A few years ago I posted about the CETI project; now Matteo Wong at the Atlantic interviews Ross Andersen about the current situation (archived):
Ross Andersen: Before attempting to translate the sperm whales’ clicks, Project CETI wants to use a model that is analogous to ChatGPT to generate sequences of them that are hopefully like those that the whales use. Putting aside the fraught and contested question as to whether ChatGPT understands human language, we know that it is quite good at predicting how our words unfold into sentences. Presumably a language model for sperm whales could do the same with the clicks.
It’s possible that these generated sequences alone would tell us something about the structure of sperm whale language. But to translate between languages, models currently need some preexisting translation samples to get going. Project CETI is hoping that they might be able to patch together the first of these required translation samples the old-fashioned way. They have been landing drones on the whales to gently suction-cup sensors onto their skin. The time-matched data that they send back helps the team attribute clicks to individual animals. It also tells them important information about what they’re doing.
The hope is that, with enough observation, they might be able to figure out what a few of the click sequences mean. They could then turn over a crude and spotty Rosetta Stone to a language model and have it fill out at least a little bit of the rest. The scientists would then check whatever it came up with against their observations, and repeat the process, iteratively, until they’ve translated the whales’ entire language.
Wong: What is it about recent advances in AI and translation software that makes this effort not, well, ridiculous? If talking to sperm whales indeed becomes possible, is there anything we can’t talk to?
Andersen: Large language models (LLMs) are getting better at translating between languages with fewer and fewer translation samples, but no one has yet done it with no samples. There is some theoretical work that suggests it could be possible, but again, this work is purely theoretical, and it assumes, among other things, that all languages share fundamental structures that can map onto each other.
As for whether it will ever work, I can’t tell you that. But for the past few years, I have been writing a book about the scientists who are searching for civilizations among the stars, and I can tell you that they are very interested in this research. If it were successful, communicating with an extraterrestrial intelligence—assuming we ever find one—would be much easier. It would also change the strategy you’d use to communicate. Instead of trying to craft a perfect, bespoke message, both sides would want to just send as much data as possible, just like you do with a large language model.
In our case, we’d probably want to send Wikipedia, along with a basic key that labeled a few common features that they would recognize—galaxy, star, planet—to get the model started. Not only would it give our interlocutors the ability to translate our language; it would give them a good sense of what our civilization is like, for better or worse. Sending that much data across cosmic distances is its own challenge, of course, but that’s a story for another time. […]
Wong: There are so many nonvocal aspects of communication—human touch and facial expression, how sperm whales roll and gather into formations. How important do you think that will be? How might we try to translate those methods of communication, if doing so is at all possible?
Andersen: We don’t really know the extent to which body language and touch play into sperm whale communication. On the one hand, we know that they’re very tactile, especially from the way that they roll together, so it’s sensible to assume that touch plays some role in how they relate to each other. On the other hand, they use sound to communicate across distances that are well beyond their range of vision, which suggests that they are able to interpret clicks without seeing any corresponding gestures or movements.
I suspect that Project CETI’s scientists would prefer to see the whales using as much body language as possible. It would give them more of those all-important contextual clues that would help us go some ways toward assembling our first spotty whale–human Rosetta Stone. […]
Early on in my reporting for this story, I called up Dan Harris, a philosopher of language at Hunter College, and he and I ended up talking a lot about whether human language and sperm whale language contain sufficient conceptual overlap to have something like a real conversation. One of the most fascinating things about a dialogue like this would be doing the slow and careful work of feeling out the conceptual edges of each other’s minds, to find out how much of a whale’s experience of the world is inexpressible in human language, and vice versa.
These animals may devote entire vocabularies to sensations that are wholly alien to us. For instance, you could imagine them being quite descriptive about all the different ways that sound waves ricochet off a fleeing squid’s contracting tentacles. As visual animals, we are very interested in representing colors; they may not be. And when you move beyond sensory stuff into higher-order concepts, it gets even trickier. Given that they don’t use tools, for instance, it would be surprising if they had a clicking sequence for “technology.”
I continue to be fascinated by these attempts while remaining deeply skeptical that anything much will come of them. Thanks, Eric!
Making the huge assumption that whales even have anything which can be called “language” in anything but a highly metaphorical sense, it shows a remarkable lack of imagination to suppose that we would be able to communicate with them even if we did know their language.
All together now: “If a lion could speak …”
we’d probably want to send Wikipedia
You fool! You’ve DOOMED US ALL!
Wikipedia is in the running for the most representative and least flattering artefact of our times, along with the Trump Phenomenon. It would be interpreted as the product of a failed civilization easy to subdue. I suppose that’s what you meant.
… and I promised on that comment thread to look at the technical details! believe it or not, that tab is still open somewhere here on the right =>
the fraught and contested question as to whether ChatGPT understands human language
to be a little intemperate: what the fuck is “contested” about that? who besides the most transparent kinds of grifters (and their dupes and enablers) is arguing that Large Plagiarism Models understand anything? it’s “fraught” because shmucks like this keep on pumping the investment bubble instead of having the slightest shred of integrity – as thinking people capable of assessing whether water is wet, let alone as journalists whose fucking job is supposed to be assessing sources.
a person can make an argument for that kind of software being useful for a task, based on what it actually is. if someone’s starting point is this kind of evasive-at-best garbage, though, any argument they make is fundamentally suspect.
Yup.
My greatest misgiving is that they’ll try and use AI to actually communicate with whales. What might that do to 1. Whale language 2. Whale behaviour and 3. Whales themselves. Will whales just shrug it off or descend into existential madness?
@rozele,
You are to be commended for the gentle civility with which you describe the mountebanks and
bunco artists.
@rozele: seconded and thirded. If anything, your civility is too “gentle”/not “intemperate” enough.
“There is some theoretical work that suggests it could be possible…” – frustrating to read without links.
“…but again, this work is purely theoretical,” – human babies learn language with rich context and possibly inborn predisposition. Practically. I wonder if this theoretical work is about learning without context (from a bit sequence) or about LLMs.
Any such links would lead to bollocks anyway.
It is (as you imply) not only theoretically possible but has regularly actually happened that human beings can learn a language without any translation samples. Not only do babies do this routinely, but field linguists have occasionally managed it too (Everett with Pirahã, for example.)
They can achieve this because they share a “form of life” (in Wittgensteinian terms) with the people whose language they are learning: not only a context, but an entire way of living. (The absence of any such commonality is what W is getting at with the “lion” thing.)
Now this is exactly what Automated Plagiarism Engines cannot do at all. They have no point of contact with external reality at all, let alone a “form of life.” They have no way of discovering what an utterance refers to. They can’t mean.
All that leaves you with is the handwaving about “all languages share fundamental structures that can map onto each other.” Except in the trivial and entirely unhelpful sense that all signalling systems whatever must logically share some features, this is a dubious proposition even for known human languages. The notion that it must apply to the “languages” even of other species is unadulterated fantasy.
It also takes as already a given that whale communication actually is a language (in the relevant) sense at all. But in reality, far from being something that can be taken as a premise, that it pretty much what actually needs to be discovered by further investigation, before any talk of translation can even make sense.
Buried under all this crap there is an assumption that whales are basically humans who look funny, have quaint exotic customs and worldviews, and live in the sea, so learning to communicate with them is the same kind of task as learning to communicate with a very exotic group of human beings. This shows a woeful lack of any ability to imagine that the way of life of a whale is utterly different from our own. It’s a sort of Disney concept of whales.
I suspect that the whole enterprise is being presented by the journalists in a way which makes it seem much stupider than it really is by all this drivel about “AI.” The stuff attributed to Dan Harris seems much more interesting and coherent: if communication with whales is possible, it will indeed first be necessary to have some handle on whether our potential conceptual worlds actually overlap – at all.
It puts me in mind a bit of Nagel’s celebrated “What’s it like to be a bat?”
https://www.sas.upenn.edu/~cavitch/pdf-library/Nagel_Bat.pdf
though the point of that paper is something rather different.
Harris himself seems to be a sensible enough sort (and clearly knows his Wittgenstein):
https://danielwharris.com/research.html
who besides the most transparent kinds of grifters (and their dupes and enablers) is arguing that Large Plagiarism Models understand anything?
Many of the machine learning engineers who understand these models best…
Here is Ilya Sutskever, Chief Scientist at OpenAI, explaining what is going on beneath the surface when these models learn to predict the next token.
I wonder if the “Chief Scientist at OpenAI” might perhaps have a pecuniary interest (so to speak) in pushing (and indeed, believing) this line?
https://quoteinvestigator.com/2017/11/30/salary/
Not all experts in the technology are in compete agreement with him …
https://osf.io/preprints/psyarxiv/4cbuv
@DE, yes, I chose not to elaborate on “context” but this context includes similar brains and many things besides what we usually include in our notion of a “sign” with its two sides..
@Seong explaining what is going on beneath the surface when these models learn to predict the next token.
What is going on beneath the surface when I say “the cow is purple” is I’m looking at (a picture of) a cow. I’m not predicting what’s a likely word to follow “the cow is …”. I see no evidence these AIs are capable of looking at the world and verifying the ‘truth’ (let alone the ‘meaning’) of the words they’re about to predict.
Since we humans have no inkling what a whale is ‘looking at’ (or listening to or whatever) when it makes its clicks, I don’t see we can configure AI to observe patterns (of what?), let alone predict them.
So(?) This is like AI predicting what a car is going to find in its headlights?
I believe whale ‘songs’ last of the order of hours; that wikip says they vary month-to-month and year-to-year. Is that what these AIs are going to predict?
BTW, from the abstract to the psyarxiv paper @DE links to (made me do a double-take)
‘deteriorate’ as transitive? That’s the historical usage (1640s, from Late Latin), but I’ve only ever seen/heard it used intransitively. I’d say ‘degrading’.
And ‘remedied’ rather than ‘remediated’.
I suspect for several of the authors, English is not their first language. Or was the abstract written by AI?
I had a Catalan colleague who said of another colleague “she is nice, but she wants to be adorated”. Since then, I have had a soft spot for the Latin causative. Also the shorter form is not fixed in English, acclimatise and orientate (Br.?) / acclimate and orient (Am.?).
We don’t even need to go as far afield as whales, lions, or bats to find seemingly unbridgeable differences in life experiences between individual minds. The last time Nagel came up in discussion here, I came to the tentative conclusion that there are human beings whose internal subjective experiences are, in some areas, unrecognizably different from my own. Discussions of what grammaticality means internally have yielded similar results. And yesterday, I was reminded of another instance that does not involve any metacognition, just sensory perception. At a meeting of the departmental committee that is elected to evaluate our peers’ teaching, we were discussing what a colleague had said about the interference patterns one hears while tuning a stringed instrument. I pointed out that not everyone hears interference beats this way. In particular, I do not. I can tune strings by ear, and I understand the physics of the beat phenomena, but beats (that slow down as two pitches approach one-another) is simply not what I perceive.
I presume that most people are familiar with the story of the Essex. But the part that comes to mind is that, just before the attack, the ship’s carpenter had been engaged in prolonged hammering while repairing one of the ship’s boats.
I wonder what it was that he said?
What if we manage to talk to whales and find out they are lying?
Latin doesn’t have a causative; what do you mean?
Very much orientated BrE vs. oriented AmE.
@dm
elaborated = from work [CAUSED]
adorated = to mouth [CAUSED]
That’s not Latin, that’s English from Latin past participles.
I have no idea why, but English has borrowed a lot of Latin past participles as present-tense verbs. Delete is another.
That is what I meant- it is an English adaptation (maybe reading the past particle as some kind of clitic form of agere), in parallel with forms like handmade .
Unfortunately “causative” is a rather fuzzy notion, along several axes.
One such axis is that in a lot of languages there is a difference between deriving verbs from adjectives or state verbs that mean “get someone/something into the state expressed by the adjective/state verb”, and deriving verbs from dynamic verbs to mean “get someone/something to do that.” Kusaal (of course) is one: it generally uses the derivational suffix -g for the first type, and -s or -l for the second.
In Africanist works (at any rate) the former group are often called “factitive”: they are the transitives corresponding to intransitive “inchoatives.” Often you get “ambitransitive” verbs that can be used either way: English likes doing this a lot, e.g. “warm (up.)” (So does Kusaal: tʋlig “warm something up, get warm.” French, on the other hand, makes a consistent syntactic distinction.)
Most of the English “-ate” type belong to the “factitive” group, and are based on Latin first-conjugation verbs derived from nominals (or are similar analogical formations created within English itself.)
Some first-conjugation verbs come straight from PIE factitives in *-ah₂-: Hittite newaḫḫi “I make new, I renovate“…
@DM, PP: Those English verbs may be formed in analogy to the Latin verbal intensiva in -(i)tare, which often look like they are formed based on the past participle.
I have no idea why, but English has borrowed a lot of Latin past participles as present-tense verbs.
Here is a brief serviceable account from the OED1 (freely available on archive.org), middle column on p. 532, under -ate, verbal formative. I wonder if this has been changed in the OED3. I am away from my OED subscription at the moment.
That entry hasn’t been revised.
@de, dm, hans, xerib
Thanks, I did not know the term factitive or the history of the -ate suffix. I am curious about some doublets, e.g., reply/ replicate (also employ/imply/implicate) and (en-dis)able / habilitate. You can say that implicate and habilitate are learned borrowings for jargon, and that the borrowing of the shorter words is through French, but there reply is répliquer (maybe there was an earlier réplier? or the French had réplier for refold so made répliquer for reply?).
English reply is from Anglo-Norman replier (no acute accent); French still has replier, but only in the sense ‘fold (again)’. The sense ‘answer’ developed in Anglo-Norman in the 1200s-1300s, but not continental French. This is why the Anglo-Norman Dictionary is essential to the OED! (OED3 s.v. reply; AND2 s.v. replier)
French répliquer is a borrowing from a sense that had developed in legal Latin, according to TLFI: Empr. au b. lat. des juristes replicare « répondre à ce qui a été répondu; reprendre un argument, répéter une explication »
Replik: “elevated response, reply; juridical response by the prosecutor to the defendant’s defense; exact copy of a work of art by the original artist”.
Spanish has the same doublet as French: patrimonial replegar (‘to fold again’; more common in the reflexive replegarse ‘to withdraw’) vs Latinising replicar (‘to respond; to replicate’)