In Search of the Language Singularity

There’s talk in the air about the Language Singularity (LS), no doubt prompted by recent big-tech advances with LLM-based generative AI services. Not quite the same as the famous technological Singularity mooted back in the 20th century by Mssrs von Neumann, Vinge, and Kurzweil, but certainly inspired by the version suggested recently by Marco Trombetti.

Marco predicted that the time will soon come when measuring the difference in translation quality between human-only and machine-only output becomes irrelevant (in most but not all cases, presumably). Machines will do as well as (trained) humans but at a far greater scale, so we’re about to achieve a language-driven interpretation of the “Singularity.”

A tech and IT encyclopedia (2014) already had an article referring to the LS as “an integration of language and technology capabilities to produce a common, easily accessible form of communication between all humans, immaterial of their home language or language capabilities.” That sounds fairly bland, plausible, and adapted to the world of standard “accessibility” rather than high-quality equivalence.

But this same article goes on to explore the “possible effects of rapid technological development on human interaction, language, and culture in a ‘globalized’ world which has unequal access to literacy and technology.”

That sounds much closer to our typical concerns about the inequality of spoken-tongue access to almost everything humans talk about. Yet it still does not specify what LS would actually change.

In other words, the term LS should theoretically refer to a sort of language event horizon — a boundary line dividing past from future — but expressed as our conception of human communication (spoken/written/signed, etc.) unhindered by language differences because enabled by transformative technology. BTW, don’t confuse that LS with the Singularity Language — a mere software resource!

It’s not yet clear exactly how measurements of translation quality can really be used to underwrite the Language Singularity event, or why translation is the best criterion for it. Perhaps we really need a specific Translation Singularity (TS) first, rather than just a general LS.

Rather than a single example as a proof of concept for the TS, a vast amount of translation activity should be put to the test across all available languages, channels, and content. The evidence that a few translation occasions can be judged by few people as constituting a Singularity moment because, for example, those people happen to know the languages involved, sounds far too weak.

There is also the problem of multiple translation outputs for a given string. Would we have attained the LS/TS when all systems deliver the same “correct” output, as if by magic? Or on the contrary, would a post-LS/TS world necessarily produce a whole range of subtler versions, each one finely tailored to specific features of the translation occasion, just as humans today will produce various near-synonymous outputs for different contexts? To achieve this, the machine would need a “world model” to understand the entire context in which any translation is required.

So here’s the real problem: Either there is a moment of truth when we pivot forever into the Singularity: the planet would echo to a constant flow of automated, quality translations, plus all other machine-driven language performances you can imagine.

Or we follow a step-by-step, possibly exponential, advance in the way we manage our automation of intelligence, language, and related powers generally today. This means ChatGPT and similar tools are the latest attempts to offload our needs to machines to achieve greater efficiency in our affairs. But they will never actually reach the LS state.

How about posting this test prompt to a hypothetical Large Translation Model: “translate this nuclear physics text from German into Dumi, a Tibeto-Burman Kiranti language spoken in the mountains of eastern Nepal.”

This would be an impossible(?) feat for humans today, and almost certainly not worth the effort anyway. But in which specific ways might a post-Singularity AI be able to deliver on this?

Note that it could not translate between existing linguistic knowns by looking up data via an LLM; it would have to invent or borrow and adapt coherent new science terms in Dumi that its 3,500 or so speakers would be able to learn to understand and manipulate. To do this it would be necessary for the AI to include an obvious skill: request aid from humans as there will not be a handy Dumi text resource.

In other words, the TS would have to be more than simply a shift in machine states — it would need to spark a new framework for cooperative operations between humans and machines, a very different enterprise to “more automation”, which includes all kinds of problems of bias, misinformation, and ignorance, and anyway already exists implicitly as a global AI project.

This future TS form of interactive agency would involve transcending the barrier of impossibility raised by mere humans (e.g. “we’ll never get a comprehensive German-Dumi language pairing”) and jointly opening up planet-wide linguistic conversion on a new scale.

The project would cover all extant human languages and their content as the machine becomes smart enough to organize more and more humans to help in the process, before replacing them.

If that is not what is meant by the AGI (artificial general intelligence)-driven LS or TS as currently specified, and I’m sure it raises the usual fears, then there’s no real singularity, no unique moment when a scale+quality leap to a “singular” level of machine capability occurs.

So the LS could mean optimally leveraging rather than replacing human skills as part of its own intelligence resource, along with data, compute power, and software engineering. But it must somehow live up to its name as a unique, value-shift event, not just a proof of gradual advance.

The trouble with all this fancy speculating is that we get perilously close to Nick Bostrom’s famous Paperclip Maximiser parable in which a rogue post-Singularity AGI decides to use the planet’s entire resources to produce nothing but… paperclips.

What if an LS/TS device got delusions of grandeur and decided that global translation needed the entire planet’s compute resources to constantly build bigger LLMs etc. and translate all spoken/written/signed content into all languages all the time so that humans have constant access to everything? There is a considerable risk that it would start gobbling up vast amounts of biological and mineral resources just to provide enough power for a huge, perpetual Translation Maximiser!

The only difference from the paperclip scenario would be that any humans still alive might actually need some of its linguistic outputs to understand what happened!