Journey down a rabbit hole with me🐇🕳

Large language models, including those in the GPT family, are best understood not as tools but as language simulators

2. My add: pretty soon we'll have fully-fledged text-to-world models that allow the user to create virtual worlds simply by describing them

in this way the language worlds that LLMs simulate will become spatial worlds that we can inhabit

3. What new dynamics will these LLM-created virtual worlds possess?

One intriguing finding from the emerging discipline that is semiotic physics (the study of the dynamics of simulated language):

in the GPT-4 world, a flipped coin is around 80% likely to land on tails
4. This strange property of simulated language must emerge out of the story worlds we created via thousands of years of writing

what else are we about to discover about semiotic physics?

and how will all this carry over into the virtual worlds that LLMs will create for us?
5. The idea that LLMs are language simulators has also been beautifully articulated by @neilhimself

6. I've long believed that the emergence of virtual worlds (what we were calling the 'metaverse' for a while there way back in 2021) marks the end of a long journey that began with language

in virtual worlds, our representations of the world become worlds unto themselves.
7. The emergence of virtual worlds created by LLMs, though, adds a strange new dimension to that thought

it turns out that our writing was a submerged kind of world building all along…

now, via LLMs and the virtual spaces they create, the worlds that have always been latent in…

Recommended by
Recommendations from around the web and our community.

Great thread!