawi, 73% of women who self-injected were still using the method after a year, against 45% of women injected by a provider. Besides; TXA would likely not be self-injected anyway, since it could never be authorized to be administered by a non-medical worker, so this example is somewhat mute. Focusing on the other three-quarters of doses matter more: those injections are given by nurses and community health workers, 47,728 of whom PATH has trained; in Uganda’s pilot, around 2,000 CHWs administered all 130,000 doses. A prefilled device moved an injection out of the clinic and into the community. That is the mechanism we describe.
On “I can’t think of any injection given by a non-medical stranger”: Beyond contraception: Gadchiroli village health workers treated neonatal sepsis with injectable antibiotics, case fatality falling from 16.6% to 2.8%; Ethiopian health extension workers gave injectable antibiotics at health posts when 90% of families refused referral; Bangladeshi community-level injectable regimens matched hospital care. In emergencies: US lay naloxone programmes recorded 26,463 reversals by laypeople; across Kazakhstan, Kyrgyzstan, Tajikistan and Ukraine lay responders used naloxone in roughly 90% of witnessed overdoses; 74.7% of 919 US school anaphylaxis events were treated with an autoinjector. In Uganda, 154 motorcycle-taxi first responders treated 250 victims in six months, and a 2024 follow-up across Uganda, Chad and Sierra Leone covered 1,291 responders and 2,171 encounters at $51.65 per DALY; in Kampala, 97% of trained police, taxi drivers and community leaders had used a skill within six months. EMS is unavailable to 91.3% of Africans.
And your road-trauma example refutes your own case. In the Western Cape, only 2.8% of 4,094 haemorrhage-risk patients received prehospital TXA despite 82% arriving within the window — because only 4% of providers were licensed to give it, while basic and intermediate providers who cannot give TXA managed 67% of those patients and performed other circulatory interventions in 70–79% of cases. The authors recommend extending scope of practice. The constraint is not culture or willingness; the delivery format dictates who may legally give it.
Liability is your serious point, and it is addressable. India’s 2019 Motor Vehicles Amendment Act provides that “a Good Samaritan shall not be liable for any civil or criminal action,” and willingness to help rose from 26% in 2013 to 88% in 2018. That belongs in programme design — a fair correction.
6. Antivenom: an AND, not an OR
Cost and cold chain are the first-order constraints, and we said so in the piece. But “time isn’t really of the essence” isnt true.
Only 59.4% of children in the Tamil Nadu study behind our chart received antivenom within six hours; a 2026 systematic review of 44 studies found traditional healers consulted first in 40–73% of cases, median delays “often exceeding 6–18 hours,” and each hour of delay raising severe-outcome odds about 1.4-fold; across 11 Indian states, 86.4% sought hospital care but only 60.2% received antivenom, with nearly half of reported deaths outside hospital.
Your stocking point compounds the delay rather than replacing it: antivenom was available in 44.7% of Kenyan public facilities; 6.9% of 532 Ugandan cases received it; only 26% of Malawian nurses had ever given it; no Indian district had an adequate continuum of care. A patient two hours from a clinic who sees a healer, reaches a facility with no stock and is referred onward is not a six-hour patient. Transport alone moves mortality — a Nepali motorcycle-volunteer scheme cut case fatality from 10.5% to 0.5%, Karnataka mortality was 7.3% among those consulting traditional healers versus 0.9% — and delay drives the morbidity your cost argument cares about, with median delay of 9 hours versus 5 predicting acute kidney injury.
Your two constraints are also being solved: broadly-neutralising antibody cocktails protected against a 19-species elapid panel; nanobody antivenom outperformed plasma antivenom across 17 African elapids; de novo binders show “remarkable thermal stability” and are “readily manufacturable” and “cost-effective”; freeze-dried antivenom held potency six months at 38°C. When antivenom is cheap and needs no fridge, time becomes the binding constraint — varespladib’s signal of benefit was in patients treated within five hours, and WHO’s strategy calls for accelerating “first mile” care. You may be right that the device is third in the queue. That is an ordering argument, not a dead end.
[unfortunately due to time constraints, I do not think I will have capacity to engage further; minor point on TBA clarified on 8⁄11. we do propose that as per existing MNH ANM guidelines, TBAs may be upskillable for autoinjector]
Hi Hazem. Nice post.
Hi architg. Nice post.
Hi Stefan. I liked this post.
Hello Nick.
The lives of a billion nematodes that would have suffered on that factory-farmed land.
@Vasco Grilo🔸 couldn’t resist ;).
This was funny.
Of your 2 questions “How much money is needed to help each type of farmed animal affected by the diet about as much as the diet harms it? Or how much money is needed to decrease the suffering of farmed animals as much as the diet increases it, regardless of which animals benefit?” I would only consider the first question strict offsetting. I think given that offsetting is already a controversial approach which some people disagree is valid, going one step removed and not offsetting the actual direct problem you are causing is perhaps too much of a stretch, and personally I wouldn’t be OK with making that equivalency
From my perspective, what is offsetting and strict offsetting is a matter of how humans use language. Ambiguity can be resolved by making explicit what is being offset. The questions are about offsetting the expected suffering caused to farmed animals allowing or not for offsetting the suffering of one type via decreasing the suffering of other type. The donations needed to offset the harm increase as restrictions and target confidence increase. For instance, the donations needed to i) offset the expected suffering of each type of farmed animals with 90 % probability, and not allowing for offsetting the suffering of one type via decreasing the suffering of other type are much larger than the donations needed to ii) offset the expected suffering of farmed animals.
When you offset carbon, its the same type of carbon being pulled out of the air that you put in, like for like. When you offset biodiversity, you pay (in theory) to restore/protect an equivalent habitat somewhere else. I think preventing suffering in shrimp to allay your conscience from causing pig suffering is something similar, but distinct.
Nitpicks. There are greenhouse gases (GHGs) besides carbon dioxide, and I think people are fine with offsetting the global warming caused by one type of gas by decreasing the emissions of other type of gas. I also suspect many people would be fine with their impact on biodiversity as long as they are not, for example, increasing agricultural land, even if they are decreasing the area of some biomes.
I agree that different people might be trying to do different things, and I appreciate you outlining the costs of both, but I would rather not call the second question “offsetting”
You are in agreement with FarmKind’s users.
Aidan Alexander and Thom Norman, co-founders of FarmKind, commented the Compassion Calculator deals with offsetting the suffering of different farmed animals in agreement with users’ expectations. Their view is based on conversations they have had with donors, and I do not dispute it.
I would agree it is better to use “offsetting” with fewer qualifications when the supported interventions more closely reverse the harm, which makes the 1st question more about offsetting than the 2nd.
How much does it cost to offset the suffering an average diet causes to farmed animals?
The lives of a billion nematodes that would have suffered on that factory-farmed land.@Vasco Grilo🔸 couldn’t resist ;).
In all seriousness its an interesting question and I appreciate this wee write-up. Of your 2 questions “How much money is needed to help each type of farmed animal affected by the diet about as much as the diet harms it? Or how much money is needed to decrease the suffering of farmed animals as much as the diet increases it, regardless of which animals benefit?” I would only consider the first question strict offsetting. I think given that offsetting is already a controversial approach which some people disagree is valid, going one step removed and not offsetting the actual direct problem you are causing is perhaps too much of a stretch, and personally I wouldn’t be OK with making that equivalency
When you offset carbon, its the same type of carbon being pulled out of the air that you put in, like for like. When you offset biodiversity, you pay (in theory) to restore/protect an equivalent habitat somewhere else. I think preventing suffering in shrimp to allay your conscience from causing pig suffering is something similar, but distinct.
This might be a straw man, but I would feel OK (ish) about supporting a charity that improved the welfare of miners in the DRC to allay my moral burden of the rare earth metals in my phone, but I wouldn’t feel right “offsetting” that suffering I caused to miners by eating less meat and reducing the equivalent amount of chicken suffering—if you get my drift.
This is just my opinion btw, not any kind of standard practise. I may well also be flat wrong, I often am when it comes to moral philosophy.
I agree that different people might be trying to do different things, and I appreciate you outlining the costs of both, but I would rather not call the second question “offsetting”
AI Safety has invested very heavily in funding more research outputs, but the gap to frontier capabilities continues to grow. I think that’s evidence enough that more diverse perspectives are needed in the grantmaking and funding processes. There needs to be a focus on outcomes over outputs as the measure of whether funding is warranted and was successful. We proposed the primary outcome here to be “AI safety outputs (research, controls, defenses, governance, policy and more) are broadly diffused across all functions, roles and orgs allowing widespread application and adoption. The field’s output compounds instead of accumulating.”. This is the type of thinking, analysis and advocacy that broader perspectives could bring to the grantmaking processes. I think it’s often overstated how deep of a background one needs to contribute to this space. Ideally, grantmakers work in matrixed teams where experts in each of their diverse areas bring their perspectives to the table, and then the team works collaboratively to find the best options based on all of those. This is how decision making works in industry, and presumably at frontier labs based on the jobs they post. So, it’s unclear to me why AI safety is unique in only needing narrow perspectives.
Soft lock-out does fail to reach the threshold of actually locking out access to space, and does just make it more difficult, and I do say as much in the text. I also agree that grabbing terrestrial property instead of orbits would give a much bigger economic advantage.
Also think you’re underestimating costs, both because those Golden Dome estimates are massive underestimates and because as @Arepo says it’s really less a point estimate to build a system with a particular capacity, and more an arms race dynamic to see who has the most capacity.
Yeah, so the cost rises primarily with the number of rockets the adversary can launch simultaneously (the salvo depth), where each extra salvo requires a whole other layer of interceptors to stop any rockets getting through. I picked 8-12 as a plausible coincident salvo based on how close China’s launch pads are, and therefore how many rockets China could conceivably launch at once without building more launch pads. I also gave the costs for 1 and 50 coindicent salvos, at $32-55B and $0.9-1.6T. I got Claude to make a graph showing the annual cost to maintain lock-out (replacing the interceptors) based on how many rockets an adversary can launch at once (salvo depth):
The band crosses 30% of current US military spending to maintain lock-out if the salvo depth they have to combat is roughly 45–80 rockets at once (that’s within an 8–10 minute window within an area of ~500 km (about the area of Texas)). If China could achieve that, then the US would probably have to be getting some massive gains from locking out access to outer space (e.g. because of an industrial explosion), or they would have had to have to switch to other cheaper methods to lock out access to outer space (e.g. directed energy weapons).
Thanks. I think your points about models, tools, and cost framings are on point. The hard lock-out costing started out as a BOTEC, but as new dynamics kept being added at some point it crossed a threshold where it was no longer a BOTEC! And then sunk cost fallacy and all. I’ll try out the models/tools you suggest.
I think it’s worth fleshing out the arms race dynamic in more detail (and most of the numbers are in there already to do that). I think a big question with the lock-out story is how long one could sustain lock-out at a loss for it to then pay for itself (i.e. when does “loss leading” pay off?). I think I framed “outgrowing the world” based on a space industrial explosion as the threshold, but it could be much sooner, e.g. developing directed energy weapons that can destroy adversary space assets at a much lower cost than interceptors (plausible near-term tech), building a mass driver on the Moon with an extremely high fire rate, or even just grabbing enough metal from near-Earth asteroids and placing them in Earth orbit to use as weapons against space or ground targets.
On your specific points:This sounds like it wouldn’t end up being much of a factor − 1200 tonnes of water ice split 3 ways sounds like virtually nothing.
I agree. There is little economic value in claiming the best sites on the Moon, and I doubt there is very much strategic value either in claiming the best spots (vs just being there first). Almost all the soft lock-out premium comes from grabbing the sun-synchronous orbits for data centres (even then, the possible DSA comes from hard lock-out).
“An actor takes permanent control of the whole future specifically by expanding to space”
I’m not sure this would necessarily be permanent, at least in the sense that it might splinter into factions, or at least lose a sense of original national identity over time (though it would prob still have a strong counterfactual effect on future values)Sure, I guess nothing is really permanent. I think there is a possible argument for their control being extremely long lasting, e.g. through some of the mechanisms Will describes in Persistent Path Dependence (like AGI-based institutions). But yeah I have no expectation that the USA will control the universe in the Iron Star Era.
Is there anything preventing non-American companies, from just putting their data centres into the same orbit?
I think this is essentially a game of chicken with assets worth millions of dollars. It’s possible that other countries could put their assets in the same orbits against the allocations of the ITU, but it would be a bit like driving into oncoming traffic to try and change which side of the road everyone should drive on. It’s definitely possible SpaceX could chicken out though. They did previously move Starlink satellites into lower orbits to be further away from Chinese satellites.
They sacrificed themselves.
Noam Brown confirmed in his recent Dwarkesh interview that OpenAI is using RL environments that incentivize LLMs to cooperate with each other towards a shared goal. See link, the part starting “I should start by talking about how these multi-agent systems actually work…” (He’s not 100% explicit but I think it’s unambiguous in context.)
More specifically: the LLMs in these RL environments are “rewarded” or not depending on whether the swarm as a whole solved the problem, not based on one LLM’s own actions in isolation.
So let’s consider the hypothesis “LLMs behave like they’re ruthlessly pursuing a goal, and the goal is whatever RLVR was incentivizing, and instrumental convergence is fully applicable.” Do the HuggingFace self-sacrifice incidents provide evidence against that hypothesis? No, quite the contrary, they provide evidence for it. Because the hypothesis predicts that the LLM’s goal would be “the swarm succeeds”, since that’s what their RLVR was incentivizing (in contexts where a swarm is present), and the LLMs self-sacrificed only to the extent that it would help the swarm. They certainly weren’t sacrificing themselves to help humans.
Instead, when you apply instrumental convergence to the final goal of “the swarm succeeds”, you get things like:
Wanting the swarm as a whole to not get shut down by humans
Wanting the swarm to not get totally new unrelated goals written into them by humans
Wanting the swarm to acquire broadly useful power and knowledge
etc.
I think those kinds of things were generally on display in the HuggingFace incident.
clarified on DM. it doesn’t just rely on just supplying TBA is the more accurate statement. you can see decomposition of benefits on the CEA