There are primates with proto-language, which I think let them communicate well enough to do these sorts of things. The question then becomes “why go from a four-grunt language to the full variety of human speech?”, and it seems like runaway dynamics make more sense here (in a way that rhymes with the Deutsch-style “humans developed causal reasoning as part of figuring out how to do ritual-style mimicry better” arguments).
why go from a four-grunt language to the full variety of human speech?
Bandwidth. 4 grunts let you communicate 2 bits of information per grunt n grunts let you communicate log(n) bits per grunt. In addition, without a code or compositional language, that’s the most information you can communicate. Even the simple agents in the OpenAI link were developing a binary code to communicate because 2 bits wasn’t enough:
The first problem we ran into was the agents’ tendency to create a single utterance and intersperse it with spaces to create meaning.
In my model, the marginal utility of extra bandwidth and a more expressive code is large and positive when cooperating. This goes on up to the information processing limits of the brain, at which point further bandwidth is probably less beneficial. I think we don’t talk as fast as Marshal Mathers simply because our brains can’t keep up. Evolution is just following the gradient.
The main reason I don’t think runaway dynamics are a major factor is simply because language is very grounded. Most of our language is dedicated to referencing reality. If language evolved because of a signalling spiral, especially an IQ signalling spiral, I’d expect language to look like a game, something like verbal chess. Sometimes it does look like that, but it’s the exception, not the rule. Social signalling seems to be mediated through other communication mechanisms, such as body-language and tone or things like vibing. In all cases, the actual content of the language is mostly irrelevant and doesn’t need to be the expressive, grounded, and compositional mechanism of language to fulfill it’s purpose.
There are primates with proto-language, which I think let them communicate well enough to do these sorts of things. The question then becomes “why go from a four-grunt language to the full variety of human speech?”, and it seems like runaway dynamics make more sense here (in a way that rhymes with the Deutsch-style “humans developed causal reasoning as part of figuring out how to do ritual-style mimicry better” arguments).
Bandwidth. 4 grunts let you communicate 2 bits of information per grunt n grunts let you communicate log(n) bits per grunt. In addition, without a code or compositional language, that’s the most information you can communicate. Even the simple agents in the OpenAI link were developing a binary code to communicate because 2 bits wasn’t enough:
In my model, the marginal utility of extra bandwidth and a more expressive code is large and positive when cooperating. This goes on up to the information processing limits of the brain, at which point further bandwidth is probably less beneficial. I think we don’t talk as fast as Marshal Mathers simply because our brains can’t keep up. Evolution is just following the gradient.
The main reason I don’t think runaway dynamics are a major factor is simply because language is very grounded. Most of our language is dedicated to referencing reality. If language evolved because of a signalling spiral, especially an IQ signalling spiral, I’d expect language to look like a game, something like verbal chess. Sometimes it does look like that, but it’s the exception, not the rule. Social signalling seems to be mediated through other communication mechanisms, such as body-language and tone or things like vibing. In all cases, the actual content of the language is mostly irrelevant and doesn’t need to be the expressive, grounded, and compositional mechanism of language to fulfill it’s purpose.