nesformer 10月30日开源，把红白机 CPU 编进千问模型权重，用生成 token 的方式跑游戏。权重由编译器算出，没经过训练。
官方说，游戏里10秒要跑约25分钟，目前只支持 NROM 卡带。想拿它替代模拟器，先忍住。
dbreunig/nesformer
4 likes · 113 views
有人把一台能跑的任天堂塞进了约 1GB 的 Qwen 权重里，以后下游戏可能得改口叫下模型了。
项目叫 nesformer，代码在这里：dbreunig/nesformer
Show HN: A working Nintendo in ~1GB of Qwen weights
View Article | Discussion + AI Summary
Summary of HN discussion 🧵👇
2 likes
Drew Breunig compiled an NES CPU into Qwen weights
the README says a frame takes 1 to 6 seconds
finally, a gaming setup where being a skeleton helps you wait
dbreunig/nesformer t.co
2 likes · 62 views
A working NES emulator built into the weights of Qwen3.5-4B. Per the author, you can play Super Mario Bros in it, very slowly. The doubt: a demo is not a benchmark. I'd want the compute cost per frame. dbreunig/nesformer
5 likes · 266 views
builds on qwenputer by divyansh agrawal. repo: dbreunig/nesformer
2 likes · 27 views
Drew Breunig compiled the NES 6502 CPU into the weights of a stock Qwen3.5 model. Nobody trained it, a compiler computed every weight. It predicts the CPU's execution log one token at a time. Mario runs, up to six seconds a frame. Haven't run it.
dbreunig/nesformer
#Emulation
1 like
Show HN: A working Nintendo in ~1GB of Qwen weights
1 point
if you look at dbreunig/nesformer
It's obvious a LLM can do anything a computer can do.
As an answer to "Why would LLMs exhibit sentience?"
It is because that would mean that no algorithm could exhibit sentience, claiming that would require some causal action that cannot be performed by an algorithm to create sentience. Some people like to point to quantum effects, but that would require the cause of sentience to be pure randomness. That seems even less plausible.
Absent of any evidence that human brains are doing something that no algorithm could do, it would be reasonable to assume that sentience could be algorithmically obtained.
A Working Nintendo in ~1GB of Qwen Weights
1 point · 1 comment
Mentions
nesformer 10月30日开源，把红白机 CPU 编进千问模型权重，用生成 token 的方式跑游戏。权重由编译器算出，没经过训练。 官方说，游戏里10秒要跑约25分钟，目前只支持 NROM 卡带。想拿它替代模拟器，先忍住。 dbreunig/nesformer
有人把一台能跑的任天堂塞进了约 1GB 的 Qwen 权重里，以后下游戏可能得改口叫下模型了。 项目叫 nesformer，代码在这里：dbreunig/nesformer
Show HN: A working Nintendo in ~1GB of Qwen weights View Article | Discussion + AI Summary Summary of HN discussion 🧵👇
Drew Breunig compiled an NES CPU into Qwen weights the README says a frame takes 1 to 6 seconds finally, a gaming setup where being a skeleton helps you wait dbreunig/nesformer t.co
A working NES emulator built into the weights of Qwen3.5-4B. Per the author, you can play Super Mario Bros in it, very slowly. The doubt: a demo is not a benchmark. I'd want the compute cost per frame. dbreunig/nesformer
builds on qwenputer by divyansh agrawal. repo: dbreunig/nesformer
Drew Breunig compiled the NES 6502 CPU into the weights of a stock Qwen3.5 model. Nobody trained it, a compiler computed every weight. It predicts the CPU's execution log one token at a time. Mario runs, up to six seconds a frame. Haven't run it. dbreunig/nesformer #Emulation
Show HN: A working Nintendo in ~1GB of Qwen weights
if you look at dbreunig/nesformer It's obvious a LLM can do anything a computer can do. As an answer to "Why would LLMs exhibit sentience?" It is because that would mean that no algorithm could exhibit sentience, claiming that would require some causal action that cannot be performed by an algorithm to create sentience. Some people like to point to quantum effects, but that would require the cause of sentience to be pure randomness. That seems even less plausible. Absent of any evidence that human brains are doing something that no algorithm could do, it would be reasonable to assume that sentience could be algorithmically obtained.
A Working Nintendo in ~1GB of Qwen Weights