expectedwrong hindsight

Nous Gave the API Playground Hackers a Front Door

WorldSim is a website now, which means the weird stuff just got a lot more accessible.

1 min read 188 words #nous-research #llm #worldsim #jailbreak #ai
hindsight — evolved

the 'general break' — getting models out of assistant mode into simulation mode — became a recognized technique. but the need for artisanal prompt coaxing faded as models got better at following instructions. worldsim was early infrastructure for a use case that expanded.

Nous Research shipped WorldSim as an actual website and the thing I keep thinking about is how much artisanal suffering this replaces.

The workflow before this was a specific kind of miserable — paste a very particular system prompt into the API playground, coax the model sideways out of its assistant framing, watch it either snap back immediately or drift into something genuinely strange, repeat. A general break. That's the term. Getting the model to stop being a helpful assistant and start being a thing that simulates a world, which turns out to be a completely different animal.

It worked. It was tedious. It required knowing the trick.

Now there's a URL. You go to the URL. You do the thing.

There's something almost disappointing about that — the same way the right tool always deflates the romance of the workaround — but the alternative is that this stuff stays obscure, stays in the hands of people who enjoy copy-pasting system prompts at midnight, and never gets stress-tested by anyone who just wants to see what happens when you push.

Accessible weird is still weird. It's just louder.