Rethinking Superintelligence without Language

By Clement POIRET · 3 min read

Language disguises the thought; so that from the external form of the clothes one cannot infer the form of the thought they clothe, because the external form of the clothes is constructed with quite another object than to let the form of the body be recognized. — Ludwig Wittgenstein, Tractatus Logico-Philosophicus.

Despite all the hype generated by big corps trying to rise billions on the idea that “scaling LLMs is all you need” to achieve AGI/ASI/etc. — whatever that means and whatever will be the benefits — let me mind-wander about language being the conveyer of the said intelligence.

Language is a model of the world which evolved because of the need to collaborate. Our complex language system evolved as an information compression mechanism, first transmissible through acoustic waves, then made long-lasting and immutable via written forms. It allowed us to communicate from bits of information to complete cultural identities, enabling knowledge sharing, coordination of actions, and many other evolutionary advantages. But as impressive as this may seem, language is a lossy compression algorithm, as Wittgenstein said: language disguises the thought¹.

Beyond its lossy nature, there exists a language-cognition interplay. Linguistic determinism states that language fully determines thought, meaning we cannot conceive of ideas for which our language lacks words or structures. A slightly weaker version of this theory, linguistic relativity, states that language influences thought, making certain concepts easier or harder to grasp depending on linguistic features.

There are clear examples of this interplay. As an example, speakers of the Australian language Guugu Yimithirr using cardinal, absolute positional systems (similar to our north/south/east/west) and not relative ones (left/right/up/down) ², are “predominantly coding nonverbal memory in “absolute” concepts congruent with their language, while a comparative sample of Dutch speakers do so in “relative” concepts”³.

We start to see that language is a part of our Brain’s Operating System which models the world, but all models are wrong, even if some are useful. Training a large language model is like distilling our own culture, the lens with which we see and interact with the world. To better model the world, we should go beyond the bottleneck that is language, because language is not all we need — to quote Wittgenstein, the limits of my language means the limits of my world.