Epistemic status: Was revealed to me in a dream. 

More seriously, this is a post (actually, stringing together two posts) I wrote based very loosely on research I did in late 2021. I do not necessarily stand by all the claims in the current post, but hope that it can still be moderately helpful to at least some readers. I think this post has some structural issues. It was not cleaned up sufficiently for my personal standards of publication. It is also in a more conversational style than I endorse, and I’ve grown to be less confident in the core metaphors. Earlier drafts also have critiques that I did not get around to addressing. I also expect it to be more generally false. However in the spirit of draft amnesty day, I am publishing it rather than leave it languishing forever in my Google Drive (and my conscience).

I think among effective animal advocates who are techno-optimists, the arguments for alternatives to factory farming are too narrow in scope, eg, plant-based meat vs cultured meat, or plant-based meat vs cultured meat vs non-food investments (eg corporate campaigns), or occasional considerations of other meat alternatives like yeast or mushrooms. 

I think this is too narrow in scope. I instead think we should have first-principles evaluation of strategies^[1] to sate people's desire for meat while avoiding animal suffering:

1. Carefully considering all the desired criteria of what people value in conventional meat
2. Enumerating all of the existing strategies, including the lesser-known ones
3. Brainstorming/exploring entirely new strategies
4. Carefully analyze the feasibility of each strategy

I believe a significant fraction of this work can be done through armchair thinking by smart generalists, but we will eventually need empirical data, computational modeling, actual experimentation and varied domain expertise. 

In the post, I explore why I believe this is the correct strategy, using the motivating example of an extremely scientifically literate person in the 1800s trying to figure out flight. Birds are an existence proof that heavier-than-air flight is possible, but not a guarantee that we can get heavier-than-air flight quickly, and certainly not a proof that our existing attempts to get heavier-than-air flight is the right approach. 

While I am more optimistic about the general framework than any operational details, I would like to sketch out a path forwards for what a research agenda/plan might look like:

1. Getting private feedback and carefully evaluating the quality of this initial plan, attracting funders and researchers as necessary <- We are here.
2. Have 1-3 people work on the initial scoping of this research for up to 6 months, trying to analyze, dissect, and red-team this argument for clear empirical or conceptual issues
3. Executive search for someone to lead a new org, or a team within an existing org (Rethink Priorities, New Science) to lead “New Meat Studies”
4. Form an initial group of 3-20 tight-knit researchers that aims to generate and quickly falsify different lines of research into viable alternatives to conventional meat
5. When a line of research looks like it cannot be clearly falsified with existing resources, spin off new nonprofits, companies, and academic research directions as necessary.
6. Repeat 4-5 as necessary with new lines of research.
7. Save ALL the farmed animals.
    

Note that there are obvious stopping points on any of #1-#7 if we discover reasons why this plan is not all-things-considered reasonable.

_____

Intro

Suppose you’re in the late 1800s, and you are extremely scientifically literate. You know in exquisite detail about electricity, mechanics, the steam engine, and all the wonders of the industrial revolution. You also know about Darwin and evolution, and (slightly anachronistically) you somehow learned Mendellian genetics too, and have strong opinions about animal breeding.
 

One day, you see some birds, and you think “flying is cool. I want to fly too.” How do you do this? Well, I claim that (in this anachronistic version of) in the 1800s, you have two clear options:

1. Evolve a bird
2. Construct a plane

Evolve a bird

By evolve, I really mean artificial selection. You want to fly, and you already know with certainty that birds can already fly. So you get some pigeons, take the largest from each generation to breed, and select bigger and bigger birds, generation after generation, until one day you have a roc-sized pigeon. You can sit on the roc and ride it to fly and stuff, which is all very cool.

Birds have many advantages over planes (1800s). For starters, they already exist. Another advantage is that you don’t need a deep understanding of aerodynamics or flight or anything, just enough to know what you want and how to make a gradient to get you there. 

They also have many disadvantages. The biggest one is that they have tightly coupled components, and you can't really put together a bird after you take it apart. As a related point, it’s also very hard to have a theoretical understanding of birds. And of course, in the real world, it turns out you’re bottlenecked by generational lengths and mutation rates (which 19th century you might not have known), so getting a roc-sized bird before you die seems unrealistic.

Construct a plane

The other approach is that you take inspiration from birds, but don’t try to replicate them. Instead, you realize that the thing you really want is flight. Birds are just an existence proof that flight is possible, but human flight might take many different forms. So you theorize and experiment with many different designs, some in your head or drawings, some via conversation with other people, some practical designs that you try to manufacture and experiment with out in the real world.

The main disadvantage of a plane is that you don’t know if you can get it to fly. From the existence of birds you know it’s possible to get heavier than air flight, but you have no initial idea if it’s a research program that takes five years or five hundred, in practice.You also don’t know if your specific ideas are any good, maybe you are trying to create a plane but the right strategy looks like a helicopter, or a rocket, or like a gravity-negater, or a really big catapult, or something else you’ve never imagined.

But there are clear advantages too. You are not constrained to the bird template, so you have a wide range of options. You can take apart and put back together planes, so you can improve one component at a time. You can make it simple and composable, so you don’t have to be very smart to work on each part. You can build a theoretical understanding of planes. Feedback loops are fast and you don’t have to wait for new generations.

And in the real world, it looks like constructing a plane was the right choice (for flight). 

Construct a bird

What you absolutely shouldn’t do, and what in foresight is really obviously silly, is to try to construct a bird.^[2] That is, try to get flight by making a mechanical implementation that imitates all the gory details of nature's implementation of birds. This is hard because having a theoretical understanding of nature is hard, and building something that has tightly coupled components is hard, and your search space of available options (constrained by both nature and your mechanical ability) is too narrow, and....

Indeed, I’m struggling to think of any major technological innovation^[3] that fits this profile (comments welcome)

(People have been trying to make mechanical birds for a long time. I believed they finally succeeded in 2011. As far as I know, there are no practical uses).

Meat

Let’s go to the main topic of discussion, which is meat. As far as I understand it, people really like the taste, price point, convenience, and nutrition profile of meat. However, there are 3 common arguments against conventional production of meat:

1. Animal suffering. Many animals are packed tightly in factory farms, and animals probably don’t like being tortured in such farms. 
2. Economics. At least naively, a cow or chicken is a really inefficient factory for turning grains into meat. 
3. Climate change/environment. Industrial factory farming produces a significant amount of greenhouse gasses and expels many pollutants, which is bad.

And indeed, these arguments are used to argue for more funding into alternative proteins, and is a core component of the theory of change for plant*-based^[4] and cultured meat. I understand this argument and respect it, but feel like these arguments are heavily underspecified. I think these are strong arguments that we should have something to replace conventional meat, but they should not anchor us to existing attempts to make alternatives. Let’s consider the existing options:

Cultured meat, aka, constructing a bird

Neil and I encountered many technical issues as we were understanding analyses of how to do scalable cultured meat. You can read a subset of the gory details here. For example, the problem of keeping large bioreactors clean from stray viral particles, the problem of getting exactly the right proportions and sufficiently pure amino acids for the nutrient profile cells need, the problem that cells are unmotivated to grow without a delicate selection of hormones and so forth.

But I think a high-level theoretical cause of all these empirical issues is that cultured meat is targeting the wrong solution to the problem. In brief, they’re trying to construct a bird. 

Cultured meat is saying “look, animals have this thing we want (meat), but also the thing we don’t want (suffering). Let’s produce meat exactly the way animals do it, but minus the suffering.” But the problem here is that nature is by default tightly coupled, and it’s not like the “meat-producing component of nature” is easily separable from the rest of it. So now you have to artificially create all the hard parts^[5] that nature solved through many millions of years in making animals, which is pretty rough. 

Consider the problem of extreme hygiene. Humbird spends 3 pages discussing all the hard issues in maintaining hygiene in animal bioreactors, as any viral or bacterial contamination is fatal. But recast in a different way, the ultimate problem is that we are trying to artificially create an immune system. And as it turns out, the immune system is really hard.^[6]

But that’s not the only issue. In yeast bioreactors, you can just dump glucose and plant material in and call it a day. This is because individual yeast cells have their own digestive systems. In animal cells, you need to separate out the plants into very specific and very pure amino acids, etc, mixes (growth media) first, and then feed the cells carefully. You are artificially creating a digestive system. But you also need the nutrients to get to the cells after you get the right mix, artificially creating a circulatory system, and to pipe all the waste products out, artificially creating a urinary system, and so forth. 

There isn’t a precise analogy to the nervous system, but both the growth factors (hormones etc that incentivize cells to grow) and the general problem of control and combining all these systems together in a proper way de facto means that the human engineers are serving as a rudimentary nervous system.

And you need to do all of this very cheaply. ^[7]

None of these problems are obviously impossible, but solving any one of them looks like it can maybe get you multiple Nobel Prizes. 

And we can be bottlenecked by just one of these problems (and several others that don't have a clear natural analogue), and at the end of the day, are we really sure that this is the easiest way to get to our desired solution of cheap, suffering-free, and environmentally sound meat analogues?

Plant-based meats (a potential method of “constructing a plane”) 

I have a much poorer understanding of the technical issues with plant-based meats. They seem fine to me in a naive read (but I would have said the same thing about cultured meat 4 months ago prior to looking into the issue). But I think both EAs and non-EAs in this space are just way too over-indexed on the existing way of doing things. 

Extending the metaphor, plant-based meats might be the planes of the future. But they might also be hot air balloons, or helicopters, or zeppelins, or something else that might have practical uses but won’t be our primary form of manmade flight (protein).

An intuition pump:

Imagine we’re in 2100, no major transformative technology happened (e.g. no transformative AI), and people are still eating meat-like products. What are your probabilities on:

1. People mostly eating cruelty-based meats?
2. People mostly eating meats produced via a) mashing a bunch of plants together or b) animal cells grown in bioreactors?  
3. Something else? 

If “something else” seems pretty likely to you, I think there’s a plausible case for EAs to try to figure out what that “something else” is (though of course the details depend on tractability, given 2021’s technical prowess and foresight). 

What we should do instead

I think what Effective Animal Advocacy (EAA) and other scientifically literate or otherwise savvy researchers in this or adjacent spaces should consider is 

1. Consider what are the components of meat that people most desire.
   1. I believe there’s some academic research on this, and informally a price/taste/convenience/nutrition/social acceptance model
      1. I think Jacob Peacock (formerly Humane Labs) has some pushback on this model iirc (not sure if I should link article here, as it’s private)
2. Do horizon scanning and/or first principles thinking about what possible technologies can get you there.

I’m more optimistic about the correctness (or at least interestingness) of this high-level process than any specific technique, but I sketch out some ideas below. I believe none of these ideas should be novel to very science/science-fiction-versed EAs, but it might be helpful to have them listed together.

Meat by Constructing a Plane

Here we consider options of selecting on the desirable properties of meat and looking at alternative ways humans can get there without (as much) suffering.

1. Make/popularize something really tasty, that doesn’t try to look like meat
   1. Tofu wasn’t invented as a meat analogue, but it arguably spread in China as a meat replacement (due to Buddhist influences) over 1000 years ago^[8]
      1. At least I don’t think tofu tastes like meat
   2. People don’t (usually) eat potato chips or oreos because they’re vegan
   3. Some (afaik, not most) people prefer the taste of impossible burgers to cow meat because it has more umami
   4. Massive spread of plant-based milk that doesn’t taste like cow milk at all.
   5. So the path forwards is try to come up with a first principles way to make or spread a new product that fills a similar culinary niche to meat, but is better (for at least some people) on grounds other than ethics
      1. Eg cheaper, tastier, healthier etc
   6. One (strong) argument against this working is the efficient market hypothesis, but it seems plausible to me that the food industry is extremely conservative and more into gradient descent than large strategic changes
2. Classical plant-based meats
   1. Discussed above
3. Algae/phytoplankton-based meats?
   1. Apparently it’s a thing but I haven’t heard of it before.
   2. I have no inside view here but might be worth looking into
4. Yeast-based meats
   1. Quorn (single-cell protein, fungi-based) is already a product
   2. Burgeoning mycoprotein industry already
   3. GM-friendly: I don’t have much of an inside view here, but I want to note that we have decades of experience genetically manipulating yeast, and we should a) be rosier on our ability to manipulate yeast than animal cells and b) direct manipulation may offer options that the coarse approaches to plant-based meats don’t.
5. Yeast production of meat ingredients
   1. I don’t know how viable this is, but in theory yeast is good for not only being a meat replacement, but also its ability to synthesize a wide range of possible products
      1. For example, you can synthesize milk with yeast, and also heme
      2. Seems intuitively plausible to me that you can do it for more meat components, though I should check with some scientists first to see how plausible this is.
6. Bacterial single-celled proteins
   1. See forum discussion here. Note that the lower bounds for price are very low. 
7. GMO plants/mushrooms/other fungi
   1. We can figure out what’s missing in existing plant-based meats and try to genetically modify plants to get the desirable properties. 
8. Hybrid GMO plant/yeast/etc options
   1. With modular (ish) components, can try to mix and match options until we get the desired flavor/taste/texture profile.
9. Atomically precise manufacture?
   1. Much more speculative, but at 80 years timescales you might be able to construct analogues of meat via nanobots (or something trending in that direction)
      1. It’s plausible that there’s nothing we can do to prepare for it now, but has anybody actually looked?
      2. It’s plausible to me that either a) timelines to reach saturation for plant-based or cultured meat is longer than this or b) they cap out at <<100% of meat consumption, so you need alternatives like nanotech
   2. There might also be chemical solutions to meat that’s weaker than fully atomically precise manufacture but look more like a chemistry problem than a food science or biomedical problem.
10. [some other examples that I removed because they might be too speculative + sensitive and I’m too lazy to reword carefully]

Meat by Evolving a Bird^[9]

Here we look at how animals produce meat and consider technical ways of getting this result by reducing the suffering in animals. Note that these solutions in the limit (usually) solve the animal suffering problem but not the environmental or economic problems, which I personally view as more minor.

1. Literally ^[10]evolving birds
   1. You can try to selectively breed animals that feel less pain/suffering 
   2. The classic reference is here.
   3. Intuitively highly plausible to me
   4. Not saying this is easy (would need a good theory or empirics on livestock suffering, for starters), but it intuitively doesn’t seem obviously harder than alternative ways to construct meat
   5. Need to be careful about downside risks like industry capture. E.g. birds are bred to chiefly look like they suffer less while suffering just as much or more
2. Genetically engineering animals to feel less suffering
   1. Now that we understand genes (sort of), we have more options for preventing suffering than nature gives us. 
   2. I suspect this is moderately hard because genetic engineering is moderately hard, but nobody said that solving factory farming is an easy problem
3. Chemically making animals happy
   1. If biology is too hard, can we maybe try chemistry? Are there cheap drugs that can (e.g.) pump chickens full of endorphins all the time?
   2. I suspect doing this is easy and doing it cheaply is very difficult/borderline impossible
4. Genetically engineer a de novo animal that has all the properties you want
   1. Haven’t thought about this much, but intuitively if you open the search space to “all possible animals that we have the genes to express” you can plausible get a wide range of properties including “enjoys life in tightly packed conditions, very efficient at turning grains into meat calories, does not have much emissions, etc”
   2. I suspect this is more like “evolving or constructing a bird-plane” since you are doing something that doesn’t quite exist in nature but there’s moderate precedent
      1. Note that “evolving a bird-plane” is metaphorically similar to modern approaches to AI, where you use gradient descent as an outer optimizer to search for a smart algorithm that has some human properties and some inhuman properties, but is not naturally decomposable or understandable to humans
   3. I suspect this is very hard
5. Brain-dead chickens
   1. Try to replace the nervous system in chickens while keeping everything else intact
   2. I don’t think this is easy, and I think if you can figure out how to do this, this is Nobel Prize worthy work, give or take
      1. But OTOH eg cultured meat arguably requires multiple Nobel Prizes
      2. Here you’re “just” replacing the nervous system rather than 4+ other systems.
   3. EDIT: Upon reflection and after going through some critical comments, I’m now updating to believing that this may be very hard to do at scale..

The path forwards

I proposed what I inside-view believe is a very good idea. Below is a series of steps put forward for falsifying this idea, being very open to being shut down at any given step. Note that I’m sketching out the plan for concreteness. I strongly believe reality will look quite different. 

1. Sharing this doc privately with [REDACTED]
2. Have 1-3 researchers (probably not me) from either RP or independent/academic researchers of similar calibre a) research this and b) try to poke holes for 1 - 30 weeks, and seeing if the arguments still hold up to theoretical and empirical evidence
   1. One obvious thing to do is checking how much of this type of work is already done in academia/industry, and how much (if at all) we think we can do better
3. Executive search for someone to either a) lead a new org or b) a research group within an existing org (eg RP or Guzey’s new thing, to conduct “new meat studies”). 
4. Form a starting group of tight-knit researchers (possibly modelled like this) that hopefully has very high epistemic standards^[11], that aims to generate and quickly falsify different lines of research into viable alternatives to conventional meat
   1. I don’t have a good idea of how hard this is. Arguably cultured meat took 20 years overall for proto-EA plus another 2.5 years for someone like Humbird to be pretty skeptical, and we’re still not too sure. On the other hand, arguably Open Phil was skeptical from Day One, and they were right to be. 
   2. Note that at this stage it would be actively counterproductive to get external publicity or funding, it’s much more important to keep things under wraps and figure out true things than optimize for external prestige and thus risking getting high on our own supply
      1. I worry that variations of the latter happen often to do-gooder initiatives
5. If/once a line of research didn’t look like it can be clearly falsified given existing resources, start spinning off an increasingly large project from it, getting EA, corporate, gov’t, and academic support as necessary
6. ???
7. Profit!
8. Repeat 5-8 as necessary, whether in sequence or in parallel
9. Save ALL the (farmed) animals

General outlook

I am very optimistic about this new research paradigm.^[12] I think there’s a high chance this is dumb for reasons I overlooked, but I inside view am much more excited about conceptual progress on this dimension than marginal $s spent on plant-based or cultured meats, and I think this project, in worlds where it isn’t obviously or subtly dumb, has a real chance of being on the causal path to victory for eventually ending factory farming. 

Very happy to receive critiques and be convinced to the contrary, of course! 

Acknowledgements

I thank the many people who have looked over and given wise feedback on earlier drafts of this post, including the (almost as) many people whose critical comments I have not addressed. I may list them by name later, if you're in this set please feel free to contact me to be listed. :)

Appendix A: Birds vs Planes and Evolve vs Construct

I ran out of steam, but if there’s enough interest at some point I should make a Lizka-style 4-quadrant diagram of birds vs planes and construct vs evolve with accompanying explanations.

The basic idea is that when we restrict ourselves to only looking at nature-inspired technological advances (the bird/plane paradigm doesn’t have a good analogy here to dishwashers), our inventions usually look like a combination of modifying birds or designing planes.

Evolving(Modifying) birds

This is where you see a tightly coupled and complex system in nature that does most of what you want, and you decide that you give up on understanding it deeply and you basically want to use it with some tweaks.

Most of modern agriculture (at least the plant and animal breeding parts) in the last 10,000 years or so looks like this. Another example is that the problem of "how do we create a general de novo solution to contagious disease" is hard. The problem of "how do you trick a human immune system to do most of the work for you without overreacting" is still hard, but often less so. And this (the story goes) is why we have vaccines for stuff like smallpox and covid. 

(Notably through most of the history of these innovations, we are able to do this without having a deep understanding of either the immune system or animal genetics)

Constructing (designing) planes

This is where you see a tightly coupled and complex system in nature that does what you want, and you decide to use nature’s existence proof as inspiration for designing something better.

Most nature-inspired inventions look like this. Maybe even most inventions in the last 200 years, I’m not sure.

Examples I can think of include planes, cars, submarines, and computers (early computer scientists knew humans can do math, but did not then decide the best strategy was to deeply understand neuroscience before you can make progress on calculating machines).  

Often you don’t need to deeply understand the natural system, but you do need to deeply understand and have a gears-level conception of the details of the artificial system you’ve created.

Evolving (gradient descenting) planes

This is where you see a tightly coupled and complex system in nature that does what you want, but for whatever reason neither modifying the bird (tweak the existing complex system) nor designing the plane (create an alternative modular system from first principles that has the desired functionality) was effective. 

Perhaps it’s possible to create an artificial evolutionary environment, where setting up the overall optimization environment to point your search towards the desired end point is more feasible than the above strategies.

Savvy readers may note that this is roughly what modern AI looks like. Some deep learning is inspired by neuroscience, sure, and some of its design is tweaked by humans, but fundamentally we understand the training environment much more deeply, and we are not constructing the networks by hand. 

On the other hand, even after they’re evolved, we still do not understand a neural network’s internals. So it’s less like evolving a plane and more like evolving a bird-plane. 

I think there are other (probably less important) problems in CS that look more directly like evolving planes, eg circuit designs. 

More speculatively, economics (and society) sort of looks like this, if you handwave a bit.

The analogy here is less perfect than the above two subsections I think.

Constructing a bird

Finally, we get to constructing a bird. This is where you see a tightly coupled and complex system in nature that does what you want, and you think your best way for getting there is by trying to emulate exactly what nature does, except via artificial/mechanical means.

I’m genuinely struggling to think of an archetypal example here that both a) exists and b) is commercially successful or otherwise important. All the obvious examples that comes to mind (cultured meat, Open Worm), just don’t seem very successful, and it seems like in most cases, either evolving a bird (if constructing an alternative system is hard) or constructing a plane (if modifying the existing systems enough is hard) are better strategies. 

Occasionally, you can neither evolve a bird (enough) or construct a plane, but you should probably look at evolving a plane (or evolving a bird-plane) before you go to constructing a bird. 

I’m not saying nothing remotely alike constructing a bird exists. Non-archetypal examples that I can think of include “easy” things where the relevant parts of the system is easy to understand (artificial fire is just like natural fire, aqueducts are like rivers, dams are like lakes, pointy weapons are maybe like claws(?)), relatively minor things with low economic import (literal mechanical birds, some solar cell designs that copied photosynthesis and are less economically viable than traditional solar cells), or both (“gravity sensors in our phones are inspired by and quite similar to amyloplasts in plants cells” H/T Holly Elmore). 

For the first class of examples, I claim that this is because there are clear reasons that if nature's implementation is simple to understand, we can copy it. But this is clearly disanalogous to flying like birds or growing meat like, well, birds. In the case of the second class, my argument is less strong, but I think it’s not a coincidence that the examples aren’t very economically viable. The theoretical thrust here is that if nature has a complex way of implementing something, it’s unlikely to be optimal to human specifications. So by the time we have the scientific understanding and technological know-how to re-implement a complex solution that nature has figured out, we likely also have the scientific/technological ability to do something better.

This is not a proof. Of course, cultured meat (and for that matter, brain emulations) can be the first major exception to this observed empirical regularity with some theoretical backing. But I wouldn’t bet it against the “field,” so to speak.

1. ^{^}

   To limit the scope of this proposal/research project, I only consider technological solutions. Reality will probably require activism and other forms of social change in addition to the technological framework.

2. ^{^}

    The fourth option in the quadrant is evolving a plane. This feels kind of silly for flight, but should make sense for some other problems, which we’ll explore later. 

3. ^{^}

   Maybe something like spidersilk in bulletproof armor? But even then it feels like you basically took the desirable property from nature (spidersilk) and constructed a plane around creating spidersilk.  Also see the Appendix.

^{^}

“plant-based*” as e.g., Impossible Foods has heme produced from GMO yeast, so is technically not 100% plant-based

^{^}

Well, almost all. For instance, you don’t have to solve motor control, which is certainly a relief.

^{^}

If us amateur epidemiologists learned anything from 2020, it’s that immune systems are highly unpredictable and very hard to tease apart (Ed Yong: “Immunology Is Where Intuition Goes to Die”).

^{^}

An obvious reductio ad absurdum of my argument is that it looks like “you can’t construct a bird” will be an argument against cultured meat existing at all, and also against pharmaceuticals that use animal cells. And clearly, cultured meat exists. But I think it’s obviously easier to do something that looks like imitating nature if you lift a bunch of constraints, eg by using a lot of expensive materials or manpower or a lot of energy. Back to the bird analogy, we did eventually create a mechanical bird in 2011, but it was a) expensive, b) not very useful, and c) much harder to get to than making planes.

^{^}

 After this post was originally written (but long before the current publication), George Stiffman wrote a detailed and well-received post on the case for rare chinese tofus.

^{^}

Note that this section is extremely speculative and easy to screw up.

^{^}

 Here I’m using “literally” as a standin for “figuratively”, which I agree is an unusual usage of the term. 

^{^}

The type of people I’m imagining for the initial team is people who have experience in a combination of FHI-style research, which looks like a lot of digging into conceptually new territory, RP-style hard empirics, and some fair amounts of computational modeling and lab science.

^{^}

UPDATE 2022/12/18: I’ve since cooled down and am only moderately optimistic about this whole thing. This is partially due to generally cooling down about the value of my own work with the benefit of hindsight and time, and partially due to increasing worries about fast TAI timelines and the relative futility of work that hits slower.

Show all footnotes