I trained the fly to play Super Mario

The whole thing is s 3D model of course as well

If you ever want to publish this so other builders can fork it onto their own games instead of rebuilding the sim from scratch, that's the kind of thing we want a home for. Would genuinely use this.

Yes I am gonna release the full repo :)

Nice, looking forward to it!

plays better than lmao

LMAO as an italian smh

I’m literally eating amatriciana rn does that help my case?

barely ahah

omg that's insane

the reward function is the part i want to see tbh. does it care about coins or just about not dying 🙂

The reward function is actually getting to the flag!

I'm worried you're having too much fun!

I want to report a bug

Stupid fly doesn't even know the secret level😭

Manmade horrors beyond our comprehension

Those damn lazy flies bother me all summer when we could play coop games together!! smdh!

Lol

😂😂😂 good one

is this created using astra?

lmaooo

It would be interesting to check the source code to understand how the fly learns. I didn’t use ML in MindFly - I tried to reproduce a published study.

That's cool, but WTF the copywriting across the UI is horrendous. LLMs need to figure that out. Stuffing complete slop everywhere. "Small brain. Big game." Eyebrow headers everywhere for no reason at all. It shouldn't require a skill to unslopify every output.

That looks really cool. That looks really cool

Out if curiosity, what’s your hardware setup? I’m thinking of dabbling in ML and AI.

Astra I guess, do you think Fable might be able to do this?

Anthropic blocks anything remotely close to training model, ML research etc so good luck with that

How did you train the model? using external compute?

nice. mine is playing n64

Doomers, pay attention. The fly race and its connectome is on a dangerous j-curve. They may be solving millennium problems soon and shortly after fixed on world domination.

How ?

oh this is brilliant