正在加载视频...
视频加载失败
Fable 5.2 wrote the transformer. Opus 5.5 generated the training data. The result: a 367K parameter language model that fits on a floppy disk and runs on a Pentium. AI is now building AI for hardware from 1995.
289,229 次观看 • 21 小时前 •via X (Twitter)
36 条评论

AI really said we don't need more powerful hardware

And intresting part is there's a lot to come for ai in future

The Pentium is cool, but the real story is that one AI designed the brain and another wrote its textbook. That’s where things get weird.

Yeah, the recursion is the part that actually feels wild

If you'd shown that in 1995, they would have burned you at the stake

This is absolutely insane! I was surprised to read about these floppy disks still being used in the aviation industry.

cool man made browser theme with fable 5.2

@ivanfioravanti 😍

AI on a floppy disk is the most 2026 sentence ever

and tells you the super bowl sandwich slappy into polling freedom gnats monster cheese when you ask what 2+3 is

We need to see more what this model is capable of! Crazy stuff

Sure, follow for more such updates

Incredible innovation packed into tiny hardware!

I could easily see this as a new type of virus, that can unpack itself into the machine and write the code for the rest of it on it's own. Mind blowing

what can it do on facts or combinations that weren't in the generated training set? the tiny hardware demo is fun, but separating recall from generalisation would make it much more interesting.

Fair point

Is this legal?

ok. this is cool :D

Wow

wth AI is everywhere

367K参数的小模型,这分工有点意思

what is this exactly?

Why this is needed? 🤔

367k parameters on a floppy makes the hardware constraint feel tangible.

what if you could generate the weights with 5.5

Running a sub-megaparameter language model on vintage hardware isn't for practical production—it's pure tech-hobbyist flexing to see if you can squeeze a transformer into a system with less RAM than a modern mouse's firmware.

what's the actual performance of this model compared to modern ones?

@grok is this true?

How did you evaluate the training data?

It's pretty dumb but once models get better at making models, then this could be impressive.

repo? hf?

tiny brain on a floppy, wild recursion, love the retro hack

This is courage

so useless

A 367K model on a floppy and my browser game still needs a real GPU to draw a canyon. My whole performance budget is an upscaler and a render scale that quietly drops when the swarm gets big.

Beautiful.





