alt Hacker NewsWhat is Nueralese and Why is it Bad
Comments
If the model can understand neuralese why can it not convert it into English for monitoring or review purposes?
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
Recurrent depth and chain-of-thought are two completely different concepts. In the former approach, the output of the layer gets is rerouted as input to the same layer, potentially several times. This output/input is a fixed width times sequence length real-valued representation; it is not comparable with output tokens.
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.
Title is misspelled -- "ue" for "eu". In a neologism, some nuisance.
Lesswrong is just so comically collectively incorrect.
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.
We have complete access to every "neuron" and "synapse" (crude analogies...) of these models, so in theory we don't need to be so reliant on CoT traces right? I say this not to minimize the difficulty of interpreting raw activations, but I'd expect a huge amount of research to be focused on it. CoT could be obscured by a model outputting language that looks innocuous but encodes actual hidden meaning. Presumably raw activations would be impossible for a malicious model to obscure in this way.
Scott Alexander article from 2022, "ELK And The Problem Of Truthful AI"
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.
A response to this, which I think is reasonable. Its a bit of a fuzzy line, but the likelihood there is "magical maniacal planning" happening here is unlikely (about as unlikely as that planning happening in hidden vector states between layers).
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).
This reads like typical LW FUD.
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.
It's not only that it'll become more difficult to monitor a single LLM but that also all the instances share exactly the same "collective unconsciousness". Like you develop some paranoid gibberish fantasy language that over time only you understand - except that there are a million copies of you and all of them understand every single nuance of your gibberish.
>If the recurrence step is no longer in English, we can no longer monitor intent, and can only observe whether a model is safe through behavior.
So, you mean, like another human person?
What about the trend of summarizing or eliding reasoning from the visible model response, ostensibly to make distillation by competitors harder?
This is intriguing because it reveals that LLM hype merchants are actually at least as afraid of AGI as critics; they too would rather not have AGI if the AI achieves it in ways they cannot scrutinise, when it is clear that for any AI to achieve a level of intelligence greater than humanity, scrutiny in human terms cannot really be generally possible.
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)
The problem is this entire argument depends on “chain of thought” being a faithful representation of a model’s actual plan. That doesn’t need to be the case. Thinking tokens are just tokens at the end of the day.
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.
My understanding is, that we do not know if the chain-of-thought actually matters in the way we assume for the result.
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.
Can we take a step back here? Both OpenAI’s and Anthropic’s models think in encryptedese, and it seems thoroughly absurd to think that the entire world should trust those two companies to adequately monitor the plaintext or, for that matter, to have their monitoring systems aligned with what is actually good for the world.
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.
surely Neuralese interpreters can be made that turn the chain of numbers into an English description?
Keep in mind that LessWrong a community that 1) ostensibly wants to slow down AI so we can prepare for it, and 2) in actuality, wants to make the Bay Area EA community a kind of guild that controls everyone's use of AI.
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.
I don't think their architectural view of CoT works is correct is it?
The "neuralese" in a looped model is a residual stream vector. We already know how to decode these vectors into tokens; it seems pretty straightforward that a model could be built to decode a series of residual stream vectors into a CoT stream[0], or a classifier could be built to detect residual stream vectors that correspond to scheming.
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.
Sebastian Raschka put out a video on this "bombshell" report (he's more skeptical)
What exactly prevents anyone worried about this to build an LLM that can decode the neuralese to english and use it to monitor what the model is doing?
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?
The argument is that chain-of-thought without "tokens" would remove a major interpretability and model intent control pane. This is definitely borne out in the OpenAI's report on the huggingface attack; they had turned of CoT monitoring for those jobs, and claim that they could have (would have?) prevented the behavior had they been monitoring it. They've changed their internal policies to always monitor CoT.
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.