This couldn't be more perfectly timed .. I have an Unreal Engine game with both VT100 terminals (for running coding agents) and Z80 emulators, and a serial bridge that allows coding agents to program the CP/M machines:
I've been playing the Z80-μLM demos in your CP/M emulator. Works great! However, I have yet to guess a correct answer in GUESS.COM! I'm not sure if I'm just not asking the right questions or I'm just really bad at it!
Those "coincidences" in Connections are really no coincidence at all, but path dependence. Breakthrough advance A is impossible or useless without prerequisites B and C and economic conditions D, but once B and C and D are in place, A becomes obvious next step.
Some of those really are coincidences, like "Person A couldn't find their left shoe and ended up in London at a coffee house, where Person B accidentally ended up when their carriage hit a wall, which lead to them eventually coming up with Invention C" for example.
Although from what I remember from the TV show, most of what he investigates/talks about is indeed path dependence in one way or another, although not everything was like that.
Indeed, part of me wants to not use imgur because we can't access it, but a bigger part of me fully supports imgur's decision to give the middle finger to the UK after our government's censorship overreach.
I love it, instant Github star.
I wrote an MLP in Fortran IV for a punched card machine from the sixties (https://github.com/dbrll/Xortran), so this really speaks to me.
The interaction is surprisingly good despite the lack of attention mechanism and the limitation of the "context" to trigrams from the last sentence.
This could have worked on 60s-era hardware and would have completely changed the world (and science fiction) back then. Great job.
Stuff like this is fascinating. Truly the road not taken.
Tin foil hat on: i think that a huge part of the major buyout of ram from AI companies is to keep people from realising that we are essentially at the home computer revolution stage of llms. I have a 1tb ram machine which with custom agents outperforms all the proprietary models. It's private, secure and won't let me be motetized.
If one would train an actual secret (e.g. a passphrase) into such a model, that a user would need to guess by asking the right questions. Could this secret be easily reverse engineered / inferred by having access to models weights - or would it be safe to assume that one could only get to the secret by asking the right questions?
I don’t know, but your question reminds me of this paper which seems to address it on a lower level: https://arxiv.org/abs/2204.06974
“Planting Undetectable Backdoors in Machine Learning Models”
“ … On the surface, such a backdoored classifier behaves normally, but in reality, the learner maintains a mechanism for changing the classification of any input, with only a slight perturbation. Importantly, without the appropriate "backdoor key", the mechanism is hidden and cannot be detected by any computationally-bounded observer. We demonstrate two frameworks for planting undetectable backdoors, with incomparable guarantees. …”
Awesome. I've just designed and built my own z80 computer, though right now it has 32kb ROM and 32kb RAM. This will definitely change on the next revision so I'll be sure to try it out.
This is excellent. Thing I’d like to do if I had time: get it running on a 48K Spectrum. 10 year old me would have found that absolutely magical back in the 1980s.
There are two things happening here. A really small LLM mechanism which is useful for thinking about how the big ones work, and a reference to the well known phenomenon, commonly dismissively referred to as a "trick", in which humans want to believe. We work hard to account for what our conversational partner says. Language in use is a collective cultural construct. By this view the real question is how and why we humans understand an utterance in a particular way. Eliza, Parry, and the Chomsky bot at http://chomskybot.com work on this principle. Just sayin'.
Pretty cool! I wish free-input RPGs of old had fuzzy matchers. They worked by exact keyword matching and it was awkward. I think the last game of that kind (where you could input arbitrary text when talking to NPCs) was probably Wizardry 8 (2001).
Luckily I have a very large amount of MSX computers, zx, amstrad cpc etc and even one multiprocessor z80 cp/m machine for the real power. Wonder how gnarly this is going to perform with bankswitching though. Probably not good.
> It won't write your emails, but it can be trained to play a stripped down version of 20 Questions, and is sometimes able to maintain the illusion of having simple but terse conversations with a distinct personality.
You can buy a kid’s tiger electronics style toy that plays 20 questions.
It’s not like this LLM is bastion of glorious efficiency, it’s just stripped down to fit on the hardware.
Slack/Teams handles company-wide video calls and can render anything a web browser can, and they run an entire App Store of apps, all from a cross-platform application.
Including Jira in the conversation doesn’t even make logical sense. It’s not a desktop application that consumes memory. Jira has such a wide scope that the word “Jira” doesn’t even describe a single product.
> Slack/Teams handles company-wide video calls and can render anything a web browser can, and they run an entire App Store of apps, all from a cross-platform application.
The 4th Gen iPod touch had 256 meg of RAM and also did those things, with video calling via FaceTime (and probably others, but I don't care). Well, except "cross platform", what with it being the platform.
My Pentium 3 in 2005 could do chat and video calls and play chess and send silly emotes. There is no conceivable user-facing reason why in 20 years the same functionality takes 30× as many resources, only developer-facing reasons. But those are not valid reasons for a professional. If a bridge engineer claims he now needs 30× as much concrete to build the same bridge as he did 20 years ago, and the reason is his/her own conveinence, that would not fly.
> If a bridge engineer claims he now needs 30× as much concrete to build the same bridge as he did 20 years ago, and the reason is his/her own conveinence, that would not fly.
By itself, I would agree.
However, in this metaphor, concrete got 15x cheaper in the same timeframe. Not enough to fully compensate for the difference, but enough that a whole generation are now used to much larger edifices.
It's pretty obvious this is just a stress test for compressing and running LLMs. It doesn't have much practical use right now, but it shows us that IoT devices are gonna have built-in LLMs really soon. It's a huge leap in intelligence—kind of like the jump from apes to humans. That is seriously cool.
i'll echo that practicality only surfaces once it is apparent what can be done. yea this feels like running "DOOM on pregnancy test devices" type of moment
Between this and RAM prices Zilog stock must be up! Awesome hack. Now apply the same principles to a laptop and take a megabyte or so, see what that does.
Nice - that will fit on a Gameboy cartridge, though bank switching might make it super terrible to run. Each bank is only 16k. You can have a bunch of them, but you can only access one bank at a time (well, technically two - bank 0 is IIRC always accessible).
You have 32KB of ROM, plus 8 Kb of ram on original game boy. Game boy color has more. Bank switching is super fast, as well. Given that models are likely streamed, I doubt the bank switching is a problem.
It depends on the model, but from my experiments (quantizing one layer of a model to 2-bit and then training the model with that layer in 2-bit to fix the damage) the first layer is the most sensitive, and yes, the last layer is also sensitive too. The middle layers take the best to quantization.
Different components of a layer also have a different sensitivity; e.g. the MLP downscale block damages the model the most when quantized, while quantizing the Q projection in self attention damages the model the least.
For future projects and/or for this project, there are many LLMs available more than good enough to generate that kind of synthetic data (20 Qs) with permissive terms of use. (So you don’t need to stress about breaking TOS / C&D etc)
All the 'Small' language models and the 'TinyML' scene in general tend to bottom out at a million parameters, hence I though 'micro' is more apt at ~150k params.
interesting, i am wondering how far can it go if we remove some of these limitations but try to solve some extremely specific problem like generating regex based on user input? i know small models(270M range) can do that but can it be done in say < 10MB range?
Generate an LLM that is designed to solve one extremely specific problem: answering the ultimate question of life, the universe, and everything.
Even with modern supercomputing the computation would be outpaced by the heat death of the universe, so token output must be limited to a single integer.
I love these thought experiments. Looking at the code size, it would have been possible for someone to come up with this back in the days, similar to the idea of a million monkeys on a typewriter eventually producing Shakespeare.
Speaking of - I remember my first digital camera (Fujitsu 1Mb resolution using SmartMedia)… it used so much power that you could take 20-30 photos and then needed to replace all 4 batteries lol
https://i.imgur.com/6TRe1NE.png
Thank you for posting! It's unbelievable how someone sometimes just drops something that fits right into what you're doing. However bizarre it seems.
I developed a browser-based CP/M emulator & IDE: https://lockboot.github.io/desktop/
I was going to post that instead, but wanted a 'cool demo' instead, and fell down the rabbit hole.
Although from what I remember from the TV show, most of what he investigates/talks about is indeed path dependence in one way or another, although not everything was like that.
The interaction is surprisingly good despite the lack of attention mechanism and the limitation of the "context" to trigrams from the last sentence.
This could have worked on 60s-era hardware and would have completely changed the world (and science fiction) back then. Great job.
Tin foil hat on: i think that a huge part of the major buyout of ram from AI companies is to keep people from realising that we are essentially at the home computer revolution stage of llms. I have a 1tb ram machine which with custom agents outperforms all the proprietary models. It's private, secure and won't let me be motetized.
“Planting Undetectable Backdoors in Machine Learning Models”
“ … On the surface, such a backdoored classifier behaves normally, but in reality, the learner maintains a mechanism for changing the classification of any input, with only a slight perturbation. Importantly, without the appropriate "backdoor key", the mechanism is hidden and cannot be detected by any computationally-bounded observer. We demonstrate two frameworks for planting undetectable backdoors, with incomparable guarantees. …”
It could with a network this small. More generally this falls under "interpretability."
You can buy a kid’s tiger electronics style toy that plays 20 questions.
It’s not like this LLM is bastion of glorious efficiency, it’s just stripped down to fit on the hardware.
Slack/Teams handles company-wide video calls and can render anything a web browser can, and they run an entire App Store of apps, all from a cross-platform application.
Including Jira in the conversation doesn’t even make logical sense. It’s not a desktop application that consumes memory. Jira has such a wide scope that the word “Jira” doesn’t even describe a single product.
The 4th Gen iPod touch had 256 meg of RAM and also did those things, with video calling via FaceTime (and probably others, but I don't care). Well, except "cross platform", what with it being the platform.
That's a bug not a feature, and strongly coupled to the root cause for slack's bloat.
By itself, I would agree.
However, in this metaphor, concrete got 15x cheaper in the same timeframe. Not enough to fully compensate for the difference, but enough that a whole generation are now used to much larger edifices.
Biggest pain point is likely the text input.
Have you experimented with having it less quantized, and evaluated the quality drop?
Regardless, very cool project.
It depends on the model, but from my experiments (quantizing one layer of a model to 2-bit and then training the model with that layer in 2-bit to fix the damage) the first layer is the most sensitive, and yes, the last layer is also sensitive too. The middle layers take the best to quantization.
Different components of a layer also have a different sensitivity; e.g. the MLP downscale block damages the model the most when quantized, while quantizing the Q projection in self attention damages the model the least.
Even with modern supercomputing the computation would be outpaced by the heat death of the universe, so token output must be limited to a single integer.
Speaking of - I remember my first digital camera (Fujitsu 1Mb resolution using SmartMedia)… it used so much power that you could take 20-30 photos and then needed to replace all 4 batteries lol