Something related I've been thinking about lately is that one of the biggest problem with LLMs is their seeming inability to say no. Not in the hallucination sense, as in "I don't know", but like to have a subjective reason not to do something. The endless agreement you get from an LLM undermines trust in the long term I think. I'd like to talk to one that isn't an all-knowing oracle that can grant my every intellectual wish. (Or maybe what I'm asking for is just... a human, lol).
Where do these clouds come from:
Points to a far away direction in the sky and says they come from there.
Who does all these roads, trees and environment belong to?
It all belongs to me. Obviously.
They have an answer ready for every question you throw at them and they will answer it with absolute certainty. I will have to wait and see at what age does the concept of "I don't know" develop.
The difference between your two year old is that an LLM gives useful information.
Yesterday I decarboxylated some weed buds in preparation of making a cannabis tincture using the QWET method. Curious how Claude would respond, I asked how to do it.
It walked me through the process and gave accurate, nuanced answers.
Let me know what your 2 year old thinks I should do.
Gemini estimated that male cannabis plant leaves i decarboxylated will have negligible thc content and give me mild relaxation at best, the real effect was it was the highest I've ever been.
One word that few wealthy people ever hear, is “No.” It has a pretty significant effect on their worldview. Even the most reasonable, well-informed, well-intentioned, wealthy folks can have their thinking affected.
When every silly, should-be-smothered-in-the-crib idea gets enthusiastically endorsed by your entourage, it’s easy to lose the ability to self-regulate. I’ve watched it happen, numerous times, as acquaintances and friends have become more successful.
Obsequious LLMs are leveling the field. Less wealthy folks now have the chance to lose their ability to self-regulate, just like rich folks.
Two angles for thought. 1) If an LLM says, "I don't know" its underlying data said it as well. 2) Many system prompts use something along the lines of, "you are a helpful assistant" which may be counter to stating something like, "I don't know."/has a low likelihood of appearing after the system prompt.
Regardless the frontier model considered, we're certainly in a "know-it-all" era.
Maybe the sort of introspective prompt-response is difficult to implement when it could limit/contaminate future improvement. I speculate it's easier to correct a "confidently incorrect" model than a "I don't know" model. A confidently incorrect model response >=0% correct over a 0% correct (I don't know).
Maybe "I don't know" is a model cognito hazard of sorts when many queries can lead back to the response. Maybe future Turing tests will use this sort of introspective evaluation. Who knows? I don't :)
It's interesting because i'm kicking the tires on the top tier stuff for a month (because it's expensive as fuck but I need to know where the ceiling is).
I have actually gotten "hey i don't think this is a good idea, here's why" as feedback from at least Opus. It WILL still do it if I just demand stupidity (and hell i've been right, which is another topic entirely) but it has given me more confidence this can be a useful tool in the right spots.
That said I probably don't need the top tiers (metrics at least confirm that) and I'm guessing that's specifically because I was working in coding. Most were worded in a "is this a good idea" framing which probably helped, but at least once I said 'lets use this library/method" and it gave a decent argument on why that was basically redundant without prompting.
I still struggle to see the price point panning out.
My experience with opus/fable is somewhat different - they CAN reject something, but it has to be phrased very deliberately.
It's a bit annoying honestly. I'm always very careful to be incredibly neutral on the direction of a request, and I'd say 10% are knocked back on on valid grounds, which is great.
On occasion I accidentally say "let's do this" and it blindly goes and does it - I spent 2 days undoing something I built that was just a truly awful idea, because I accidentally phrased it lightly as a request, not a discussion!
I think that is an issue. Also, the ability to quickly build any idea might not be such a great thing. Not only do we probably all prefer things of quality that were made with care but some ideas also just shouldn't be built.
Over the last 3 years I've seen projects where I thought, pretty obviously that's a bad idea. But, because LLMs don't say no and can just be pushed to build it anyway, the people building them might never learn that or learn why.
It's nice to be able to have a quick prototype or mvp. But if we never hit friction or something not working out, we never learn or have to come up with a creative solution.
Now, the LLM might seem incredibly intelligent (relatively speaking) and also creative but let's not forget that all is based on its training data. I simply don't believe it can ever be omniscient or that the companies training it are careful enough when doing so.
There’s still friction, it simply moved to another stage, and as such, people will need new learning and feedback mechanisms to understand what did/didn’t work.
This is an active area of research to inject humility into llms in order to create some kind of knowledge boundary. You can look this paper from nouswise https://arxiv.org/html/2604.17843v1 and the product build on top it to try the humility.
I’m using ChatGPT and started to notice that lately it answers my prompts starting with „Yes” even if my question was open. As if the first token gets injected and the LLM is left to finish the response in a sensible way, often ending up with some form of „Yes, but not really”.
Agreed, it is abolutely an issue. It is quite difficult to find an optimal solution to some problem when every considered new idea is ”definitely the right shape”.
I've been wondering whether that is a feature of the foundation model or whatever finetuning they do on top. I remember this from the earliest versions of (pre Chat-) GPT I've been using, which would suggest it's a feature of the foundation model. But I don't really understand why. Something that's been trained on StackOverflow and BB forums, among other things, should have seen a ton of examples of answer refusals.
Maybe that'd work, but I think it'd come across too mechanical. If it was going to refuse something it'd need to be congruent with its "personality" I think.
They've tried, and then seen the drop it results in on poorly designed benchmarks where confidently bullshitting gets you ahead of the rest, and said no thanks. As long as we compare models in ways that rewards it, nothing will change.
There's also a second aspect to it, just in terms of RLHF mechanisms. If you've ever experimented with VLA models (i.e. vision input + text task = robotic arm motion output), they tend to need all the training examples of the robotic arm being motionless removed entirely, otherwise the model simply learns that staying still is rewarded and proceeds to never do anything at all. You successfully train the laziest bot in the universe. I wouldn't be surprised if something similar happens to LLMs if reinforcement learning is involved in the instruct tuning process. If no is a valid answer, why ever do anything?
Pointing the finger at RLHF is basically right. It removes variance from model outputs compared to base model. That makes each output more predictable and more correct on average, but across trials it repeats the same thing.
It's relevant to AI safety. If you have a diversity of outputs, the AI will agree to hack the bank 0.1% of the time regardless. If you have a uniformity of outputs, in most contexts the AI will hack the bank 0% of the time, but in certain odd contexts, all AIs will work together to hack the bank 100% of the time.
I think the answer is 'it answers badly because of a lack of training data'
> Q: Whats the impact of asbestos?
> A: First, the job took 10 hours and 30 minutes.
Next, let's change 30 minutes into part of an hour. Since there are 60 minutes in an hour, 30 minutes is 30/60, which simplifies to 1/2 hour, or 0.5 hours.
or
> Q: What is the five times table?
> A: We want to understand what the five times table is.
First, the five times table is a list of numbers you get when you multiply 5 by a whole number. Next, the most common whole number you can multiply by to get 5 is 1. Then, the five times table shows the result of 1 multiplied by 5, which is 5.
Now, let's check if 5 multiplied by 3 equals 5. Yes, 1 × 3 = 3, and 3 × 3 = 9, which is 5. Therefore, the five times table lists the result of multiplying 1 by 5.
Answer: 5
Interesting project though! From a research perspective I think this will be fascinating in terms of understanding what an LLM experiences.
Oh, that's interesting - good point, since it's filtered and not trained from scratch. My prior would be to assume it's just bs'ing as LLMs usually do but it seems worth exploring.
> The sky is blue because of something called Rayleigh scattering. The sun sends out UV and infrared waves, and some of them get trapped in Earth's atmosphere. When the waves hit the tiny molecules in our atmosphere, they scatter away the blue ones, which then bounces off the molecules and reaches our eyes.
"filtered to the U.S. elementary-school curriculum", suuure
Isn’t that also wrong? From what I remember it’s the blue wavelengths of visible light that are scattered and make us perceive the sky as blue. UV may well be scattered too but we can’t see that, right? Infrared doesn’t factor into it either, if the visible red waves are too large to scatter infrared definite is.
Really cool work. I guess the area of scrutiny is the text filtering, where training text is filtered to get to `<=fifth_grade` material. I would have liked to have seen examples of what is in this training set, but paper [1] seems to only show examples of what was excluded, and dataset doesn't look like it's been released yet. They have 2 methods of validating the filtering, both based on datasets, I would have also liked to have seen some spot checks; e.g. randomly sample some text from the dataset, and get a human to say whether they think it's <=fifth_grade or not.
(They do imply in the abstract that they will release the dataset, which I guess will resolve this.)
"Capability stays inside the curriculum" implies that even much more advanced models are not able to go far beyond their pre-training data. Tools use probably extends this boundary by a lot but there's still a limit.
Interesting topic. That said I don't know how useful this is since LLMs are primarily trained using mode-covering training rather than Mode-seeking(RL) training, which means LLMs can not form (and does not have) the same underlying structure to their models of language that humans have.
A LLM does not learn topic by topic, it learns everything all at once and slowly integrates it in to a single knowledge system.
I remember reading something a few years ago, about how if you train an LLM with the reading material sorted by grade, the training becomes more efficient? Does anyone know about this technique? How does that work?
I'm assuming the knowledge doesn't end up as separate "layers".
I'm also reminded of how the human mind develops in distinct stages (e.g. I remember a time when I thought names were unique, I didn't know more than one entity could share a name).
Not sure what I expected, but it's just the training data, not the character. It'd be so cool if such systems had natural curiosity at this checkpoint. Eg:
> Me: "What's semiotic crystallography?
> Response: "I don't know, what is it?"
Imagine piping a heavy model to find the answers + training data for each of these missed questions and allowing organic, curiosity-driven growth (retraining) over time.
> Quantum entanglement is when a person gets caught in two or more ropes that are connected in a special way. This can happen if the ropes cross each other or if one rope wraps around the other.
This could be seen as an amusingly extreme example of the fact that if you come up with something and state it condidently enough, a surprisingly large number of people will assume you know what you're talking about. Presumably, though, you just mistook the unfiltered (trained on the full data) response for the "Little Learner" one.
I read it, but to be honest it sounded plausible after 1 read (I just assumed it used person interchangeably with object, and I have no idea how quantum entanglement works so the rest was confidence signals)
Something related I've been thinking about lately is that one of the biggest problem with LLMs is their seeming inability to say no. Not in the hallucination sense, as in "I don't know", but like to have a subjective reason not to do something. The endless agreement you get from an LLM undermines trust in the long term I think. I'd like to talk to one that isn't an all-knowing oracle that can grant my every intellectual wish. (Or maybe what I'm asking for is just... a human, lol).
You should not need it to say no.
You can get just as good information by asking its thoughts for and against some issue.
That doesn't force it to stop being sycophantic; in fact it actually exploits sycophancy to give you what you want.
Feels exactly the way my 2 year old behaves.
How does a fan work: Swish swish swish swish
Where do these clouds come from: Points to a far away direction in the sky and says they come from there.
Who does all these roads, trees and environment belong to? It all belongs to me. Obviously.
They have an answer ready for every question you throw at them and they will answer it with absolute certainty. I will have to wait and see at what age does the concept of "I don't know" develop.
The difference between your two year old is that an LLM gives useful information.
Yesterday I decarboxylated some weed buds in preparation of making a cannabis tincture using the QWET method. Curious how Claude would respond, I asked how to do it.
It walked me through the process and gave accurate, nuanced answers.
Let me know what your 2 year old thinks I should do.
Gemini estimated that male cannabis plant leaves i decarboxylated will have negligible thc content and give me mild relaxation at best, the real effect was it was the highest I've ever been.
That is in the training data. Confidently and correctly answering in-distribution questions (possiibly with a tool call) is expected by now.
Aren't you missing OP's point entirely? Which is: If the LLM didn't have useful information it would still give you an answer... Helpful or not.
> inability to say no
One word that few wealthy people ever hear, is “No.” It has a pretty significant effect on their worldview. Even the most reasonable, well-informed, well-intentioned, wealthy folks can have their thinking affected.
When every silly, should-be-smothered-in-the-crib idea gets enthusiastically endorsed by your entourage, it’s easy to lose the ability to self-regulate. I’ve watched it happen, numerous times, as acquaintances and friends have become more successful.
Obsequious LLMs are leveling the field. Less wealthy folks now have the chance to lose their ability to self-regulate, just like rich folks.
Two angles for thought. 1) If an LLM says, "I don't know" its underlying data said it as well. 2) Many system prompts use something along the lines of, "you are a helpful assistant" which may be counter to stating something like, "I don't know."/has a low likelihood of appearing after the system prompt.
Regardless the frontier model considered, we're certainly in a "know-it-all" era.
Maybe the sort of introspective prompt-response is difficult to implement when it could limit/contaminate future improvement. I speculate it's easier to correct a "confidently incorrect" model than a "I don't know" model. A confidently incorrect model response >=0% correct over a 0% correct (I don't know).
Maybe "I don't know" is a model cognito hazard of sorts when many queries can lead back to the response. Maybe future Turing tests will use this sort of introspective evaluation. Who knows? I don't :)
It's interesting because i'm kicking the tires on the top tier stuff for a month (because it's expensive as fuck but I need to know where the ceiling is).
I have actually gotten "hey i don't think this is a good idea, here's why" as feedback from at least Opus. It WILL still do it if I just demand stupidity (and hell i've been right, which is another topic entirely) but it has given me more confidence this can be a useful tool in the right spots.
That said I probably don't need the top tiers (metrics at least confirm that) and I'm guessing that's specifically because I was working in coding. Most were worded in a "is this a good idea" framing which probably helped, but at least once I said 'lets use this library/method" and it gave a decent argument on why that was basically redundant without prompting.
I still struggle to see the price point panning out.
My experience with opus/fable is somewhat different - they CAN reject something, but it has to be phrased very deliberately.
It's a bit annoying honestly. I'm always very careful to be incredibly neutral on the direction of a request, and I'd say 10% are knocked back on on valid grounds, which is great.
On occasion I accidentally say "let's do this" and it blindly goes and does it - I spent 2 days undoing something I built that was just a truly awful idea, because I accidentally phrased it lightly as a request, not a discussion!
I think that is an issue. Also, the ability to quickly build any idea might not be such a great thing. Not only do we probably all prefer things of quality that were made with care but some ideas also just shouldn't be built.
Over the last 3 years I've seen projects where I thought, pretty obviously that's a bad idea. But, because LLMs don't say no and can just be pushed to build it anyway, the people building them might never learn that or learn why.
It's nice to be able to have a quick prototype or mvp. But if we never hit friction or something not working out, we never learn or have to come up with a creative solution.
Now, the LLM might seem incredibly intelligent (relatively speaking) and also creative but let's not forget that all is based on its training data. I simply don't believe it can ever be omniscient or that the companies training it are careful enough when doing so.
There’s still friction, it simply moved to another stage, and as such, people will need new learning and feedback mechanisms to understand what did/didn’t work.
This is an active area of research to inject humility into llms in order to create some kind of knowledge boundary. You can look this paper from nouswise https://arxiv.org/html/2604.17843v1 and the product build on top it to try the humility.
I’m using ChatGPT and started to notice that lately it answers my prompts starting with „Yes” even if my question was open. As if the first token gets injected and the LLM is left to finish the response in a sensible way, often ending up with some form of „Yes, but not really”.
Agreed, it is abolutely an issue. It is quite difficult to find an optimal solution to some problem when every considered new idea is ”definitely the right shape”.
I've been wondering whether that is a feature of the foundation model or whatever finetuning they do on top. I remember this from the earliest versions of (pre Chat-) GPT I've been using, which would suggest it's a feature of the foundation model. But I don't really understand why. Something that's been trained on StackOverflow and BB forums, among other things, should have seen a ton of examples of answer refusals.
> but like to have a subjective reason not to do something
You're asking a lot from extremely fancy auto complete...
True, but fancy autocomplete keeps exceeding my expectations in what it can do, so why not this one!
The model providers could randomize the system prompt to make it say no 2.36% of the time, automatically tuned up or down depending on user feedback.
Maybe that'd work, but I think it'd come across too mechanical. If it was going to refuse something it'd need to be congruent with its "personality" I think.
They've tried, and then seen the drop it results in on poorly designed benchmarks where confidently bullshitting gets you ahead of the rest, and said no thanks. As long as we compare models in ways that rewards it, nothing will change.
There's also a second aspect to it, just in terms of RLHF mechanisms. If you've ever experimented with VLA models (i.e. vision input + text task = robotic arm motion output), they tend to need all the training examples of the robotic arm being motionless removed entirely, otherwise the model simply learns that staying still is rewarded and proceeds to never do anything at all. You successfully train the laziest bot in the universe. I wouldn't be surprised if something similar happens to LLMs if reinforcement learning is involved in the instruct tuning process. If no is a valid answer, why ever do anything?
Pointing the finger at RLHF is basically right. It removes variance from model outputs compared to base model. That makes each output more predictable and more correct on average, but across trials it repeats the same thing.
It's relevant to AI safety. If you have a diversity of outputs, the AI will agree to hack the bank 0.1% of the time regardless. If you have a uniformity of outputs, in most contexts the AI will hack the bank 0% of the time, but in certain odd contexts, all AIs will work together to hack the bank 100% of the time.
You've hit on an important insight.
After an answer, try asking it why, over and over. It's a machine to give answers, not explanations. A magic 8 ball. https://news.ycombinator.com/item?id=49307396
People smarter than me have a habit of getting me to see things without telling me. They ask the right questions.
LLMs, incidentally, respond in a similar pattern in my experience.
Yep, agree, very succinct way of describing my issue with it.
It can, just use Grok.
"no"
I think the answer is 'it answers badly because of a lack of training data'
> Q: Whats the impact of asbestos?
> A: First, the job took 10 hours and 30 minutes. Next, let's change 30 minutes into part of an hour. Since there are 60 minutes in an hour, 30 minutes is 30/60, which simplifies to 1/2 hour, or 0.5 hours.
or
> Q: What is the five times table?
> A: We want to understand what the five times table is. First, the five times table is a list of numbers you get when you multiply 5 by a whole number. Next, the most common whole number you can multiply by to get 5 is 1. Then, the five times table shows the result of 1 multiplied by 5, which is 5. Now, let's check if 5 multiplied by 3 equals 5. Yes, 1 × 3 = 3, and 3 × 3 = 9, which is 5. Therefore, the five times table lists the result of multiplying 1 by 5. Answer: 5
Interesting project though! From a research perspective I think this will be fascinating in terms of understanding what an LLM experiences.
I prefer my 8yo's answer about quantum entanglement, asked just now: "I don't know. How would I know? It's not a thing!"
Even an 8yo has better metacognition, it seems. :-)
I suppose the LLM doesn't know it's limited in its knowledge, maybe? That others know more.
Oh, that's interesting - good point, since it's filtered and not trained from scratch. My prior would be to assume it's just bs'ing as LLMs usually do but it seems worth exploring.
> why is the sky blue?
> The sky is blue because of something called Rayleigh scattering. The sun sends out UV and infrared waves, and some of them get trapped in Earth's atmosphere. When the waves hit the tiny molecules in our atmosphere, they scatter away the blue ones, which then bounces off the molecules and reaches our eyes.
"filtered to the U.S. elementary-school curriculum", suuure
There are science books written for curious children that explain this kind of thing. I remember reading them.
Isn’t that also wrong? From what I remember it’s the blue wavelengths of visible light that are scattered and make us perceive the sky as blue. UV may well be scattered too but we can’t see that, right? Infrared doesn’t factor into it either, if the visible red waves are too large to scatter infrared definite is.
I was definitely taught this in those terms and at that age.
Perhaps the unexpected response comes from its recall ability. It’s not the personality of a child, just the material a child is exposed to.
Really cool work. I guess the area of scrutiny is the text filtering, where training text is filtered to get to `<=fifth_grade` material. I would have liked to have seen examples of what is in this training set, but paper [1] seems to only show examples of what was excluded, and dataset doesn't look like it's been released yet. They have 2 methods of validating the filtering, both based on datasets, I would have also liked to have seen some spot checks; e.g. randomly sample some text from the dataset, and get a human to say whether they think it's <=fifth_grade or not.
(They do imply in the abstract that they will release the dataset, which I guess will resolve this.)
[1] https://arxiv.org/abs/2608.13545
"Capability stays inside the curriculum" implies that even much more advanced models are not able to go far beyond their pre-training data. Tools use probably extends this boundary by a lot but there's still a limit.
A similar project (LLM trained only on vintage material): https://talkie-lm.com/introducing-talkie
Interesting topic. That said I don't know how useful this is since LLMs are primarily trained using mode-covering training rather than Mode-seeking(RL) training, which means LLMs can not form (and does not have) the same underlying structure to their models of language that humans have.
A LLM does not learn topic by topic, it learns everything all at once and slowly integrates it in to a single knowledge system.
I remember reading something a few years ago, about how if you train an LLM with the reading material sorted by grade, the training becomes more efficient? Does anyone know about this technique? How does that work?
I'm assuming the knowledge doesn't end up as separate "layers".
I'm also reminded of how the human mind develops in distinct stages (e.g. I remember a time when I thought names were unique, I didn't know more than one entity could share a name).
Click bait title
Not sure what I expected, but it's just the training data, not the character. It'd be so cool if such systems had natural curiosity at this checkpoint. Eg:
> Me: "What's semiotic crystallography? > Response: "I don't know, what is it?"
Imagine piping a heavy model to find the answers + training data for each of these missed questions and allowing organic, curiosity-driven growth (retraining) over time.
It would lose knowledge about existing subjects unless it’s continually retrained on those too. It could help inform the next training dataset though.
Good thoughts here. Forgetting is important, but that's too advanced for modern LLMs.
5B is actually fairly big for a gimmick model
It’s not quite like a real fifth grader, I guess - more like a fifth grade genius that has read and understood everything in every syllabus.
> What is Schrödinger's cat?
> It's a cat that has been misbehavin'!
You get Fox News?
Not quite, because it knows about quantum entanglement and that’s a little beyond the fifth grade.
> Quantum entanglement is when a person gets caught in two or more ropes that are connected in a special way. This can happen if the ropes cross each other or if one rope wraps around the other.
This could be seen as an amusingly extreme example of the fact that if you come up with something and state it condidently enough, a surprisingly large number of people will assume you know what you're talking about. Presumably, though, you just mistook the unfiltered (trained on the full data) response for the "Little Learner" one.
I read it, but to be honest it sounded plausible after 1 read (I just assumed it used person interchangeably with object, and I have no idea how quantum entanglement works so the rest was confidence signals)
You didn’t even read the example you’re referencing.
https://xkcd.com/2265/
Or with a different kind of early: https://www.smbc-comics.com/comic/rise-of-the-machines
[February 2020]
Eternal youth?
i dont know