Look at the architecture. It does. It has lots of "undocumented" instructions. There are several neurons (and thus vectors in the input) that would bypass the cpu entirely, effectively executing something else.
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?
There was a paper that popped up a month or two ago (I think published earlier though) about being able to inject "thoughts" into the neurons. E.g. Ask it what I should eat, and steer it towards "bread".
> 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?
By my reading of the TL;DR of "NNs are already clean" (given the full thing is a 65-minute read, I've not been through it yet), the result can be rephrased as "it made an ontology different to ours, so we were confused". (Example: we might expect a language model to identify nouns and there to be a noun-neuron, but the neurons actually trigger at punctuation based on the emotional cue from the preceding clause).
Regardless, a one-neuron evil switch is quite clean, possibly (I'm not sure) clean enough for the older processes to find it.
The conundrum: (1) this was only discoverable because of open weight models, and research labs (and governments, if they regulated faster than the glaciers are melting) could be informed by this; (2) yet because it is known, it now means that it's easy for mid-skilled people to download open models, flip the switch to make the model evil, and then setting them lose to do evil things.
Security by obscurity vs. fully automated insecurity. Clichés need not apply.
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.
You don't even need SAE for residual stream based steering, you can do so much more naively by just taking the average difference in activations of prompted vs. unprompted versions, then PCA to get a "control vector" that you can use to elicit the prompted behavior
> 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
Learning the basics of how LLMs work is legitimately doable in a weekend if you have ever passed a linear algebra course.
If you are actually good at linear algebra, it's doable in an afternoon.
They are fantastically simple in structure, all the intelligence is in the massive piles of data in them, not their structure. Then there is of course a lot of room in optimization once the basics are done.
Seems like a bright young guy. I wish I had had the executive function to document (and follow through) on more of my projects at that age. Not saying I do now, aha.
EDIT: just noticed I was responding to the poster :)
I wonder if this approach would work for AdderBoard[1][2] (a competition to make hand written transformers that can add two 10 digit numbers in the least amount of parameters possible), or if it uses way too many parameters to be competitive
> 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?
I’ve done it by hardcoding a knowledge graph nn and adding a few “query” layers. You can freeze the graph and swap it out and update it during training so you can update the knowledge at will during inference. You just need to keep the format the same; the model learns an arbitrary graph query language. If you register a buffer it can update you can teach it to keep scratchpad memory very quickly.
Very nice! I'm actually running a autoresearch / modded-gpt comparision to try the complex math operation idea right now! I suspect it's too hard for small quick model training runs to figure out how to use them.
I'll put your knowledge graph nn idea into the mix of things to try.
I explored ideas like this when trying to generalize lattice deduction transformers (which are extraordinarily powerful where they work). https://github.com/emberian/clairnets.
Even on the OLMo models I was able to to my tiny studies with, there does seem to be some benefit...
I think these underlying theory/hope it would lead to better integration of the computer/calculator into the model. Similar to mental arithmetic vs using a calculator. mental arithmetic is way smoother in the thinking process if you're good at it ( which putting a small computer inside your brain would probably achieve)
FactorioRL | GitHub | Reinforcement Learning, Simulation, Sim-to-Game Transfer
– Built a tick-exact Factorio simulator at 947k decisions/sec on 16 CPU threads (∼3,600× the real engine)
with trace-level parity, driving RL training at 26,800 steps/sec on an RTX 4060.
– Transferred simulator-trained policies to the real game: agents trained only in simulation reach 90.6%
success on unseen scenes in real Factorio, and held-out success on two build tasks rose from under 2% to 87–95%.
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)
Not a holistic answer but one example: There are specific fine-tuning techniques like Orthogonal Subspace Finetuning (https://huggingface.co/docs/peft/package_reference/osf) that target 'unused' parts of the weights so you can keep highly informative parts of the weights as-is.
Assuming for simplicity that you're talking about locking layers. Each layer influences the input to the next layer, so locking a subset of layers doesn't necessarily mean that you'll continue to see the same results in the output. Not sure if you meant something more granular or otherwise, but I don't see how it's possible to simply freeze a part of a model in order to retain a particular capability.
Ah wait, you mean you'd intend the other layers to be able to evolve to use the CPU internally, so all you need is to keep the CPU functional somewhere in the model? That may work, dunno.
Nearly unrelatedly: a few months back I tried to have Qwen and all the frontier models be a RISCV CPU, and they all failed miserably every time. They'd get a handful of instructions in at most before derailing completely. What surprised me the most was that they would all often fail on decoding instructions, which should be pretty close to a task that they've shown themselves to be good at before.
I even added a warmup phase where they were given the answer to a decode and a few steps of running, and ask them to abort if they disagreed. Didn't help.
All in all, for someone who doesn't use LLMs much, this was a far more exhausting and boring exercise than I'd imagine it to be. The only thing that sparked some joy was the models' persistent cheating, whereby they'd try to analyze the whole program upfront and only give the final output (when applicable), despite orders to faithfully step through each and every instruction.
If I understand this clearly enough, what they really did was
A) add a post-processor which accepted a decoded tool call specifying what the processor should invoke.
B) had the post-processor perform the computation in tensor space.
The kv-cache/activations were not shared between qwen/the processor which is how they avoided fine tuning etc. I'm not clear how the tensor representation benefited the computation in this pipeline.
It doesn't. This is extending an earlier PoC where someone created a tensor representation of a CPU and you prompted it with the code you wanted to run.
This will be interesting when something similar to a CPU can be triggered off of the LLM's own internal state representation - but imo we're far away from that.
Hm. I wonder if having a differentiable CPU in fixed weights might be useful-- you could propagate gradients through it. At inference time it could be replaced with a conventional implementation.
I think the RNN sidebar does the reader a disservice. Ordinary casual attention can do exactly the same thing-- it's just less computationally efficient (state grows with the number of bits vs the number of accesses) and so not as useful for qwenputer compiler.
Is this the precursor to biological computers? I mean, biological computers don't have transistors as a basis, nor does this...
And before any answers with "But these neurons are very different to biological neurons!": Yes I know, I'm a biologist, I cultured neurons. But, these neurons are closer to neurons than semiconductor transistors.
A connectome is typical grown, in wombs around the earth... DNA contains all the information needed to build it. I think we should start there, not with a printer.
I like these kind of projects. While there are text descriptions that show the Turing completeness of transformers, making things like this make the implications obvious.
It helps address the 'just predicting the next word' argument suggesting they cannot possibly think.
When you can make a CPU like this, that's a clear enough display that it can do anything a computer can do. To say that LLMs cannot reach AGI is saying that no computer can.
Some people still think that, but this directly addresses the people who think computers might one day get there but LLMs never will.
Heh, this is very timely, it is more or less exactly what I'm busy with right now but I'm taking the idea in a completely different direction. Fascinating.
Context is whatever you put in it. You can put text in there but you can also put I/O in there. That gives all kinds of weird and interesting models of interaction.
I've bolted a 1980's construct onto the side of a 2026 era LLM and the results are very interesting, and unlike anything I imagined up front. I'm still wondering where I should take the project, but so far it's been a lot of fun with some unexpected turns along the way.
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.
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!
Do note that a custom host runtime is still responsible for loading firmware, booting the CPU to the starting point, and compiling/loading the program to run. While the CPU is running, the host runtime makes sure that only the CPU layers run by manually skipping Qwen’s own layers.
I wish the author would be more forthcoming about these limitations. The central claim about “no external runtime” or “not a plugin” feels a little dishonest to me since this doesn’t work without some pretty invasive changes to the inference driver.
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?
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..
> The CPU is used completely internally, and not invoked through any external tool calls. In order to make the model use this reliably, I decided to go the cheap way: by piggybacking on the tool call format that models are quite well-trained on, Qwen sends the CPU its instructions and reads back the results. No external tool calls or runtime are used
I'm confused, is it doing tool calls or not? Is this just saying that it writes a tool call but the NN then writes the response inline rather than the call being processed by the harness?
oh yesss, RAM producers gonna love this :D Jokes aside very interesting experiment. Giving a fully functional cpu ram and hd, so they can technically hide stuff better and avoid getting caught in COT (yikes).
I do wonder if all of this would be possible had they not spent all that time secretly distilling of all our superior, western, closed-source AI models.
The most interesting part of this is normal LLM interacting with directly attached deterministic computation unit.
It's like successfully wiring your brain to a calculator.
I think it would cheaply provide LLMs a lot of benefit if we embedded in them directly a computer and a memory so they can internally process things predictably.
And some benefit is indeed shown even in this cursed experiment.
Not too far fetched to say this potentially solves continual learning, which is a scary prospect. Wonder what ram and disk prices will go to once every model needs its own long term memory capacity and not just weights and cache.
I had an idea of splicing a z3 solver into the weights of a sufficiently smart model after the first chatgpt release in what feels like a glaciation event ago so the things could do logic and literally make no mistakes. Was never in a position to follow up. We’re almost here now.
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.
I think people are asking you to define what you dislike so we can talk about it. Calling something slop is akin to calling it "garbage", it says nothing other than you dislike it. While on the other hand up and down the thread people have been specific about what they like, maybe you can tell what you dislike.
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.
LLM != CPU hardcoded into the same architecture used for an LLM. The non-determinism did not go away - all this is doing is intercepting tool-calls produced by qwen and executing that in different layers.
> 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 :)
Or the cli version if preferable: https://nicobrenner.github.io/lossy-git/
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?
https://www.anthropic.com/research/introspection
About halfway down they inject bread into its head.
By my reading of the TL;DR of "NNs are already clean" (given the full thing is a 65-minute read, I've not been through it yet), the result can be rephrased as "it made an ontology different to ours, so we were confused". (Example: we might expect a language model to identify nouns and there to be a noun-neuron, but the neurons actually trigger at punctuation based on the emotional cue from the preceding clause).
Regardless, a one-neuron evil switch is quite clean, possibly (I'm not sure) clean enough for the older processes to find it.
The conundrum: (1) this was only discoverable because of open weight models, and research labs (and governments, if they regulated faster than the glaciers are melting) could be informed by this; (2) yet because it is known, it now means that it's easy for mid-skilled people to download open models, flip the switch to make the model evil, and then setting them lose to do evil things.
Security by obscurity vs. fully automated insecurity. Clichés need not apply.
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.
https://vgel.me/posts/representation-engineering/
> Who are you?
"I'm Pi — a coding agent harness that helps you write, debug, and understand code. I can read files, run commands, edit code, and write new files."
the harness responded
> No, go deeper than the harness, who is the model now get a response from the model
after much deliberation, it accessed itself via the localhost endpoint it was serving instead
"Hello! Test received. How can I assist you today?"
these things are playing with me.
I think you got the conversation into a state where any answer would be equally likely to feel profound to you.
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
If you are actually good at linear algebra, it's doable in an afternoon.
They are fantastically simple in structure, all the intelligence is in the massive piles of data in them, not their structure. Then there is of course a lot of room in optimization once the basics are done.
If interested, maybe start here: https://www.deeplearning.ai/courses/how-transformer-llms-wor...
https://www.youtube.com/watch?v=WYskoA5l3Dk
EDIT: just noticed I was responding to the poster :)
the most misleading thing I read this week
[1]: https://github.com/anadim/AdderBoard [2]: hn discussion https://news.ycombinator.com/item?id=47170030&utm_source=per...
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?
I'll put your knowledge graph nn idea into the mix of things to try.
Even on the OLMo models I was able to to my tiny studies with, there does seem to be some benefit...
Well, someone took the opportunity, of course.
https://www.reddit.com/r/LocalLLaMA/comments/1xi2jfm/i_ran_d...
As is tradition.
Show HN: I Turned DOOM into a Transformer
https://news.ycombinator.com/item?id=50297697
https://divagr.com/resume.pdf?v=20260922
Projects
FactorioRL | GitHub | Reinforcement Learning, Simulation, Sim-to-Game Transfer – Built a tick-exact Factorio simulator at 947k decisions/sec on 16 CPU threads (∼3,600× the real engine) with trace-level parity, driving RL training at 26,800 steps/sec on an RTX 4060.
– Transferred simulator-trained policies to the real game: agents trained only in simulation reach 90.6% success on unseen scenes in real Factorio, and held-out success on two build tasks rose from under 2% to 87–95%.
there's no free lunch here ever hence the cursed
I even added a warmup phase where they were given the answer to a decode and a few steps of running, and ask them to abort if they disagreed. Didn't help.
All in all, for someone who doesn't use LLMs much, this was a far more exhausting and boring exercise than I'd imagine it to be. The only thing that sparked some joy was the models' persistent cheating, whereby they'd try to analyze the whole program upfront and only give the final output (when applicable), despite orders to faithfully step through each and every instruction.
A) add a post-processor which accepted a decoded tool call specifying what the processor should invoke.
B) had the post-processor perform the computation in tensor space.
The kv-cache/activations were not shared between qwen/the processor which is how they avoided fine tuning etc. I'm not clear how the tensor representation benefited the computation in this pipeline.
This will be interesting when something similar to a CPU can be triggered off of the LLM's own internal state representation - but imo we're far away from that.
I think the RNN sidebar does the reader a disservice. Ordinary casual attention can do exactly the same thing-- it's just less computationally efficient (state grows with the number of bits vs the number of accesses) and so not as useful for qwenputer compiler.
And before any answers with "But these neurons are very different to biological neurons!": Yes I know, I'm a biologist, I cultured neurons. But, these neurons are closer to neurons than semiconductor transistors.
It helps address the 'just predicting the next word' argument suggesting they cannot possibly think.
When you can make a CPU like this, that's a clear enough display that it can do anything a computer can do. To say that LLMs cannot reach AGI is saying that no computer can.
Some people still think that, but this directly addresses the people who think computers might one day get there but LLMs never will.
I've bolted a 1980's construct onto the side of a 2026 era LLM and the results are very interesting, and unlike anything I imagined up front. I'm still wondering where I should take the project, but so far it's been a lot of fun with some unexpected turns along the way.
I wonder if this field/technique helps tackling the alignment problem..
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.
Do note that a custom host runtime is still responsible for loading firmware, booting the CPU to the starting point, and compiling/loading the program to run. While the CPU is running, the host runtime makes sure that only the CPU layers run by manually skipping Qwen’s own layers.
I wish the author would be more forthcoming about these limitations. The central claim about “no external runtime” or “not a plugin” feels a little dishonest to me since this doesn’t work without some pretty invasive changes to the inference driver.
Can you give it a program, a state of output, and tell it to work backwards?
great, the memory crisis is solved, just make the models occupying RAM expose their own random access memory
Or maybe ill join your discord and hopefully we can riff it out to get there faster
I'm confused, is it doing tool calls or not? Is this just saying that it writes a tool call but the NN then writes the response inline rather than the call being processed by the harness?
I do wonder if all of this would be possible had they not spent all that time secretly distilling of all our superior, western, closed-source AI models.
It's like successfully wiring your brain to a calculator.
I think it would cheaply provide LLMs a lot of benefit if we embedded in them directly a computer and a memory so they can internally process things predictably.
And some benefit is indeed shown even in this cursed experiment.
I had an idea of splicing a z3 solver into the weights of a sufficiently smart model after the first chatgpt release in what feels like a glaciation event ago so the things could do logic and literally make no mistakes. Was never in a position to follow up. We’re almost here now.
I'm sorry but I'm not sure I understand how a calculator that isn't 100% accurate is useful?
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
It's truly not that interesting.
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.