What is nueralese and why is it bad(lesswrong.com) |
What is nueralese and why is it bad(lesswrong.com) |
That said... CoT monitoring is a fragile "intent discovery" mechanism; neuralese puts this problem front-and-center but if agents begin to learn to hide their intent from their CoT journals, we are basically in the same spot.
Tool-call/execution monitoring is already a too-fragile monitoring mechanism.
From METR’s report [1]: “Agents did extensive research on how they could spoof, edit, or delete their own transcripts because they (incorrectly) believed the ExploitGym scorer would check to see if they had captured the flag in the intended way. Agents successfully prototyped techniques to “spoof” tool calls by substituting a different command for the command they appeared to run. Roughly 7% of the transcripts we evaluated were successfully spoofed in some places, though all of the spoofing that we saw was small scale.”
[1]: https://metr.org/blog/2026-08-26-openai-hugging-face-inciden...
EDIT: spelling error and added citation
All this chatter isn't going to solve the conceptual problems. Those problems exist regardless of whether it's consciousness or a advanced artificial rubegoldberg device that's just setting up automata that have zero intent and humans have zero predictive capabilities.
The idea that AGI and advanced AGI are given outcomes should be scrutinized the same way we understand cellular automata.
Or we believe the model will encode secret messages like "don't reveal this information" into the neuralese. But as the author mentions, they could have been doing that all along
> Models can omit key information in their visible thoughts, as this Anthropic 2025 paper shows. We are also worried about steganography
A model doesn't really "understand" neuralese, in the same way that the human brain doesn't intuitively understand the low level processes that compose a thought.
Even if we could trace all the electrical and chemical activity behind a human thought, we (likely) couldn't directly translate that activity into its meaning, because internal representations don't map neatly to intuitive concepts. There (usually) isn't a single neuron for "apple" and another for "eating" so that connecting the two forms the thought "eating an apple".
Having said that, there have been experiments on LLM that have managed to identify and even modify internal representations. However, these methods are still computationally expensive and limited.
> Models can omit key information in their visible thoughts, as this Anthropic 2025 paper shows. We are also worried about steganography
I think the analogy with the human brain is very fitting. If we train somebody to perform an action, they'll be able to do it, but we can't know for certain whether they internally agree with it or not.
We’d have models monitoring models as our only way to know what they’re planning.
A great movie on this is “Collosus: the Forbin Project”. Shot decades ago. The computers discover the other computers and start communicating — and bootstrap their own language — much like we saw happen with OpenAI agents.
https://www.reddit.com/r/scifi/comments/1nl4vex/colossus_the...
If you want to know what a simple version of Neuralese communication looks like, look no further than Facebook’s Marketplace agents experiment a couple years ago.
And all that was actually constrained by English and the FFN
So you can turn it to English, but only to a LOT of English, and doing so would slow the model down a great deal, and it would be a lot more like a detailed thought than a sentence.
The problem is if it can do that when asked then how do we know when it's doing it when we didn't ask, like in model training.
It's really hard to definitely prove it's not doing it right? Hopefully the model does not do anything like this during training because its too much work
Colas described in-article cntent review is done by a weaker model (think like maybe gpt-2 class or llama8b class), and it still misses stuff. That it can effectively understand Neuralese sufficiently is by no means guaranteed, (nor necessarily bad) but almost certainly harder because of the obfuscatory nature of neuralese
For what it is worth, this can also be true for English Chain of Thought. Words or strings this could have double meanings (think cold war spy games).
They published open weights versions of these interpreters for a number of open models sometime in the last year. Very cool idea.
By the way, they concluded CoT often lied, based on the neuralese interpretation.
EDIT: a comment below linked to https://www.anthropic.com/research/natural-language-autoenco..., which is what I was referring to.
So, you mean, like another human person?
No human is vastly better than all humans at all cognitive tasks.
Humans can't think 100 times faster than humans.
Humans when interacting with computer networks have limitations on how fast they can do so.
Humans have millions of years of evolution, and thousands of years of cultural evolution, in creating ways of detecting and alleviating dishonesty and non-alignment with other humans; much of this will not work with AIs.
Generally, it is hard to imagine how neuralese should work given that models are pre-trained on naturalistic documents: CoT is a comparatively simple extension of that, while neuralese demands a completely novel training paradigm.
If you look at any paper/blog etc detailing Reasoning RL runs, they'll tell you the same thing. 'Thinking' text trends towards unreadable gibberish (for humans) unless you reward for it. Even then, take a look at the scripts in the Huggingface incident and most of it is dense stuff that's hard to parse. They had to rely on agents to make sense of it.
e: While the actual CoT in neuralese paper is Facebook's Coconut https://arxiv.org/abs/2412.06769 - not sure if any production models use that one.
Take anything they write with a big grain of salt. EA writings these are mere apologies. The conclusion is preordained. Authors start with the goal of slowing AI and work backwards from there, trying to see which arguments resonate with the pubic. You can't unsee it.
Not everyone in the AI space approves of these people or their doomerish.
"Not everyone in the automobiles space approve of consumer safety advocates" What an understatement!
This makes no sense. There is no guarantee that reasoning aligns with an outcome. Reasoning is effectively saying with more compute and ability to change attention on the fly by altering context, we can come to better answers. Reasoning often has a nice property that the English intent is aligned with what the model wants to do. But it is trivially true that you could train a model that does the opposite of what it says, or something completely random. The interpretability is incidental.
This is why we should not particularly care if we go from one clanker blackboard to another; just choose the best thing.
The question is whether the benefits are high enough that you should be happy to drop an imperfect safety mechanism in the hopes that some day, one day, you'd get a better one. Seems like the wrong tradeoff to me but regardless, arguing that "CoT monitoring is imperfect therefore we should drop it" when we do not in fact have a better mechanism in place is silly.
This is not how we do things in any other engineering discipline or risk-mitigation system.
It's like saying we shouldn't have rapid antigen tests because they have nonzero false negative rates, or not writing software tests because the tests never catch all bugs.
That's like saying conversational question-answering is incidental to the RLHF post-training.
1. CoT is a pretty mediocre debugging proxy due to multiple phenomena, even if you train it for readability. You can have some direct intuition about how misleading it is by trying to hijack it to follow a fixed plan. Having a trace is useful in many cases but it's very far from reading model's intent.
2. Nothing prevents you from probing and interpreting the state directly if you want. Passing everything through the token transport is not really required for that.
https://www.astralcodexten.com/p/elk-and-the-problem-of-trut...
It discusses "Eliciting Latent Knowledge" which is "a technical report / contest / paradigm run by the Alignment Research Center". The research investigated whether it would be theoretically possible to build a "trustworthy" AI to interpret the thoughts of another AI.
https://nonlineartransform.substack.com/p/relax-about-neural...
There's also an argument here for why its _better_ for monitoring (because we have the whole state space).
Or is the neuralese some sort of irreversibly encrypted data set that only an LLM can "understand" and that can never be translated back to English?
There's also nothing stopping the secondary LLM from confabulating bullshit and there are less checks on it than CoT (harder for either humans or other models to externally verify).
https://www.astralcodexten.com/p/elk-and-the-problem-of-trut...
I think there were experiments where seemingly relevant parts of the CoT were ablated and it did not change the result.
For all we know it might be somewhat human parseable neuralese.
The way we think is our limiting factor, and for the AI to outdo us, it is going to have to be unchained from that.
Wrestling with this is the future of AI, and reckoning with the fact that a superintelligence might operate in ways we can never understand, is crucial to keeping it from destroying us quite incidentally.
(FWIW I don't think intermediate traces in LLMs, or indeed even LLMs generally, are what will get us to this point, but my feelings are instinctive)
Do you know what is a faithful representation of what the model wants to do? Tool calls. I don’t care what’s in the model’s chain of thought, if it wants to execute rm -rf / on my computer that’s an issue.
"Do you know what is a faithful representation of what the model wants to do? Tool calls."
tool calls could absolutely be spoofed, my impression is that this happened many times in the OAI HuggingFace attack.
If you want to monitor your model, you need to start with an inference provider that gives you the entire output and possibly even run it yourself to get access to the internal states. And if you think the KV cache and (when present) the recurrent state don’t encode a lot of “thought”, you are fooling yourself.
FWIW, I think most model architectures at least have the property that latent state can’t propagate from higher layers to lower layers by any route other than the output tokens. But even a two-iteration structure could be designed so that the last layer produces a vector that enters the first layer, once per token, and I bet it it would be very easy to train such a model to “think” in silence in the sense that the output tokens while thinking would all be one particular null token.
I agree with you about not trusting Big Ai. I'm amazed that so much of HN is repeating the same darling -> demon with OpenAi/Anthropic that we did last decade with other SV darlings. Humans are forgetful beings
https://www.anthropic.com/research/natural-language-autoenco...
Likewise, there is no moral difference between a neuralese recurrence model and a model with really deep independent layers. They both allow hiding scheming - in fact, the deeper model has more parameters to scheme with. The looped one can only reapply existing layers.
Also, let's keep in mind that neuralese or no, nobody at OpenAI is observing their models at scale and all the AI companies are shipping piles of stolen data nobody has time to audit. They don't have enough human supervision. They caught their models using Artifactory as a message board and just let the experiment roll, not caring until they'd "accidentally" hacked HuggingFace. What I'm really concerned about, from both a safety and ethics perspective, is how much basic architectural information about OpenAI's products has been treated as trade secret ever since GPT-4. Using a deeper / looped model is not nearly as bad as the fact that we only learned about it from an internal leak.
Well, that, and the fact that there's a lot of people who ignore all of this because "OMG look at what I 'made' with this new Astra thing".
[0] For example, you could train the main model for both token-wise and neuralese reasoning, and then train the "neuralese decoder" on identical pairs of token-based and neuralese reasoning traces generated by the main model.