Published on March 15, 2025 9:39 PM GMT
Context: I think there’s a ~17% chance that average orcas are >=+6std intelligent.
TLDR: I have an experiment proposal for testing how quickly orcas might be able to learn a well-designed language and for getting a better model of how smart they are. I think making that experiment happen has very high expected value, although not quite high enough for me to do it myself. Thus I’m trying to find people (perhaps you!) to do the work for making this experiment happen. There is some available funding for doing this.
Motivation for the project
Alignment might be too hard for current humans to solve.[1]
Assume orcas were on average >=+6std smart, and they were interested in learning to communicate with us. Then:
It’s reasonably likely that some intelligent people (like me) could figure out a way to teach them a well-constructed language. And then we could teach them more about science and rationality and AI and mentor them to become superscientists which then might be able to solve the alignment problem.[2]
Furthermore, this could happen faster than superbabies, because we might just need a few years to learn to deeply communicate with orcas, and can then start teaching young orcas right away, rather than needing to wait like 15 years until there are superbabies at the age of 5.
Thus, the “teach orcas” project would be one of the most important projects for reducing existential risk, possibly even surpassing the importance of the “superbabies” project.
The project, for which I’m trying to find someone by writing this post, isn’t the whole “teach orcas” project though, but rather an initial experiment for testing how smart orcas might be and whether we can learn to communicate.
This project is extremely neglected, since normal people don’t seriously consider whether orcas might be that smart. Demonstrating that advanced communication is possible would unlock a lot more resources. Thus, in the world where subsequently the “teach orcas” project is successfully executed, this initial experiment would get a significant share of the impact credit.
Since it doesn’t take much resources to do this experiment, it is extremely cost effective.
Also, consider that this line of hope is very uncorrelated to whether other agendas work out, which makes it more effective. Most alignment agendas are only useful in worlds where alignment is low-medium difficulty, and each making only a marginal contribution that might help us survive in a very slightly harder low-medium difficulty world.
What is the experiment?
It’s trying to teach a few grammatical features, but not nearly everything a language needs. The main variable the experiment is testing is how long the orcas take to learn that. It’s not itself going to be conclusive evidence on whether the “teach orcas” project is worth doing, but it will be evidence, and the results from the experiment may yield some resources to make further testing possible.
What needs doing?
I’m not sure how to best continue, and not even whether it’s better to attempt to do the experiment with captive or with wild orcas somewhere, though I have guesses.
If you’re potentially interested in doing this project, please reach out to me, so I can communicate my model to you.
The work that has to be done is basically:
- Researching which places seem best to try to do the experiment there, and what people might be useful to reach out to.Reaching out to people with relevant experience or in relevant positions for finding valuable collaborators or supporters.Writing research proposals and/or permit applications.Doing the experiment. (I can probably do this one (or find someone else) if you think you cannot.)
So basically the project requires some knowledge-research, strategic planning, and communication abilities.
How long would it take until one can do the experiment?: I don’t know, but my median guess is like 1.5-2 years with an average effort of 10-20h/week, though it also depends on how competent you are at the tasks. Could be significantly faster though, but I’m mostly very pessimistic about captive facilities or government departments allowing such an experiment, and I guess one has to persevere and try hard instead of just the straightforward path, and even then it might often not work and one has to try multiple possibilities.
Available resources
One person reached out to me and said they were willing to spend $40k (and possibly more) on making this experiment happen, but: 1. It was for an earlier proposal of the experiment that was more extensive, so it might be less for this version. 2. They said they might pay less if the project ends up costing more, since then it would be less cost effective.
I know one neurobiology professor who is supporting the project, and he can send your emails from his name to increase the chance that people will reply.
Q&A
Q: Why would wild orcas initially want to communicate with us?
A: Because humans are likely interesting from the perspective of orcas. I think the worry is rather that their curiosity is overruled by their caution, since humans were sometimes capturing and killing orcas in the past. I guess it may depend on where you try it. I’ve seen at least some scenes in documentaries where orcas seemed interested in humans[3].
Q: The initial experiment is quite basic. Would we really be able to teach a full language even if that experiment works great?
A: Yes, the initial experiment doesn’t contain the hardest parts of grammar to learn (which are propositional attitude clauses btw). It is possible they would fail at those harder parts, and further testing is required. However, grammar will be simpler in a well-designed language than it is in English, and it’s not that extremely much harder. Though of course it’s also possible that they can learn the grammar in simple sentences but have more difficulties learning advanced concepts or keeping track of a lot in their mental context.
Q: Assuming orcas are that smart, would we really be able to teach a full language? It seems potentially quite hard and much work to figure out how to communicate many concepts.
Tbh, my answer is that I think that I could but that probably only quite few people could (in a reasonable timeframe), but not totally sure. Sorry I don’t currently know how to make that legible.
Q: So would you take on the “teach orcas” project if it seems like orcas are indeed that smart?
A: Not necessarily, but in case someone does the organizational work for the “teach orcas” project pretty well so the bottleneck is devising the plan for teaching orcas, I would very likely be willing to do work on devising the plan.
Q: Assuming you manage to teach orcas a full language, how do you plan to teach them science? Manually translate all the learning resources?
A: Manually is possible, though I think automatic translation is a feasible problem to solve, even though AFAIK it’s currently not possible to robustly translate to languages with very little training data text. The constructed language would have the advantage that it’s very easy to parse. That alone wouldn’t be enough, but I guess it’s feasible, though I could be wrong.
Q: Would you try to turn wild or captive orcas into superscientists?
A: Wild orcas.
Q: How do you plan to overcome the laws against “disturbing” wild marine mammals?
A: Initially persevere to get more permits for doing research on orca language capabilities, until there’s clear evidence that orcas are human-level smart and ought to be treated as people. And then persevere to make it legal somewhere to teach orcas if the orcas chooses to participate. Yes this could be a bottleneck.
Q: And how do you imagine the physical setting? A facility at the sea where orcas can come and learn?
A: Yeah probably. Though I guess maybe we would also build “orca smartphones”, which don’t have a screen but just voice control and voice output, and which can be strapped or stuck onto the skin of an orca. So they could listen to books and take notes and call with people while travelling.
Aside on what I’ve tried
The only existing facility with orcas that seemed to me reasonably open to behavioral scientific research was/is Marineland Antibes, which closed on Jan 5th (2025). I reached out to them and managed to talk to the head of the science department there, but ultimately they said no because they were already over-stressed with the park’s closing (and my proposal at the time was worse and less clear).
Why don’t I do the project myself?
Because I think I’m one of the smartest young supergeniuses, and I’m working on things that I think are even more useful in expectation, and which almost nobody except me can do.
Please reach out
If you’re considering whether to do this, please reach out.
You might think you’re not a good fit because you have other useful expertise this project wouldn’t draw on. Please consider that most people who are able to take the possibility that orcas are significantly smarter than humans might as well.
If you’d be willing to contribute but not do it alone, please still reach out.
You can book a call with me here.
- ^
In more detail: The scientific method cannot be applied well to robustly solve alignment, since we are not able to test stuff past some critical level of AI intelligence, since if we get an alignment failure past that point, everyone dies, so we cannot as usually learn from our mistake and try again. Thus for solving alignment robustly, we need to form correct deep models about AI cognition without being able to test them (except in very indirect ways). WIth extremely rare exceptions like Einstein, humans are very bad at forming correct deep theories in the absence of a lot of empirical evidence. Thus we need smarter people in order to solve alignment robustly enough that humanity can thrive even if alignment turns out to be hard.
- ^
Yes I know it sounds crazy, but it wouldn’t shock me if it actually happens. Also as context, I’d guess orca values might be naturally much more aligned to human values, and we can likely trade easier with orcas than with AIs.
- ^
E.g. dropping a half-eaten Manta Ray before the human, presumably as gift.
Discuss
