Follow up to: My previous posts on orca intelligence.

TLDR: I now think it’s <1% likely that average orcas are >=+6std intelligent.

(I now think the relevant question is rather whether orcas might be >=+4std intelligent, since that might be enough for superhuman wisdom and thinking techniques to accumulate through generations, but I think it’s only 2% probable. (Still decently likely that they are near human level smart though.))

1. Insight: Think about societies instead of individuals

I previously thought of +7std orcas like having +7std potential but growing up in a hunter-gatherer-like environment where the potential isn’t significantly realized and they don’t end up that good at abstract reasoning. I imagined them as being untrained and not knowing much. I still think that a +7std human who grew up in a hunter-gatherer society wouldn’t be all that awesome at learning math and science as an adult (though maybe still decently good).

But I think that’s the wrong way to think about orcas. If the average orca was +6std intelligent, it would be a +6std society, where cultural evolution would happen on ultrasteroids, knowledge and skill accumulates quickly and gets taught effectively, so orcas would end up extremely competent. I think the equilibrium for +6std societies is very likely that they act sorta like a macroagent pursuing the multi-agent optimum, aka sorta like dath ilan^[1], even if they don’t have science or a growing economy. The amount of expertise that can be transmitted by smart people is pretty large even without writing.

When I realized that, I stopped thinking about “what would I expect to see if there existed +7std orca individuals” (where I implicitly imagined +0std societies) and shifted to thinking “what would I expect to see from small +6std societies without writing and science”, where in my opinion, the latter implies an extremely much higher level of competence, though I’m probably bad at communicating why.

I previously thought that it’s not clear that orcas could catch more prey if they were smarter. (I was mistaken as I realized today, but that’s for later.) Now I stopped thinking just about food but also about other dimensions, and I remembered that the SRKWs (southern resident killer whales) are a reasonably small and reasonably closed population. I asked claude whether they have low genetic variance and yes they seem to have. Thus, this a divergence from the multi-agent optimum: They ought to interbreed more with other populations. Individual +7std orcas couldn’t realistically figure out why low interbreeding with other populations would be bad, but +6std societies totally should I think.^[2]

They also ought to be able to demonstrate their intelligence to humans, though it’s plausible that they judge this as undesirable and therefore don’t, but it’s still evidence.

More generally, thinking in terms of +6std societies made me think that somehow I would expect to see something more impressive from wild orcas, even though it’s hard to say what exactly. But they don’t appear to be that smart.

Similar things are true for +4std orcas, thought to a lesser extent. I also don’t expect orcas to be +4std.

2. Evidence: Northern resident orcas seem better at catching some fish species than southern resident orcas

Basically, I checked today why SRKWs (souther resident killer whales) mostly eat chinook salmon, and the reason seems to be that they are not that good at catching other fish/salmon. The SRKWs are somewhat undernourished. NRKWs still eat a lot of chinook, but also a lot of chum salmon and some more fish species.

Claude also mentioned that other salmon species, like chum, in the SRKW territory are less depleted than chinook, and that research suggests that it might help the SRKW food situation if the SRKWs could adapt their hunting strategies to also better catch other fish. (Some salmon species may just be harder to catch, but it at least seems probable that NRKWs are better at catching chum salmon.)

Thus it seems likely that SRKWs could capture more fish if they could execute the techniques of the NRKWs.

It's possible that habitat factors make catching other fish easier in northern territory^[3], and that they cannot go into the northern territories and learn from the NRKWs how to catch fish there because the northern orcas don't want them there because that would be additional competition for catching fish. though this itself sorta seems like suboptimal coordination.

Though my main guess is that better techniques would at least help slightly.

This would imply:

1. Orcas cannot quickly reinvent advanced hunting techniques.
2. Orcas aren’t good enough at coordination to share relevant techniques with other populations.^[4]
3. The equilibrium of SRKW hunting techniques before human interference was basically “only be very good at hunting chinook” rather than “be able to hunt all fish species around”. This suggests a learning bottleneck.

(Tbc, this is still in the range of stupidity of humans. orcas could still be human level smart or maybe +1std on average, but it seems significant evidence against them being significantly smarter.)

3. Ryan Greenblatt’s Fermi Estimate

Shortly after I had the two updates above, Ryan Greenblatt posted his Fermi Estimate, which also updated me a bit further.

How could I have thought that faster?

1. Insight on intelligence of societies vs individuals:
   1. Check whether the world you’re imagining is actually a stable equilibrium.
2. Evidence on suboptimal hunting technique sharing:
   1. Properly imagine what you would do if you were an orca. (Don’t hold back parts of your intelligence because you think orcas couldn’t realistically know some stuff.)
   2. Look at a concrete situation orcas are in and see what you would do.
      1. Yes their problems are to a large extent about getting enough food, and what do I know about hunting techniques? But apply full creative problem solving instead of just focusing on the hunting technique part.
   3. -> generalization: When predicting what other smart minds (like AIs) might do, actually apply your full intelligence by imagining myself in their situation.
3. Ryan’s estimate
   1. I think a key mistake here was that I assumed the difference in intelligence between chimps and humans is much much larger than differences within the human variation. While it still seems hard to assess, the difference is perhaps less huge than I thought.
      1. I didn’t try to come up with a measure of intelligence that transcends species, since IQ mostly just makes sense within a species.
      2. I also just thought that within a species, other factors are more important than brain size / number of cortical neurons, and it didn’t occur to me to try to extrapolate within human correlation between brain size and intelligence to how smart other species might be.
   2. Lesson: Try to map out my models precisely with numbers, even if it feels hard to put numbers on it (like how many standard deviations chimps are below humans). Try to formalize your intuitions and thereby bring assumptions out in the open.
      1. Also think whether there is more available data you can use.
      2. Ask “What would you expect to be different in the world where X vs not X”, and don’t just think in terms of causal downstream consequences but also in terms of underlying laws that govern and what those might imply in other places.
4. Other stuff I could’ve done better:
   1. First evaluate the feasibility of your project in more detail before investing significant resources in it.
      1. (Note: I initially thought it might be cheap to test, so maybe wasn’t all that bad.)
   2. Check potential legal obstacles earlier.
   3. Be more pessimistic about how quickly projects will go.
   4. Maybe other people had better intuitions that we would be supposed to see something more impressive from orcas if they actually were that smart, and maybe I could’ve convinced them to patiently discuss our disagreements and find cruxes and maybe their underlying reasons would’ve crystallized into something more useful than their stated arguments.
      1. (Note: I don’t really expect that would’ve helped much, or maybe only with people who’s time is extremely valuable,  though I will perhaps try next time.)

It’s sorta suspicious that I only realized those now, after I officially dropped the project^[5]. It suggests I might’ve been biased towards thinking orcas are very smart, but I don’t quite see concretely how I was. I now started training my introspection, and will train myself to notice when my thoughts seem less than maximally truth seeking and have some other motivations attached to them. But during the project I wasn’t good enough to notice any such directionality, and I don’t know yet what to tell my past self so it would actually do better generally in this way, rather than just benefitting from hindsight.

(I still need to plan how to train myself to not make similar mistakes in the future (e.g. thinking through multiple example scenarios of how failure and success might look like), but I will do that early April.)

Implications for the orca language project

I guess don’t bother doing this for x-risk reduction.

It’s still plausible to me that teaching orcas language would work, though it might take a lot longer than if they had been super smart. But it would be interesting. The experiment I proposed still would be interesting in order to test how feasible that might be.

1. ^{^}

   To clarify, they don’t end up there because of altruism, it's because they set up institutions and contracts so you end up at the pareto frontier, even if everyone is selfish. E.g. in the prisoner's dilemma, two selfish parties still have it in their own interest to sign a contract like "i will cooperate if and only if the other person also signs this contract". You might of course need institutions for what happens when the contract is broken. Orcas could invent their own decentralized crypto currency and contracts that when an orca breaches a contract a significant amount of money will be transferred to the orcas who were damaged. Or sth like that.

2. ^{^}

   Aka, I don't think computers and being able to do molecular biology experiments are that important for figuring out interbreeding. Nor writing if you have a decent memory. Yeah humans are bad at deriving correct theories without overwhelming experimental evidence, though even some humans like Einstein were able to.

3. ^{^}

   From claude:
   1. Foraging behaviors and techniques NRKWs have been observed using different hunting strategies than SRKWs in some contexts. For example, when Fraser River salmon migrate, NRKWs often intercept them in more northern waters where the salmon are more concentrated in narrower channels. This allows for more efficient hunting compared to the more dispersed salmon populations in the broader Salish Sea where SRKWs typically hunt.

   2. Habitat differences NRKWs range extends further north along the British Columbia coast into waters with different bathymetry (underwater topography). These areas include more narrow channels, inlets, and passages that may facilitate certain hunting techniques like corralling fish against shorelines or into bays.

4. ^{^}

   Though not sure about that one. It’s possible the northern orcas just act rationally and have nothing to gain from teaching the southern orcas even if they trade fairly, because perhaps both populations indirectly feed upon the same fish populations.

5. ^{^}

   Not super suspicious, since I already quit at the end of January and postponed writing the post until now, and since I find my other research more exciting anyway, but still. I think maybe I just got better at noticing when sth isn’t actually an equilibrium, and the second insight perhaps followed because I started imagining something actually smart. But still pretty suspicious.