Language models don't have to reason to be useful

We don't demand reasoning from deterministic computers. Why do we demand reasoning from probabilistic ones?
The question of whether a computer can think is no more interesting than the question of whether a submarine can swim
- Dijkstra

Dijkstra would be disappointed by our obsession with whether language models can reason. We're so focused on anthropomorphizing language models that we miss the more banal but valuable qualities.

In practice, whatever language models are now — glorified probabilistic lookup, n-grams on steroids, spicy autocomplete — is enough for sufficiently motivated product builders to build something magical.

World simulation is one of the more banal but valuable qualities of language models we've neglected in our obsession with reasoning. Recently, a few early projects have been pushing the boundaries of world simulation in exciting new directions.

WebSim is one example I've been obsessed with.

Websim is a web browser that seemingly lets you browse the internet. If you put in a URL, it returns a webpage — except that webpage does not exist. It is fully hallucinated by a language model — in this case, Claude 3.

Here is WebSim generating a blog post:

Here is WebSim generating a somewhat functional Figma prototype.

WebSim today feels like a toy that could easily become something more profound.

The general capability WebSim highlights — world simulation — works as well for simulating a made-up internet as it does for simulating software. WebSim's long-forgotten predecessor was an early experiment that simulated an entire virtual machine inside ChatGPT.

We can extend world simulation to other tasks. Consider this recent paper where 'doctor agents' collaborate with simulated 'patient agents' to enhance care delivery. The benefits of world simulation in professional training are evident. Most coaching/training involves tasks that aren't performed in isolation. Simulated worlds provide a superior environment for modeling the interactions trainees are likely to encounter in real-world scenarios. This applies not just to doctors but also to sales reps, police officers, therapists, and more. 

World simulation only scratches the surface of what we can use language models for today — right now. Not in some distant future where models reason.

We don't demand reasoning from deterministic computers. Why do we demand reasoning from probabilistic ones?