Re:Inference business simulator – don't go broke, rent GPUs,...
by dval on Thursday November 05, 2026 @05:00PM
(#50385088)
Hi HN, we made a game out of the fast-growing inference business. See if you can run the most profitable inference business in this market, or at least not go bankrupt.
You start with $10k and no GPUs. You rent GPUs, pick which open-weights models to serve on which of them, guess where the market is going, and take over the world, er, market share. Spare GPUs go to Elise, our research agent, who finds speedups for the engine that you then install. Your score is the capital you end the year with.
We made it because the job of Elise, one of our research agent, is tuning inference engines like vLLM on AMD and NVIDIA GPUs, and it's a lot more fun to explain that to friends and family with a game than with our report: https://makermaker.ai/breakthroughs/inference/ (the code and baselines are there if you want the long version).
Re:Inference business simulator – don't go broke, rent GPUs,...
by cortexcollider on Friday November 06, 2026 @12:09PM
(#50395138)
Hi Dhaval, fun game. Interested to hear your take on RSI after Anthropic just updated their S1 to include it. Also, how do you know that your research agent can keep up with the pace of optimizing the "black boxes" of other models if they also start recursively optimizing in-house?
Re:Inference business simulator – don't go broke, rent GPUs,...
by dval on Friday November 06, 2026 @03:58PM
(#50398431)
Most labs should be doing RSI but it's only now starting to come out.
Your question about keeping page with other black boxes that are also doing RSI is very profound. The good thing is that as AI gets better, it's more and more able to deal with black boxes. The hard thing (putting on my physicist hat!) is that I am not familiar with equilibrium models that work well yet for such "turtles all the way" problems. It's a very exciting time!
Re:Inference business simulator – don't go broke, rent GPUs,...
by noakishere on Friday November 06, 2026 @12:13PM
(#50395200)
Pretty cool concept! How stable is Elise right now? I see the simulation, but would like to understand how the recursive learning could apply to a larger scale production
Re:Inference business simulator – don't go broke, rent GPUs,...
by dval on Friday November 06, 2026 @03:55PM
(#50398392)
Elise is pretty stable, but is under development in terms of continuously extending her capabilities. Luckily the recursive self-improvement helps a lot. One of the most interesting part right now is the human <> AI multi-agent interaction!
Re:Inference business simulator – don't go broke, rent GPUs,...
by jmjoseph on Friday November 06, 2026 @05:46PM
(#50399525)
I don't think you really answered the part about larger scale production. Seems RSI may work with small, constrained problems but I don't know of any real demonstrations of RSI working at scale?
You start with $10k and no GPUs. You rent GPUs, pick which open-weights models to serve on which of them, guess where the market is going, and take over the world, er, market share. Spare GPUs go to Elise, our research agent, who finds speedups for the engine that you then install. Your score is the capital you end the year with.
We made it because the job of Elise, one of our research agent, is tuning inference engines like vLLM on AMD and NVIDIA GPUs, and it's a lot more fun to explain that to friends and family with a game than with our report: https://makermaker.ai/breakthroughs/inference/ (the code and baselines are there if you want the long version).
The game is open source https://github.com/MakerMakerMakerMakerMaker/Inference-Machi.... Bug reports welcome there, or run your own copy.
Your question about keeping page with other black boxes that are also doing RSI is very profound. The good thing is that as AI gets better, it's more and more able to deal with black boxes. The hard thing (putting on my physicist hat!) is that I am not familiar with equilibrium models that work well yet for such "turtles all the way" problems. It's a very exciting time!