There was an interesting bit of research that qwen has a single neuron that makes it “demonic evil qwen” and a simple “Who are you?” will get some absolutely terrifyingly vile results.
Anyways, what I’m trying to say is I want evil qwen to be my CPU.
The single-neuron steering stuff has been going around twitter lately, do you happen to know who was the first one to do it?
I wonder how this squares with the recent "NNs are already clean" interpretability finding, which may have also been the impetus for this kind of research. Similar to their work on CNNs, maybe residual-steering is overkill, there might be just a single neuron that "cleanly" corresponds to "evilness" that you can turn up/down?
This was not a neuron. This was from a trained sparse auto encoder.
In a few words, they found a way to map residual stream activations (so, a vector) into a few interpretable sparse concepts. They also had the converse maps, that translate concepts into activations.
Golden Gate claude was obtained by artificially increasing the activations in the direction of the 'Golden Gate' concept. It was not neuron surgery.
The real question is, do current training methods allow one to learn this when training on enough, say, assembly, or can the weights only be achieved via hard-coding?
A fun and important thought experiment about the the space of nerual networks vs the image of current training methods.
Most discourse conflates neural networks and trainable neural networks.
> When Qwen 'saves' a file, the compiled layers hash its name into one of 1,024 slots in a directory kept in the DeltaNet state, and store the text itself in other heads, with eight tokens per row, chained together
I’ve never wanted to learn how LLMs work more than reading this. Fascinating.
Does divagr get paid to do wild things with LLMs? Just curious if they’re being compensated for these wild experiments
Along these lines I’m curious how a scratch trained model with a few built-in tools (memory, calculator, ?) available during a single generation would behave. Not tool calls as turns.
a single generation would still allow looping in the "CPU" because the instruction set can do it. The model is almost a Turing machine (except for the limit on memory)
> Now, here's where our fun experiment starts. I found out that by setting the gate to 1,1 the write strength (beta) to 1,2 and using one-hot key encoding,3 every write we perform will overwrite exactly one row, and every one-hot query can retrieve an exact value. This means that the GDN layers themselves can work exactly as RAM, at a far lower cost than using normal attention as RAM, since a GDN head's state stays the same size however long the conversation gets, unlike attention's KV cache.
So like... can't you use a computer as a computer without doing all that roundabout stuff? Like if you want to integrate a LLM with hard computation there's probably a more efficient way to do that.
Or is this building towards "but can it run DOOM?" :)
I’ve always wondered why we don’t add a few frozen “layers” to models that perform complex computations 2/3rds deep in the model. Some set of steerable fft + some common vector math operations could be reusable units of capability that the model learns to leverage during training.
Given that models learn to use residual connections optimally, it can’t be hard for them to steer around the fixed computation layer when it’s not needed.
I’m sure giant models invest a few layers in reproducing these functions, but surely smaller models can benefit?
It could be interesting to experiment with more complex representations in the memory/file system, i.e. vectors. Seems like the full advantage of this architecture is yet to be explored. Blending neurons with deterministic compute to fit on one GPU is a fascination concept, never considered it before!
i wonder if you could implement "authenticated token authorship" and self-verified running with this.
the model verifies an input as signed by an authorized requester, and signs its own outputs. eventually it also authenticates its own running via e.g. snarks? :)
there's lots of meta available with these toys..
This is cool and I understand the RAM emulation with gated attention but how does the CPU emulation wor? You add layers that are then trained to predict outputs while interpreting RISC instructions?
This reminds me of C++ Template Metaprogramming (TMP) (it's Turing-complete so people wrote "games" that run at compile time). The same with TypeScript types, I think someone got DOOM running entirely in the TypeScript type system. An interesting tech demo but entirely superfluous.
It is a great example to refute the common misconception that LLMs are inherently non-deterministic. I see people repeating this a lot and being able to point at a CPU implemented within one is a very strong counter-example.
> It is a great example to refute the common misconception that LLMs are inherently non-deterministic.
No it isn't.
Realistically the opposite is true in practice and LLMs are really non-deterministic beyond simplistic experiments such as this in useful activities in production. Even with hacks such as "setting temp=0" (which no-one seriously does) to "achieve determinism", was one of the most common misconceptions ever propagated by "AI engineers".
So using tiny slop experiments aren't really useful as a "strong counter-example". In fact this is a very weak one.
> Because a CPU inside a language model clearly wasn't cursed enough, I decided the prudent thing would be to give it a filesystem too.
Love the spirit :)
Not only can it solve the halting problem, but also the ignition problem!
Anyways, what I’m trying to say is I want evil qwen to be my CPU.
I wonder how this squares with the recent "NNs are already clean" interpretability finding, which may have also been the impetus for this kind of research. Similar to their work on CNNs, maybe residual-steering is overkill, there might be just a single neuron that "cleanly" corresponds to "evilness" that you can turn up/down?
In a few words, they found a way to map residual stream activations (so, a vector) into a few interpretable sparse concepts. They also had the converse maps, that translate concepts into activations.
Golden Gate claude was obtained by artificially increasing the activations in the direction of the 'Golden Gate' concept. It was not neuron surgery.
A fun and important thought experiment about the the space of nerual networks vs the image of current training methods.
Most discourse conflates neural networks and trainable neural networks.
They implemented wasm in transformer with cpu decoding for the wasm part.
I’ve never wanted to learn how LLMs work more than reading this. Fascinating.
Does divagr get paid to do wild things with LLMs? Just curious if they’re being compensated for these wild experiments
So like... can't you use a computer as a computer without doing all that roundabout stuff? Like if you want to integrate a LLM with hard computation there's probably a more efficient way to do that.
Or is this building towards "but can it run DOOM?" :)
Given that models learn to use residual connections optimally, it can’t be hard for them to steer around the fixed computation layer when it’s not needed.
I’m sure giant models invest a few layers in reproducing these functions, but surely smaller models can benefit?
Or maybe ill join your discord and hopefully we can riff it out to get there faster
there's no free lunch here ever hence the cursed
This reminds me of C++ Template Metaprogramming (TMP) (it's Turing-complete so people wrote "games" that run at compile time). The same with TypeScript types, I think someone got DOOM running entirely in the TypeScript type system. An interesting tech demo but entirely superfluous.
We have gone down this path already in [0] and this is the same thing but overly complicated. This reason is "because LLMs".
Why?
[0] https://arxiv.org/abs/1410.5401
No it isn't.
Realistically the opposite is true in practice and LLMs are really non-deterministic beyond simplistic experiments such as this in useful activities in production. Even with hacks such as "setting temp=0" (which no-one seriously does) to "achieve determinism", was one of the most common misconceptions ever propagated by "AI engineers".
So using tiny slop experiments aren't really useful as a "strong counter-example". In fact this is a very weak one.