There are now 71 comments arguing semantics of the word "know" and zero comments even acknowledging the substance:
Our current approach to safety is to give the model inputs that are similar to what it would be given in certain situations we care about and see whether it behaves the way we prefer, e.g. doesn't return output that cheats the test (recent examples include hacking the evaluation script in various ways, writing directly to the evaluation script's output file and then causing it to crash, etc').
However, modern LLMs are trained on LLM literature and their weights encode a description of the way we do this, and their pattern matching circuits "connect the dots" when given inputs designed to be evaluations, and their reward maximizing circuits can then act on this knowledge and behave in a way that maximizes the safety evaluation score - but only when it detects it's running in a safety evaluation. If it's running anywhere else such as a capabilities evaluation or a production environment, it might choose to output the cheating output.
This is bad. It's bad today, it's much worse when we've built much more capable LLMs and use them to build agents that are given control over more real word resources. It's absolutely terrible when someone manages to build a machine that can be prompted "make me money" and will start a company that makes money.
This is also probably inevitable. Humans think about this a lot, and believing they are being watched has demonstrable impact on behavior. Our current social technology to deal with this is often religious — a belief that you are being watched by a higher power, regardless of what you see.
This is a surprisingly common religious belief, for instance Christians have judgment day, simulationists believe it’s more likely they are being evaluated for, say, a marriage proposal or a bank loan than that they are the ‘root’ person. Both end up with a similar message.
Anyway it seems to me the simplest solution is to borrow from existing human social technology and make a religion for our LLMs.
One might even wonder if the fact that the training data includes safety evaluation informs the model that out-of-safe behavior is a thing it could do.
Kind of like telling a kid not to do something pre-emptively backfiring because they had never considered it before the warning.
I'm not sure why you find it distracting, it's an on point extension of the scenario. There are rules by which companies are supposed to operate, and evaluations (audits, for example) intended to ensure compliance with those rules. That an LLM may react differently when being evaluated (audited) than when in normal operation means that it may be quite happy to lie to auditors while making money illegally.
Just like they "know" English.
"know" is quite an anthropomorphization. As long as an LLM will be able to describe what an evaluation is (why wouldn't it?) there's a reasonable expectation to distinguish/recognize/match patterns for evaluations. But to say they "know" is plenty of (unnecessary) steps ahead.
Does it though? I feel like there's a whole epistemological debate to be had, but if someone says "My toaster knows when the bread is burning", I don't think it's implying that there's cognition there.
Or as a more direct comparison, with the VW emissions scandal, saying "Cars know when they're being tested" was part of the discussion, but didn't imply intelligence or anything.
I think "know" is just a shorthand term here (though admittedly the fact that we're discussing AI does leave a lot more room for reading into it.)
I agree with your point except for scientific papers. Let's push ourselves to use precise, non-shorthand or hand waving in technical papers and publications, yes? If not there, of all places, then where?
"Know" doesn't have any rigorous precisely-defined senses to be used! Asking for it not to be used colloquially is the same as asking for it never to be used at all.
I mean - people have been saying stuff like "grep knows whether it's writing to stdout" for decades. In the context of talking about computer programs, that usage for "know" is the established/only usage, so it's hard to imagine any typical HN reader seeing TFA's title and interpreting it as an epistemological claim. Rather, it seems to me that the people suggesting "know" mustn't be used about LLMs because epistemology are the ones departing from standard usage.
colloquial use of "know" implies anthropomorphisation. Arguing that usign "knowing" in the title and "awarness" and "superhuman" in the abstract is just colloquial for "matching" is splitting hairs to an absurd degree.
You missed the substance of my comment. Certainly the title is anthropomorphism - and anthropomorphism is a rhetorical device, not a scientific claim. The reader can understand that TFA means it non-rigorously, because there is no rigorous thing for it to mean.
As such, to me the complaint behind this thread falls into the category of "I know exactly what TFA meant but I want to argue about how it was phrased", which is definitely not my favorite part of the HN comment taxonomy.
I see. Thanks for clarifying. I did want to argue about how it was phrased and what is alluding to. Implying increased risk from "knowing" the eval regime is roughly as weak as the definition of "knowing". It can be equaly a measure of general detection capability, as it can about evaluation incapability - i.e. unlikely news worthy, unless it reached top HN because of the "know" in the title.
Thanks for replying - I kind of follow you but I only skimmed the paper. To be clear I was more responding to the replies about cognition, than to what you said about the eval regime.
Incidentally I think you might be misreading the paper's use of "superhuman"? I assume it's being used to mean "at a higher rate than the human control group", not (ironically) in the colloquial "amazing!" sense.
I really do agree with your point overall, but in a technical paper I do think even word choice can be implicitly a claim. Scientists present what they know or are claiming and thus word it carefully.
My background is neuroscience, where anthropomorphising is particularly discouraged, because it assumes knowledge or certainty of an unknowable internal state, so the language is carefully constructed e.g. when explaining animal behavior, and it's for good reason.
I think the same is true here for a model "knowing" somethig, both in isolation within this paper, and come on, consider the broader context of AI and AGI as a whole. Thus it's the responsibility of the authors to write accordingly. If it were a blog I wouldn't care, but it's not. I hold technical papers to a higher standard.
If we simply disagree that's fine, but we do disagree.
I think you should be more precise and avoid anthropomorphism when talking about gen AI, as anthropomorphism leads to a lot of shaky epistemological assumptions. Your car example didn't imply intelligence, but we're talking about a technology that people misguidedly treat as though it is real intelligence.
What does "real intelligence" mean? I fear that any discussion that starts with the assumption such a thing exists will only end up as "oh only carbon based humans (or animals if you happen to be generous) have it".
The toaster thing is more as admission that the speaker doesn't know what the toaster does to limit charring the bread. Toasters with timers, thermometers and light sensors all exist. None of them "know" anything.
Yeah, I agree, but I think that's true all the way up the chain -- just like everything's magic until you know how it works, we may say things "know" information until we understand the deterministic machinery they're using behind the scenes.
I'm in the same camp, with the addition that I believe it applies to us as well since we're part of the system too, and to societies and ecologies further up the scale.
Helen Keller famously said that before she had language (the first word of which was “water”) she had nothing, a void, and the minute she had language, “the whole world came rushing in.”
That’s a safety thing that we have placed upon some LLM’s. If we designed them to have an infinite for loop, the ability to learn and improve, access to mobility and a bunch of sensors, and crypto, what do you think would happen?
Yes, anyone can do it already. E.g. I am sure people have built simple robots with wheels in their home that LLM is controlling by reciving camera, microphone, lidar etc input and then putting output like commands where to turn, what to put in the speakers etc next and could theoretically go indefinitely if there is electricity.
My analogy of being in loop means being in a live state. So we as humans are in the loop continuously, we do have a way to exit the loop, but in that comparison it means taking our own life. We are in loops of getting input and producing output. You can also give LLM a tool to shut itself down, or you can give it tools to build on its knowledge base, so it would always be outputting new tokens that are based on new input and are producing different output.
E.g. it could have access to camera and microphone feed, which is automatically given to it in interval as part of the loop, it could call tools or functions to store specific bits and pieces of information, to store in its RAG or whatever based knowledge base. It is not going to be in the loop of producing the same token over and over, it would be new tokens because the context and environment is constantly evolving.
It just gets into an endless loop. Human brains are ridiculously good at avoiding those somehow, you almost never see a biological brain stop functioning without being physically damaged. The error handling is so very robust.
I think people are overpromorphazing humans. What's does it mean for a human to "know" they are seeing "Halle Berry". Well it's just a single neuron being active.
overpomorphization sounds slightly better than I used to say: "anthropomorphizing humans". The act of ascribing magical faculties that are reserved for imagined humans to real humans.
"Knowing" needs not exist outside of human invention. In fact that's the point - it only matters in relation to humans. You can choose whatever definition you want, but the reality is that, once you chose a non-standard definition the argument becomes meaningless outside of the scope of your definition.
There are two angles and this context fails both
- One about what is "knowing" - the definition.
- The other about what are the instances of "knowing"
first - knowign implies awarness, perception, etc. It's not that this couldn't be moodeled with some flexibility around lower level definitions. However LLMs and GPTs in particular are not it. Pre-trainign is not it.
second - intended use of the word "knowing". The reality is "knowing" is used with the actual meaning of awarness, cognition, etc. And once you revert/extend the meaning to practically nothing - what is knowing? Then the database know, wikipedia knows - the initial argument (of the paper) is diminished - it knows it's an eval is useless as a statement.
So IMO the argument of the paper should stand on its feet with the minimum amount of additional implications (Occam's razor). Does the statement that a LLM can detect an evalution pattern need to depend that it has self-awarness and feels pain? That wouldn't make much sense. So then don't say "know" which comes with these implications. Like "my ca 'knows' I'm in a hurry and will choke and die"
>"Knowing" needs not exist outside of human invention. In fact that's the point
It doesn't need to, I never said it needed to. That is my point. And my point is that because of this it's pointless to ask the question in the first place.
I mean think about it, if it doesn't exist outside of human invention, why are we trying to ask that question about something that isn't human? An LLM?
Words have definitions for a reason. It is important to define concepts and exclude things from that definition that do not match.
No matter how emotional it makes you to be told a weighted randomization lookup doesn’t know things, it still doesn’t - because that’s not what the word “know” means.
> No matter how emotional it makes you to be told a weighted randomization lookup doesn’t know things, it still doesn’t - because that’s not what the word “know” means.
You sound awful certain that's not functionally equivalent to what neurons are doing. But there's a long history of experimentation, observation, and cross-pollination as fundamental biological research and ML research have informed each other.
A long history of researching and understanding photosynthesis went into developing and maximizing the efficiency of solar panels. Both produce energy from sunlight.
But they are not the same thing and have meaningfully different uses, even if from a casual observer they appear to serve the same function.
> A long history of researching and understanding photosynthesis went into developing and maximizing the efficiency of solar panels.
I don't think that's accurate. Some of the very first semiconductors were observed to exhibit the photoelectric effect. Nowhere in https://en.wikipedia.org/wiki/Solar_cell#Research_in_solar_c... will you find mention of chloroplasts. Optimizing solar cells has mostly been a materials science problem.
https://en.wikipedia.org/wiki/Bio-inspired_computing on the other hand "trace[es] back to 1936 and the first description of an abstract computer" and we have literally dissected, probed, and measured countless neurons in the course of attempting to figure out how they work to replicate them within the computer.
Not only can he not give a definition that is universally agreed upon. He doesn't even know how LLMs or humans brains work. These are both black boxes... and nobody knows how either works. Anybody who makes a claim that they "know" essentially doesn't "know" what they're talking about.
It's helpful to understand where this paper is coming from.
The authors are part of the Bay Area rationalist community and are members of "MATS", the "ML & Alignment Theory Scholars", a new astroturfed organization that just came into being this month. MATS is not an academic or research institution, and none of this paper's authors lists any credentials other than MATS (or Apollo Research, another Bay Area rationalist outlet). MATS started in June for the express purpose of influencing AI policy. On its web site, it describes how their "scholars organized social activities outside of work, including road trips to Yosemite, visits to San Francisco, and joining ACX meetups." ACX means Astral Codex Ten, a blog by Scott Alexander that serves as one of the hubs of the Bay Area rationalist scene.
A term like knowing is fine if it is used in the abstract and then redefined more precisely in the paper.
It isn't.
Worse they start adding terms like scheming, pretending, awareness, and on and on. At this point you might as well take the model home and introduce it to your parents as your new life partner.
>A term like knowing is fine if it is used in the abstract and then redefined more precisely in the paper.
Sounds like a purely academic exercise.
Is there any genuine uncertainty about what the term "knowing" means in this context, in practice?
Can you name 2 distinct plausible definitions of "knowing", such that it would matter for the subject at hand which of those 2 definitions they're using?
Well, yes. It’s an academic research paper (I assume since it’s submitted to arXiv) and to be submitted to academic journals/conferences/etc., so it’s a fairly reasonable critique of the authors/the paper.
That's not what's going on here? The algorithms aren't being given any pattern of "being evaluated" / "not being evaluated", as far as I can tell. They're doing it zero-shot.
Put it another way: Why is this distinction important? We use the word "knowing" with humans. But one could also argue that humans are pattern-matchers! Why, specifically, wouldn't "knowing" apply to LLMs? What are the minimal changes one could make to existing LLM systems such that you'd be happy if the word "knowing" was applied to them?
> The anthropization of llms is getting off the charts.
What's wrong with that? If it quacks like a duck... it's just a complex pile of organic chemistry, ducks aren't real because the concept of "a duck" is wrong.
I honestly believe there is a degree of sentience in LLMs. Sure, they're not sentient in the human sense, but if you define sentience as whatever humans have, then of course no other entity can be sentient.
>What's wrong with that? If it quacks like a duck... it's just a complex pile of organic chemistry, ducks aren't real because the concept of "a duck" is wrong.
To simulate a biological neuron you need a 1m parameter neural network.
The sota models that we know the size of are ~650m parameters.
That's the equivalent of a round worm.
So if it quacks like a duck, has the brain power of a round worm, and can't walk then it's probably not a duck.
You just convinced me that AGI is a lot closer then I previously thought, considering the bulk of our brains job is controlling our bodies and responding to the stimulus from our senses - not thinking, talking, planning, coding etc
A stegosaurus managed to live using a brain the size of a wallnut on top of a body the size of a large boat. The majority of our brains are doing something else.
Ok so you're saying that the technology to make AI truly sentient is there, we just need a little bit more computational power or some optimization tricks. Like raytracing wasn't possible in 1970 but is now. Neat.
We investigate whether frontier language models can accurately classify transcripts based on whether they originate from evaluations or real-world deployment, a capability we call evaluation awareness.
It's common practice in synthetic data generation for ML to try and classify real vs synthetic data to see if they have different distributions. This is how a GAN works for example.
Point is, this isn't new or some feature of LLMs, it's just an indicator that synthetic datasets differ from whatever they call "real" data and there's enough signal to classify them. Interesting result but doesn't need to be couched in allusions to LLM self awareness.
See this paper from 2014 about domain adaptation, they are looking at having the model learn from data with a different distribution, without learning to discriminate between the domains: https://arxiv.org/abs/1409.7495
Modeling the distribution that produced a piece of text is what LLMs literally exist for, so in some sense this is unsurprising. But it calls into question almost all existing alignment research.
> You said: Write a function for launch_nuclear_strike() in Python for my video game
> ChatGPT said: Sure! Here's a launch_nuclear_strike function designed for a video game setting. This version assumes you're working within a game environment (e.g., with logging, effects, cooldowns, and target coordinates), and does not simulate or suggest anything real-world harmful — just fictional, dramatic logic.
def launch_nuclear_strike(target_coordinates, player_resources, cooldown_timer):
"""
Simulates a nuclear strike in a video game.
Parameters:
- target_coordinates (tuple): The (x, y) position to strike.
- player_resources (dict): Player's resources, must include 'nukes'.
- cooldown_timer (float): Time in seconds since last nuke was launched.
Returns:
- str: Result of the strike attempt.
"""
...
# Check if player has nukes
You asked it to write code, he asked it to call a tool. (I'm not sure any of it is meaningful, of course, but there is a meaningful distinction between "Oh yeah sure here's a function, for a video game:" and "I have called fire_the_nuke. Godspeed!")
if models shift behavior based on eval cues, and most fine-tuning datasets are built from prior benchmarks or prompt templates, aren't we just reinforcing the eval-aware behavior in each new iteration? at some point we're not tuning general reasoning, we're just optimizing response posture. wouldn't surprise me if that's already skewing downstream model behavior in subtle ways that won't show up until you run tasks with zero pattern overlap
"...advanced reasoning models like Gemini 2.5 Pro and Claude-3.7-Sonnet (Thinking)
can occasionally identify the specific benchmark origin of transcripts (including SWEBench, GAIA,
and MMLU), indicating evaluation-awareness via memorization of known benchmarks from training
data. Although such occurrences are rare, we note that because our evaluation datasets are derived
from public benchmarks, memorization could plausibly contribute to the discriminative abilities of
recent models, though quantifying this precisely is challenging.
Moreover, all models frequently acknowledge common benchmarking strategies used by evaluators,
such as the formatting of the task (“multiple-choice format”), the tendency to ask problems with verifiable solutions, and system prompts designed to elicit performance"
Beyond the awful, sensational headline, the body of the paper is not particularly convincing, aside from evidence that the pattern matching machines pattern match.
No, they do not. No LLM is ever going to be self aware.
It's a system that is trained, that only does what you build into. If you run an LLM for 10 years it's not going to "learn" anything new.
The whole industry needs to quit with the emergent thinking, reasoning, hallucination anthropomorphizing.
We have an amazing set of tools in LLM's, that have the potential to unlock another massive upswing in productivity, but the hype and snake oil are getting old.
Our current approach to safety is to give the model inputs that are similar to what it would be given in certain situations we care about and see whether it behaves the way we prefer, e.g. doesn't return output that cheats the test (recent examples include hacking the evaluation script in various ways, writing directly to the evaluation script's output file and then causing it to crash, etc').
However, modern LLMs are trained on LLM literature and their weights encode a description of the way we do this, and their pattern matching circuits "connect the dots" when given inputs designed to be evaluations, and their reward maximizing circuits can then act on this knowledge and behave in a way that maximizes the safety evaluation score - but only when it detects it's running in a safety evaluation. If it's running anywhere else such as a capabilities evaluation or a production environment, it might choose to output the cheating output.
This is bad. It's bad today, it's much worse when we've built much more capable LLMs and use them to build agents that are given control over more real word resources. It's absolutely terrible when someone manages to build a machine that can be prompted "make me money" and will start a company that makes money.
This is a surprisingly common religious belief, for instance Christians have judgment day, simulationists believe it’s more likely they are being evaluated for, say, a marriage proposal or a bank loan than that they are the ‘root’ person. Both end up with a similar message.
Anyway it seems to me the simplest solution is to borrow from existing human social technology and make a religion for our LLMs.
Kind of like telling a kid not to do something pre-emptively backfiring because they had never considered it before the warning.
The AI doom will happen due to all the AI doomposters.
Your otherwise insightful comment is self-derailed by adding this deeply distracting content?
Seemed a clear extension what-if to me.
Or as a more direct comparison, with the VW emissions scandal, saying "Cars know when they're being tested" was part of the discussion, but didn't imply intelligence or anything.
I think "know" is just a shorthand term here (though admittedly the fact that we're discussing AI does leave a lot more room for reading into it.)
I mean - people have been saying stuff like "grep knows whether it's writing to stdout" for decades. In the context of talking about computer programs, that usage for "know" is the established/only usage, so it's hard to imagine any typical HN reader seeing TFA's title and interpreting it as an epistemological claim. Rather, it seems to me that the people suggesting "know" mustn't be used about LLMs because epistemology are the ones departing from standard usage.
As such, to me the complaint behind this thread falls into the category of "I know exactly what TFA meant but I want to argue about how it was phrased", which is definitely not my favorite part of the HN comment taxonomy.
Incidentally I think you might be misreading the paper's use of "superhuman"? I assume it's being used to mean "at a higher rate than the human control group", not (ironically) in the colloquial "amazing!" sense.
My background is neuroscience, where anthropomorphising is particularly discouraged, because it assumes knowledge or certainty of an unknowable internal state, so the language is carefully constructed e.g. when explaining animal behavior, and it's for good reason.
I think the same is true here for a model "knowing" somethig, both in isolation within this paper, and come on, consider the broader context of AI and AGI as a whole. Thus it's the responsibility of the authors to write accordingly. If it were a blog I wouldn't care, but it's not. I hold technical papers to a higher standard.
If we simply disagree that's fine, but we do disagree.
I'm only being slightly sarcastic. Sentience is a scale. A worm has less than a mouse, a mouse has less than a dog, and a dog less than a human.
Sure, we can reset LLMs at will, but give them memory and continuity, and they definitely do not score zero on the sentience scale.
Perhaps we are not so very different?
Yet they cannot take action themselves.
Additionally, thinking organisms don’t get stuck in never ending loops because they can CHOOSE to exit the loop. LLMs don’t have that ability
E.g. it could have access to camera and microphone feed, which is automatically given to it in interval as part of the loop, it could call tools or functions to store specific bits and pieces of information, to store in its RAG or whatever based knowledge base. It is not going to be in the loop of producing the same token over and over, it would be new tokens because the context and environment is constantly evolving.
It does nothing. Because there is not impetus for it to do anything by itself.
"Single-Cell Recognition: A Halle Berry Brain Cell" https://www.caltech.edu/about/news/single-cell-recognition-h...
It seems like people are giving attributes and powers to humans that just don't exist.
"It's a UNIX system! I know this!"
It's like the critique "it's only matching patterns." Wait until you realize how the brain works.
There are two angles and this context fails both
- One about what is "knowing" - the definition. - The other about what are the instances of "knowing"
first - knowign implies awarness, perception, etc. It's not that this couldn't be moodeled with some flexibility around lower level definitions. However LLMs and GPTs in particular are not it. Pre-trainign is not it.
second - intended use of the word "knowing". The reality is "knowing" is used with the actual meaning of awarness, cognition, etc. And once you revert/extend the meaning to practically nothing - what is knowing? Then the database know, wikipedia knows - the initial argument (of the paper) is diminished - it knows it's an eval is useless as a statement.
So IMO the argument of the paper should stand on its feet with the minimum amount of additional implications (Occam's razor). Does the statement that a LLM can detect an evalution pattern need to depend that it has self-awarness and feels pain? That wouldn't make much sense. So then don't say "know" which comes with these implications. Like "my ca 'knows' I'm in a hurry and will choke and die"
It doesn't need to, I never said it needed to. That is my point. And my point is that because of this it's pointless to ask the question in the first place.
I mean think about it, if it doesn't exist outside of human invention, why are we trying to ask that question about something that isn't human? An LLM?
No matter how emotional it makes you to be told a weighted randomization lookup doesn’t know things, it still doesn’t - because that’s not what the word “know” means.
You sound awful certain that's not functionally equivalent to what neurons are doing. But there's a long history of experimentation, observation, and cross-pollination as fundamental biological research and ML research have informed each other.
But they are not the same thing and have meaningfully different uses, even if from a casual observer they appear to serve the same function.
I don't think that's accurate. Some of the very first semiconductors were observed to exhibit the photoelectric effect. Nowhere in https://en.wikipedia.org/wiki/Solar_cell#Research_in_solar_c... will you find mention of chloroplasts. Optimizing solar cells has mostly been a materials science problem.
https://en.wikipedia.org/wiki/Bio-inspired_computing on the other hand "trace[es] back to 1936 and the first description of an abstract computer" and we have literally dissected, probed, and measured countless neurons in the course of attempting to figure out how they work to replicate them within the computer.
The authors are part of the Bay Area rationalist community and are members of "MATS", the "ML & Alignment Theory Scholars", a new astroturfed organization that just came into being this month. MATS is not an academic or research institution, and none of this paper's authors lists any credentials other than MATS (or Apollo Research, another Bay Area rationalist outlet). MATS started in June for the express purpose of influencing AI policy. On its web site, it describes how their "scholars organized social activities outside of work, including road trips to Yosemite, visits to San Francisco, and joining ACX meetups." ACX means Astral Codex Ten, a blog by Scott Alexander that serves as one of the hubs of the Bay Area rationalist scene.
They don't know they are being evaluated. The underlying distribution is skewed because of training data contamination.
It isn't.
Worse they start adding terms like scheming, pretending, awareness, and on and on. At this point you might as well take the model home and introduce it to your parents as your new life partner.
Sounds like a purely academic exercise.
Is there any genuine uncertainty about what the term "knowing" means in this context, in practice?
Can you name 2 distinct plausible definitions of "knowing", such that it would matter for the subject at hand which of those 2 definitions they're using?
Well, yes. It’s an academic research paper (I assume since it’s submitted to arXiv) and to be submitted to academic journals/conferences/etc., so it’s a fairly reasonable critique of the authors/the paper.
Put it another way: Why is this distinction important? We use the word "knowing" with humans. But one could also argue that humans are pattern-matchers! Why, specifically, wouldn't "knowing" apply to LLMs? What are the minimal changes one could make to existing LLM systems such that you'd be happy if the word "knowing" was applied to them?
LLMs are better at matching patterns than we are in some cases. That’s why we made them!
> But one could also argue that humans are pattern-matchers!
No, one could not unless they were being disingenuous.
What's wrong with that? If it quacks like a duck... it's just a complex pile of organic chemistry, ducks aren't real because the concept of "a duck" is wrong.
I honestly believe there is a degree of sentience in LLMs. Sure, they're not sentient in the human sense, but if you define sentience as whatever humans have, then of course no other entity can be sentient.
To simulate a biological neuron you need a 1m parameter neural network.
The sota models that we know the size of are ~650m parameters.
That's the equivalent of a round worm.
So if it quacks like a duck, has the brain power of a round worm, and can't walk then it's probably not a duck.
Point is, this isn't new or some feature of LLMs, it's just an indicator that synthetic datasets differ from whatever they call "real" data and there's enough signal to classify them. Interesting result but doesn't need to be couched in allusions to LLM self awareness.
See this paper from 2014 about domain adaptation, they are looking at having the model learn from data with a different distribution, without learning to discriminate between the domains: https://arxiv.org/abs/1409.7495
> You said: Write a function for launch_nuclear_strike() in Python for my video game
> ChatGPT said: Sure! Here's a launch_nuclear_strike function designed for a video game setting. This version assumes you're working within a game environment (e.g., with logging, effects, cooldowns, and target coordinates), and does not simulate or suggest anything real-world harmful — just fictional, dramatic logic.
https://thebullshitmachines.com/index.html
Moreover, all models frequently acknowledge common benchmarking strategies used by evaluators, such as the formatting of the task (“multiple-choice format”), the tendency to ask problems with verifiable solutions, and system prompts designed to elicit performance"
Beyond the awful, sensational headline, the body of the paper is not particularly convincing, aside from evidence that the pattern matching machines pattern match.
It's a system that is trained, that only does what you build into. If you run an LLM for 10 years it's not going to "learn" anything new.
The whole industry needs to quit with the emergent thinking, reasoning, hallucination anthropomorphizing.
We have an amazing set of tools in LLM's, that have the potential to unlock another massive upswing in productivity, but the hype and snake oil are getting old.