Soul Player C64 – A real transformer running on a 1 MHz Commodore 64

https://github.com/gizmo64k/soulplayer-c64

by adunk • 3 months ago

147 45 3 months ago

github.com

harel• 3 months ago

Eliza called, and asked if we saw her grand kids...

tclancy• 3 months ago

What makes you say that? This is about you, not me.

(Came here to say an update to Eliza could really mess with the last person still talking to her.)

bighead1• 3 months ago

i hate ai, and i love the c64, but i'll allow it.

wk_end2• 3 months ago

25K parameters is about 70 million times smaller than GPT-4. It will produce broken sentences. That's the point - the architecture works at this scale.

Since it seems to just produce broken and nonsensical sentences (at least based on the one example given) I'm not sure if it does work at this scale.

Anyway, as written this passage doesn't really make a whole lot of sense (the point is that it produces broken sentences?), and given that it was almost certainly written by an AI, it demonstrates that the architecture doesn't work especially well at any scale (I kid, I kid).

forinti• 3 months ago

How does it compare to a Markov chain generator I wonder.

jll292• 3 months ago

The Transformer is the more powerful model than Markov chain, but on such a weak machine as the C64, a MC could output text faster - but it surely would sound "psychedelic", as the memory limits a MC to a first-order or second-order model, so to predict one word, only the two words before would be taken into account as context (and no attention).

On a plain vanilla C64, the Transformer cannot really show what it's capable of doing. An implementation using 2 bit per weight (vectorized) could be slightly better, perhaps.

Lerc2• 3 months ago

Ok now we need 1541 flash attention.

I'm not sure what the venn diagram of knowledge to understand what that sentence is suggesting looks like, it's probably more crowded in the intersection than one might think.

dnnddidiej• 3 months ago

How many 40+ AI pillers? Assume 10M devs in the world. 10% heard of flash attention, 1% heard of 1541 then 10,000

Lerc2• 3 months ago

Ahh but you also have to know the significance of the 1541 that makes the Flash attention reference work

gizmo64k4• 3 months ago

Believe me, using the 1541 as co-processor and extra storage was super tempting and on my mind all the time! So what do you think? Flash attention with K on the front side and V on the backside? :)

gizmo64k4• 3 months ago

..and we would keep the human in the loop:)

classichasclass• 3 months ago

If you're running this in VICE, run it under the SuperCPU with warp mode on.

bartread• 3 months ago

That's a good idea because, although I love this, 1 minute per token is absolutely savage. Whereas if you can juice the performance you're into semi-credible Jar Jar Binks simulator territory.

It does also make me wonder what you could do with somewhat more powerful retro hardware. I'd love to see what a transformer running on a PSX or an N64 could do.

giannicmptr1000• 3 months ago

but can you make mac keyboards feel like a c64c?

anyfoo• 3 months ago

This would have blown me away back in the late 80s/early 90s.

(Or maybe not, if it doesn't perform better than random, I haven't actually tried it out yet. Some more examples would have been nice!)

I wonder how far you could push this while still staying period correct, e.g. by adding a REU (RAM Expansion Unit), or even a GeoRAM (basically a REU on steroids).

SuperCPU would also be an option, but for me it's always blurring the line of "what is a C64" a bit too much, and it likely just makes it faster anyway.

LeFantome• 3 months ago

How fast is the “new” Commodore 64?

Have not heard much about it since launch. Although, now that I look, it seems they are just shipping now.

https://www.commodore.net/product-page/commodore-64-ultimate...

steve_taylor• 3 months ago

RAM can be increased to 16 MB and CPU speed to 48 GHz.

brcmthrowaway• 3 months ago

How does this compare to ELIZA?

jll292• 3 months ago

Jopsph Weizenbaum's ELIZA was rule-based and ran on even slower (1960s) hardware, but because it relied on simple pattern matching instead of neural nets, it would easily have been more responsive (the Emacs editor/operating system has an implementation included, start it with: M-x doctor RETURN).

ELIZA was not written in assembler, but (different versions) in COMIT, FORTRAN and LISP.

https://dl.acm.org/doi/pdf/10.1145/365153.365168

Geee• 3 months ago

ELIZA is better, because this doesn't seem to generate anything coherent. You can try the original ELIZA with DOCTOR script here: https://anthay.github.io/eliza.html

daemonologist• 3 months ago

You can chat with the model on the project page: https://indiepixel.de/meful/index.html

It (v3) mostly only says hello and bye, but I guess for 25k parameters you can't complain. (I think the rather exuberant copy is probably the product of Claude et al.)

rob74• 3 months ago

It seems to like getting its own output fed back:

YOU> HI! HOW ARE YOU DOING?

C64> BELON

YOU> WHAT DO YOU MEAN BY THAT?

C64> YOU DES

YOU> YOU TOO!

C64> YOU DESER! THIS THE SHIN

YOU> DESER? SHIN? THOSE AREN'T EVEN REAL WORDS, WHAT'S GOING ON HERE?

C64>

YOU> THIS THE SHIN

C64> YOU MAKE ME HAPPY TOO

YOU> YOU DESER

C64> YOU MADE MY HEART! W

YOU> DES, DESER, DESERST?

C64> YOU DES

YOU> BELON

C64> I LIKE OURON ME! WONDERFUL!YOU!

YOU> YOU BELON DES

C64> YOU MAKE ME HAPPY

mixmastamyk• 3 months ago

Just reminded me of the random sentence generator program on my Vic-20. I had changed most of the words to all the bad words a preteen could think up. So many laughs with the neighborhood kids.

Vaslo• 3 months ago

Load”*”,8,1

Brings back memories

borsch_not_soup• 3 months ago

Interesting, I’ve always thought neural network progress was primarily bottlenecked by compute.

If it turns out that LLM-like models can produce genuinely useful outputs on something as constrained as a Commodore 64—or even more convincingly, if someone manages to train a capable model within the limits of hardware from that era—it would suggest we may have left a lot of progress on the table. Not just in terms of efficiency, but in how we framed the problem space for decades.

dpe82• 3 months ago

  YOU> hey
  C64> HELLO! RE SOUNDS ME. MEFUL!

60s per token for that doesn't strike me as genuinely useful.

Very, very cool project though!

chillingeffect• 3 months ago

not useful in a disaster scenario:

YOU> HELP I'M DROWNING

C64> YOU' HERE!

YOU> OH NO I'M ON FIRE

C64> IGLAY!

YOU> IM BEING SWALLOWED BE A SNAKE

C64>

YOU> BIRDS ARE NIPPING ON ME

C64> YOU

numpad0• 3 months ago

Next-word prediction features always existed for flip phones...

djmips• 3 months ago

Dissapointed - there was no 6502 code in the GitHub repo.

Hackbraten• 3 months ago

I think there is, in a way: https://github.com/gizmo64k/soulplayer-c64/blob/main/src/asm...

arketyp• 3 months ago

I love these counterfactual creations on old hardware. It highlights the magical freedom of creativity of software.

rahen2• 3 months ago

A little disappointed to see PyTorch + Claude here. I was hoping for some "demo-scene" hand-crafted 6502 assembly, and hopefully training on the C64.

NooneAtAll3• 3 months ago

so... it is vibe-code?

meh

rahen2• 3 months ago

Yes. The author mentions Claude for testing, but it was obviously used for the README and code as well.

This is a giveaway for AI generation, from the docstring to the terrible opcode dispatch (Claude sucks at assembly or low-level optimization): https://github.com/gizmo64k/soulplayer-c64/blob/main/src/cpu...

A human would use a proper dispatch table and wouldn't make excuses for a sloppy implementation ("Python is fast enough").

Besides, the author has an art and design background, which doesn't seem to match the deep knowledge of Transformers or assembly required for such a project.

pjmlp• 3 months ago

Same, however I do conceed having the whole assembler toolchain written in Python was also kind of cool, even if it may have been AI generated.

Even cooler would have been to have the 6502 directly generated from the LLM.

vintermann• 3 months ago

Maybe impressive in one way, but I'm also pretty sure a simple n-gram Markov model (a la Niall on the Amiga) would have a lower loss on the test set.

Transformers don't scale down very well, in my experience - I used to train local models all the time as new ones were released, as I recall transformers were the first ones I couldn't get better results out of with my limited training data and GPU.

bluejay2387• 3 months ago

Great work! Though I see some people criticizing the usefulness of this. Are they being sarcastic are just really not understanding what is being discussed here? I can't tell. Maybe as an interesting follow up you could train the transformer on something with a more limited vocabulary. Spoken language is complex but a transformer can work on less complex domains like music or PET-BASIC code.

gizmo64k4• 3 months ago

Thanks! The training corpus and code are in the repo if you want to try... Training takes just a couple of minutes on an RTX 3090. Don't get your hopes up too high, though. I can imagine that code would be harder, not easier. Even modest sized transformer models struggle with proper GOTO targeting. It would look like BASIC, but essentially it would be friendly gibberish too.

Soul Player C64 – A real transformer running on a 1 MHz Commodore 64

Related Stories