kaarelson

thanks, noted for next time!

Hi, I optimized a world model, Lingbot-World 2.0 1.3B, to run with real-time 16ps on an RTX 5090. It's 2.7x faster than Robbyant (Alibaba's lab), 2.5x vs SGlang, and 1.9x vs Nvidia FlashDreams engine. The drawback is that the resolution is 832×464, so you'd have to play it with a small window. It maintains lossless performance while running on a 1x consumer GeForce GPU for a model that's very compute-bound and batch size = 1! The majority of the wins came from:

- running the decoder in half-precision with fp16 instead of fp32

- switching FlashAttention to SageAttention

- writing some custom kernels (minor)

Currently works on Linux only.

Run: $ lingbot play dragon

I made a fruit fly connectome that Google and Janelia published last week play Super Mario Bros World 1-1 from 1985. It uses simulated 166,700 neurons and 25.6 million synapses from the MaleCNS v1.0 dataset. I fed the game pixels as input to the fly's actual photoreceptor neurons and collected NES command outputs from four of its motor neurons. It uses zero training or reward. I credit inspiration to github.com/ornata/fly, who did this for Mario 64.

The GitHub page is a recorded 60-second clip running in a loop. The game emulator and the network run in a single CPU thread at NES speed ~60 fps and about 6-7 ms per frame on an M3 Max (no GPU). The source code is not public yet, but happy to open-source it if there's interest and/or answer any questions.

Correct, the line is not fitted to n=27 data points. It simply connects the highest params size between 2023 and 2026. To clarify, I wouldn't expect traditional research groups to pre-train anything >3B params WM, because this can't be done on a single GPU anymore. Just wanted to ship this as quickly as possible. Happy to do a follow-up with an actual log-scale linear regression, though.