Hacker News Reader

Comment by throwa356262 | Hacker News Reader

"LLM backends: Anthropic, OpenAI, OpenRouter."

And here I was hoping that this was local inference :)

Sure. Why purchase a H200 if you can go with an ESP32 ^^

Blowing more than 800kb on essentially an http api wrapper is actually kinda bad. The original Doom binary was 700kb and had vastly more complexity. This is in C after all, so by stripping out nonessential stuff and using the right compiler options, I'd expect something like this to come in under 100kb.

yeah i sandbagged the size just a little to start (small enough to fit on the c3, 888 picked for good luck & prosperity; I even have a build that pads to get 888 exactly), so i can now try reduce some of it as an exercise etc.

but 100kb you’re not gonna see :) this has WiFi, tls, etc. doom didn’t need those

Doom had the benefit of an OS that included a lot of low-level bits like a net stack. This doesn’t! That 800kB includes everything it would need from an OS too.

yah my back of the envelope math..

the “app logic”/wrapper pieces come out to about 25kb

WiFi is 350 Tls is 120 and certs are 90!

Maybe you’re misremembering or referring to Doom (2016). The original Doom was developed for DOS and id had to build a lot of its own network stack. BSD style socket based networking wasn’t a given in DOS.

Still, zclaw is an impressive achievement.

> vastly more complexity.

Doom is ingenious, but it is not terribly complex IMHO, not compared to a modern networking stack including WiFi driver. The Doom renderer charm is in its overall simplicity. The AI is effective but not sophisticated.

The whole ESP32 Libraries are kind of bloated. To enable Bluetooth, WiFi or HTTP handling, you need to embed some large libraries

haha well I got something ridiculous coming soon for zclaw that will kinda work on board.. will require the S3 variant tho, needs a little more memory. Training it later today.

Sounds interesting, please keep us posted.

I dont think I have an S3, but plenty of C3. I thought they had the same amount of memory

right, 888 kB would be impossible for local inference

however, it is really not that impressive for just a client

It's not completely impossible, depending on what your expectations are. That language model that was built out of redstone in minecraft had... looks like 5 million parameters. And it could do mostly coherent sentences.

  > built out of redstone in minecraft

Ummm...

  > 5 million parameters

Which is a lot more than 888kb... Supposing your ESP32 could use qint8 (LOL) that's still 1 byte per parameter and the k in kb stands for thousand, not million.

https://www.youtube.com/watch?v=VaeI9YgE1o8

Yes I know how much a kilobyte is. But cutting down to 2 million 3 bit parameters or something like that would definitely be possible.

And a 32 bit processor should be able to pack and unpack parameters just fine.

Edit: Hey look what I just found https://github.com/DaveBben/esp32-llm "a 260K parameter tinyllamas checkpoint trained on the tiny stories dataset"

  > But cutting down to 2 million 3 bit parameters or something like that would definitely be possible.

Sure, but there's no free lunch

  > Hey look what I just found

I've even personally built smaller "L"LMs. The first L is in quotes because it really isn't large (So maybe lLM?) and they aren't anything like what you'd expect and certainly not what the parent was looking for. The utility of them is really not that high... (there are special cases though) Can you "do" it? Yeah. I mean you can make a machine learning model of essentially arbitrary size. Will it be useful? Obviously that's not guaranteed. Is it fun? Yes. Is it great for leaning? Also yes.

And remember, Tiny Stories is 1GB of data. Can you train it for longer and with more data? Again, certainly, BUT again, there are costs. That Minecraft one is far more powerful than this thing.

Also, remember that these models are not RLHF'd, so you really shouldn't expect it to act like you're expecting a LLM to work. It is only at stage 0, the "pre-training", or what Karpathy calls a "babbler".

A reminder that what I said was "not completely impossible, depending on what your expectations are"

And I was focused more on the ESP32 part than the exact number of bytes. As far as I'm concerned you can port the model from the minecraft video and you still win the challenge.

Also, that last link isn't supposed to represent the best you can do in 800KB. 260k parameters is way way under the limit.

That bar has no lower bound though so of course we're talking past one another

Also we're talking about an esp32. They aren't magic

Being able to talk back and forth with coherent sentences has a lower bound, and it's close to the limit of this hardware.

Something that can actually be an "assistant" has its own lower bound, probably a little harder but mostly a matter of training it differently.

I disagree, in the future it might be possible. But perhaps not in English, but in some more formal (yet fuzzy) language with some basic epistemology.

I mean, there is a lambda calculus self-interpreter in 29 bytes. How many additional logical rules are required for GAI inference? Maybe not that many as people think. Understanding about 1000 concepts of basic english (or say, lojban) might well be sufficient. It is possible this can be encoded in 800kB, we just don't know how.

Reply on news.ycombinator.com