Thanks for asking, here are some ideas I personally wish funders would consider at least investigating. The epistemic status of some of these ideas is not great, and I never attempted any robust analysis on the expected values of these potential interventions/causes, but I hope they are worth investigating.
Stop/slow down the development of remotely controllable insects (not just for AW reasons; such a tech can be misused for surveillance, military uses, and bioterrorism)
Stop the spread of caged broiler systems, particularly in African countries, and other LMICs.
Help the most (probably) farmed fish on earth (in terms of number of individuals, counting from actual quantities sold), pond loaches, who suffer from mortality rates like 20-80%, stressful and long transports, and long, excruciating slaughters.
FWI is investigating potential interventions, both on the on-farm condition, and slaughter
Using “Virtual Control Groups” to reduce the number of animals used in the control groups in preclinical studies and basic research
think about it, the control group should be similar across different experiments using the same strain of animal. But animals in the control arm also suffer, because they have to be captivated in small spaces, eat boring food, fed “vehicle chemicals” meant to carry the drug in the test arm, and also have their blood drawn. So the idea is: why don’t we reuse historical control group data (it’s more complicated than that, but that’s the spirit)
Building an LLM-based agent that can allow regulators and drug discovery researchers to have an easy-to-use AI agent with a text-based chat interface (credit to Alexandra Hammond, who developed the idea with me). This idea specifically tackles the problem that many regulators, and some researchers, are reluctant to use ML/DL models that can predict toxicities of chemicals to replace animal tests because there are too many different models, with each of them working for only a narrow range of chemicals. This idea also partly solves the “blackbox concern” excuse used by some regulators to not adopt AI alternatives to animal tests. This idea might be achievable from two different architectures, or the combination of them
Allow the LLM-based agent to access and utiliize various ML/DL models that can be found online, open-sourced, and either teach researchers and regulators how to apply them to certain kinds of chemicals, or even just give the prediction directly in the chat.
Train a multi-modal LLM, with one of its modes being toxicity data.
It seems to me that many alt-protein companies, particularly cultivated meat ones, are interested in using AI to take their research to a new level. But since each of these companies data are proprietary, none of them can get their models to become a large model. What if we fund an open-source large database for alt proteins which everyone can use to train models to help their work. Or maybe we even just fund both the database and the training of a “large-altpro-model”.
AI to help wild animal welfare. A particular intervention that might serve many values, such as signalling effect, signposting, experimentation, and piloting, is to use AI-controlled drones that can identify animals who are in serious injuries or diseases so bad that it means they will not survive for long (and will die from starvation, dehydration, being eaten alive, or multiple organ failures). The drone would have the capability to euthanize the animal on the spot, preferably with methods that would not leave a trace (e.g. if a chemical was used, it has to be non-toxic to at least predators and scavengers nearby). The design is to try to make sure there is as little counterfactual impact on the situation, other than the alleviation of suffering by killing the animal sooner
Note: My original idea was to even limit the use case to animals with broken spines or multiple broken limbs so that counterfactually, not even the location of the death would differ.
Supporting the research of using AI-controlled drones, preferably of insect sizes, for investigating cruelty in factory farms
Of course, don’t use remote-controllable insects.
The idea could be expanded to investigating any animal cruelty cases
Controversial and risky, socially, politically, and legally
Thanks for the nudge! I have something really important coming up in July, possibly the most important thing I might do so far in my career. I will consider after that. Feel free to nudge me again in August!
Thanks Fai. The most farmed fish on earth is likely feeder fish for mandarin fish, numbered in the trillions (~1B mandarin fish x fed 2-4k feeder fish each). We’re working on it at Myrias.
Yes feeder fish for mandarin fish is a big category. But my understanding is that many species’ fry are used, including pond loaches (rarely though, I believe). I am not sure the majority of them are one species (i.e. mud carp).
Also, since we need to count fry to come up with a few trillion figure for feeder fish for mandarin fish, we also need to count pond loach numbers by the fry stage. Estimates of pond loach survival rates from fry to sellable fish vary widely, from 2% to 10%. Given that the number of pond loach slaughtered each year is roughly 10B, that’s ~100B-500B pond loaches slaughtered each year.
Yeah, maybe mud carp or some other species is no.1, but I am guessing it is also possible pond loach is still no.1 or close.
Yeah, from what I heard it’s mostly mud carps (土鲮), which are likely in the trillions. FWIW, pond loaches contain many different species too. Anyhow, both very numerous and neglected, deserving of some serious effort.
Yes, thanks for the reminder. I have long (incorrectly) thought pond loach is just one species, until Ryan pointed out that there are at least 4 (but seems like only two are commercially popular).
From what I learned, even though mud carp should be the biggest used fry for mandarin fish feed, many other species such as other carps and tilapia are also used in significant amounts. But in terms of cause priortization/conceptualization, grouping them together makes perfect sense!
Thanks, Fai. I like your ideas. Virtual Control Groups and LLM-agents are especially interesting to me right now. I want to look into the state of digital twins for various animals. This could not only obviate some animal testing, but also facilitate better translational medicine between humans, farmed animals, companion animals, and wild animals. It might also help us model levels of suffering and welfare improvement associated with interventions like novel pesticides, without the need for physical experimentation. Models would probably mostly cover individual physiology, but could also model population dynamics on farms or in the wild. Aware of anything on these fronts?
Also, your comment on pond loaches reminded me of our ~2021 discussions around animals in the long term future in space. I am planning on revisiting some of those topics.
Thanks for asking, here are some ideas I personally wish funders would consider at least investigating. The epistemic status of some of these ideas is not great, and I never attempted any robust analysis on the expected values of these potential interventions/causes, but I hope they are worth investigating.
Stop/slow down the development of remotely controllable insects (not just for AW reasons; such a tech can be misused for surveillance, military uses, and bioterrorism)
Stop the spread of caged broiler systems, particularly in African countries, and other LMICs.
Help the most (probably) farmed fish on earth (in terms of number of individuals, counting from actual quantities sold), pond loaches, who suffer from mortality rates like 20-80%, stressful and long transports, and long, excruciating slaughters.
FWI is investigating potential interventions, both on the on-farm condition, and slaughter
Using “Virtual Control Groups” to reduce the number of animals used in the control groups in preclinical studies and basic research
think about it, the control group should be similar across different experiments using the same strain of animal. But animals in the control arm also suffer, because they have to be captivated in small spaces, eat boring food, fed “vehicle chemicals” meant to carry the drug in the test arm, and also have their blood drawn. So the idea is: why don’t we reuse historical control group data (it’s more complicated than that, but that’s the spirit)
Building an LLM-based agent that can allow regulators and drug discovery researchers to have an easy-to-use AI agent with a text-based chat interface (credit to Alexandra Hammond, who developed the idea with me). This idea specifically tackles the problem that many regulators, and some researchers, are reluctant to use ML/DL models that can predict toxicities of chemicals to replace animal tests because there are too many different models, with each of them working for only a narrow range of chemicals. This idea also partly solves the “blackbox concern” excuse used by some regulators to not adopt AI alternatives to animal tests. This idea might be achievable from two different architectures, or the combination of them
Allow the LLM-based agent to access and utiliize various ML/DL models that can be found online, open-sourced, and either teach researchers and regulators how to apply them to certain kinds of chemicals, or even just give the prediction directly in the chat.
Train a multi-modal LLM, with one of its modes being toxicity data.
It seems to me that many alt-protein companies, particularly cultivated meat ones, are interested in using AI to take their research to a new level. But since each of these companies data are proprietary, none of them can get their models to become a large model. What if we fund an open-source large database for alt proteins which everyone can use to train models to help their work. Or maybe we even just fund both the database and the training of a “large-altpro-model”.
AI to help wild animal welfare. A particular intervention that might serve many values, such as signalling effect, signposting, experimentation, and piloting, is to use AI-controlled drones that can identify animals who are in serious injuries or diseases so bad that it means they will not survive for long (and will die from starvation, dehydration, being eaten alive, or multiple organ failures). The drone would have the capability to euthanize the animal on the spot, preferably with methods that would not leave a trace (e.g. if a chemical was used, it has to be non-toxic to at least predators and scavengers nearby). The design is to try to make sure there is as little counterfactual impact on the situation, other than the alleviation of suffering by killing the animal sooner
Note: My original idea was to even limit the use case to animals with broken spines or multiple broken limbs so that counterfactually, not even the location of the death would differ.
Supporting the research of using AI-controlled drones, preferably of insect sizes, for investigating cruelty in factory farms
Of course, don’t use remote-controllable insects.
The idea could be expanded to investigating any animal cruelty cases
Controversial and risky, socially, politically, and legally
Nudge to spend another hour on this and turn it into a post?
Thanks for the nudge! I have something really important coming up in July, possibly the most important thing I might do so far in my career. I will consider after that. Feel free to nudge me again in August!
Wow very cool, best of luck!
Thanks Fai. The most farmed fish on earth is likely feeder fish for mandarin fish, numbered in the trillions (~1B mandarin fish x fed 2-4k feeder fish each). We’re working on it at Myrias.
Yes feeder fish for mandarin fish is a big category. But my understanding is that many species’ fry are used, including pond loaches (rarely though, I believe). I am not sure the majority of them are one species (i.e. mud carp).
Also, since we need to count fry to come up with a few trillion figure for feeder fish for mandarin fish, we also need to count pond loach numbers by the fry stage. Estimates of pond loach survival rates from fry to sellable fish vary widely, from 2% to 10%. Given that the number of pond loach slaughtered each year is roughly 10B, that’s ~100B-500B pond loaches slaughtered each year.
Yeah, maybe mud carp or some other species is no.1, but I am guessing it is also possible pond loach is still no.1 or close.
Yeah, from what I heard it’s mostly mud carps (土鲮), which are likely in the trillions. FWIW, pond loaches contain many different species too. Anyhow, both very numerous and neglected, deserving of some serious effort.
Yes, thanks for the reminder. I have long (incorrectly) thought pond loach is just one species, until Ryan pointed out that there are at least 4 (but seems like only two are commercially popular).
From what I learned, even though mud carp should be the biggest used fry for mandarin fish feed, many other species such as other carps and tilapia are also used in significant amounts. But in terms of cause priortization/conceptualization, grouping them together makes perfect sense!
Thanks, Fai. I like your ideas. Virtual Control Groups and LLM-agents are especially interesting to me right now. I want to look into the state of digital twins for various animals. This could not only obviate some animal testing, but also facilitate better translational medicine between humans, farmed animals, companion animals, and wild animals. It might also help us model levels of suffering and welfare improvement associated with interventions like novel pesticides, without the need for physical experimentation. Models would probably mostly cover individual physiology, but could also model population dynamics on farms or in the wild. Aware of anything on these fronts?
Also, your comment on pond loaches reminded me of our ~2021 discussions around animals in the long term future in space. I am planning on revisiting some of those topics.