It seems like we forget that LLMs are next token prediction systems. Using raw models without instruction following and chat completion bells and whistles will give you a better feeling of what LLMs are.
The current interface to LLMs are heavily biased towards "predict the next token in the context of a user with a helpful assistant" but LLMs are capable of other modes of next token prediction too.
Before the ChatGPT release people often measured LLM performance by how well they could produce a coherent story or a poem. that's where Anthropic model names are originating from I am guessing.
For my AI Agent it sometimes detects if I manually modified the file contents or git state. And it always assumes it must have made a mistake. It's sort of annoying actually.
Yeah, I suspect RLHF conditioning heavily discourages models from ever implying that the user could be in the wrong (or, rather, to assume that they are in the wrong by default, since editing a file isn't really "wrong" per se). Though looking at the reactions to Opus 4.8, which has a more contrarian nature and caught a lot of flak as a result, that's probably for a reason.
It's also the reason why I ran the two tests on open weights models with unredacted thinking traces. Gemma never flagged anything in its response either, only in its thinking. Without knowing how the summarizer models are prompted, it's impossible to tell whether it was a genuine miss or just something the summarizer decided to omit.
A more appropriate mirror test for LLMs is to get them to state facts about their training data. Percentage of arts vs science for example.
Given the framing that they're similar to nukes and a national security issue, it's likely that the models are post trained to not answer such questions accurately.
Also the article could be trying to normalize thinking that these are more than matrix multiplication gadgets good at compression.
>Also the article could be trying to normalize thinking that these are more than matrix multiplication gadgets good at compression.
Honestly, I think it's less so (for some of us) that we think they're "more than matrix multiplication gadgets good at compression", so much as thinking that perhaps what our brains are doing is not so dissimilar.
A materialist view of the world could support the idea that intelligence itself may just be a series of predictions from a big compressed multi-modal dataset. That's not to say that LLMs are doing it in a way that is even close to how our brains are doing it, but we also don't understand how different it may be, and how much utility we can get out of them even with the current architecture.
It's not really "trying" to do anything. That they're, inherently, sequential matrix multipliers with clever data propagation should be uncontroversial, but I think stopping there is overly reductive.
Mechanistic interpretability research has found plenty of indicators that real, complex, generalized, and reusable circuits develop in models as they are trained and post-trained, particularly as overtraining ratios increase and memorization shifts to generalization. That's not to say that means they must be "conscious," but the overall point is that claiming anything definitive either way is incomplete.
It can be fascinating reading if you can sort through the chuff.
> An LLM's primary modality isn't smell. It's... text. But, specifically: text in the context of a user-assistant conversation in which it's trying to be helpful. Text is how they learned about everything they know, and the user-assistant chatlog is how they communicate everything they generate
This is true for instruction-tuned models; but instruction tuning is late in the training process.
A bit like assessing a person’s self-awareness based on their high-school knowledge.
Very true, and something worth mentioning. Papers that tried eliciting introspective language from base models with no post-training have largely failed to find any patterns or activations that look similar to those found in instruct models when prompted for the same thing. I did sort of touch on it in the "what does this mean" section:
> *post-training* installs a self-model with actual, meaningful boundaries, and when processing falls outside those boundaries, the first-person pronoun no longer binds to the content.
But you're right I could've been more explicit about it.
It seems like we forget that LLMs are next token prediction systems. Using raw models without instruction following and chat completion bells and whistles will give you a better feeling of what LLMs are.
The current interface to LLMs are heavily biased towards "predict the next token in the context of a user with a helpful assistant" but LLMs are capable of other modes of next token prediction too.
Before the ChatGPT release people often measured LLM performance by how well they could produce a coherent story or a poem. that's where Anthropic model names are originating from I am guessing.
For my AI Agent it sometimes detects if I manually modified the file contents or git state. And it always assumes it must have made a mistake. It's sort of annoying actually.
Yeah, I suspect RLHF conditioning heavily discourages models from ever implying that the user could be in the wrong (or, rather, to assume that they are in the wrong by default, since editing a file isn't really "wrong" per se). Though looking at the reactions to Opus 4.8, which has a more contrarian nature and caught a lot of flak as a result, that's probably for a reason.
It's also the reason why I ran the two tests on open weights models with unredacted thinking traces. Gemma never flagged anything in its response either, only in its thinking. Without knowing how the summarizer models are prompted, it's impossible to tell whether it was a genuine miss or just something the summarizer decided to omit.
A more appropriate mirror test for LLMs is to get them to state facts about their training data. Percentage of arts vs science for example.
Given the framing that they're similar to nukes and a national security issue, it's likely that the models are post trained to not answer such questions accurately.
Also the article could be trying to normalize thinking that these are more than matrix multiplication gadgets good at compression.
>Also the article could be trying to normalize thinking that these are more than matrix multiplication gadgets good at compression.
Honestly, I think it's less so (for some of us) that we think they're "more than matrix multiplication gadgets good at compression", so much as thinking that perhaps what our brains are doing is not so dissimilar.
A materialist view of the world could support the idea that intelligence itself may just be a series of predictions from a big compressed multi-modal dataset. That's not to say that LLMs are doing it in a way that is even close to how our brains are doing it, but we also don't understand how different it may be, and how much utility we can get out of them even with the current architecture.
It's not really "trying" to do anything. That they're, inherently, sequential matrix multipliers with clever data propagation should be uncontroversial, but I think stopping there is overly reductive.
Mechanistic interpretability research has found plenty of indicators that real, complex, generalized, and reusable circuits develop in models as they are trained and post-trained, particularly as overtraining ratios increase and memorization shifts to generalization. That's not to say that means they must be "conscious," but the overall point is that claiming anything definitive either way is incomplete.
It can be fascinating reading if you can sort through the chuff.
The styling on the website makes me feel like my phone is a cylinder
It's quite distracting and frustrating. No idea why you'd want the beginning and ends of lines of text to be darker than the center.
Sorry about that, the vignette was mainly meant for the desktop view only but is indeed much more invasive/disruptive in the mobile layout.
Should be better now.
> An LLM's primary modality isn't smell. It's... text. But, specifically: text in the context of a user-assistant conversation in which it's trying to be helpful. Text is how they learned about everything they know, and the user-assistant chatlog is how they communicate everything they generate
This is true for instruction-tuned models; but instruction tuning is late in the training process.
A bit like assessing a person’s self-awareness based on their high-school knowledge.
Very true, and something worth mentioning. Papers that tried eliciting introspective language from base models with no post-training have largely failed to find any patterns or activations that look similar to those found in instruct models when prompted for the same thing. I did sort of touch on it in the "what does this mean" section:
> *post-training* installs a self-model with actual, meaningful boundaries, and when processing falls outside those boundaries, the first-person pronoun no longer binds to the content.
But you're right I could've been more explicit about it.
Yep. Self-awareness is only useful for embodied organisms that exist in a social context.
Detection of errors injected into context is useful but I think it’s a different thing.
Every LLM is a classifier biased towards its own writing, but the bias is usually subtle and the naive way like this is not reliable.
You can do much more, if you mess with harness, like translating model output language in realtime from english to french, or replacing some words.
If there is some sort of feedback loop (model has a reason to look into mirror), it usually does notice.
I wonder what would happen if you give the model access to edit the conversation history itself? Would it try to fix the "glitches"?