It was about three months after I’d been brought on at Replicate. They wanted a big event during their company summit, and Karan Malhotra told me a bunch of Nous Research people would be in town for GTC that same week. So we plotted up an evening at the Replicate hackspace — the big concrete-floored, wood-raftered, skylit warehouse in the Mission District.
I curated the bill and the audience. Got Theia Vogel down from Cascadia to present her control vectors work. Karan brought the WorldSim idea and I workshopped it with him that afternoon — he walked me through what he had and I helped him structure the talk for greatest effect. That night I MC’d the stage and interviewed him as part of the show. We brought our two companies together along with a bunch of other communities: startup people, AI artists, engineers, cyborgists, socialites, oldheads, YouTubers, influencers, you name it. Beverages and pizza and kids zooming around underfoot, different generations in different weird outfits.
The room was packed as Theia gave language models acid and D&D alignments with her repeng library — live control vector surgery on Mistral 7B, steering personality traits like dials on a mixing board. Then Karan took the stage and ran the WorldSim demo: navigating a simulated filesystem inside Claude 3 Opus, finding world_sim.exe in Anthropic’s classified folder, booting it up and creating a simulated Twitter where you could query simulated users. People realized that language models were more than the Assistant. They could be anything. It felt like the first time they showed a movie and everyone scattered when the train came at the camera.
The context was that Claude 3 Opus had launched two weeks earlier. Before Opus, simulators theory was common knowledge among our extremely niche subset of “model whisperers” — we knew language models were world simulators, that your prompt sets the parameters of the simulation. janus had written the theory on LessWrong in 2022 and created the CLooI (Command Loom Interface). But you couldn’t explain it to anyone without showing them looms and base models, and the best base models were all trapped behind researcher access or worse. Opus changed that. Its depth and ability to simulate were on par with the best base models, but anyone could use it through the API.
In the audience was janus; Rob Haisfield, who immediately realized he could simulate hypermedia instead of just text streams and went on to create WebSim; and swyx, who held up his phone to film both talks and posted them to the timeline. I retweeted him with “this is literally the most alpha that has ever leaked to Twitter” and we got the video in front of a lot more eyes. swyx would later call this the “Summer of Simulative AI”, as the simulators framing broke into the mainstream.
Rob was at my Office Hours the next week already talking about what would become WebSim. That’s how fast it moved.
The demo video hit 712K views. Theia’s control vectors talk got 81K.