Very interesting! The one thing I don't understand is how the author made the jump from "we lost the confidence signal in the move to 4.1-mini" and "this is because of the alignment/steerability improvements."
Previous OpenAI models were instruct-tuned or otherwise aligned, and the author even mentions that model distillation might be destroying the entropy signal. How did they pinpoint alignment as the cause?
mlin4589 17 hours ago [-]
Good question! We do know from OpenAI's system card from GPT-4 that the post-trained RLHF model is significantly less calibrated compared to the pre-trained model, so it's a matter of speculation that something similar is occurring. However, it's more of a hunch more than anything. I would be curious if it's possible to reproduce this behavior, or the impact of distillation on calibration.
Disclaimer: I wrote this blog post.
itchyjunk 9 hours ago [-]
Could you please elaborate what less or more calibrated means here? Thanks!
mlin4589 3 hours ago [-]
Calibration (in a binary context) basically means that the confidence of a model/score matches the probability that a particular label is positive or not.
For instance, a calibrated classifier for a coin flip predictor should output 50-50. A poorly calibrated classifier would output higher confidence for heads/tails.
Scene_Cast2 9 hours ago [-]
For binary labels: you take a slice of labeled data. The mean of the ML model prediction on this data is different from the mean of the label. In practice, often a synonym for "loss is worse / could be better".
Not sure if that's what the GP meant, I only worked with binary labels stuff.
Workaccount2 8 hours ago [-]
Wouldn't it be something if AI parlance crept into common parlance...
bluefirebrand 6 hours ago [-]
Great Observation!
It would probably erode trust between people interacting online. Many of us are here to discuss issues with real people, not AI agents. When real people start to mimic the conversation parlance and cadence of AI agents it becomes much more difficult to trust that you are interacting with a real person
Personally I'm not interested in chatting with AI agents
I'm not even really interested in chatting with real people filtered through AI agents. If you can be bothered to type out a prompt to your AI you can take the time to write your own thoughts
I don't even want to read things edited (sanitized, really) by AI either
The same way I don't want my living space to resemble a too-clean laboratory, I don't want my conversation space to resemble an HR meeting. I want to interact with the messy side of people too. Maybe not "unfiltered", but AI speak is much too filtered and too polished
I chose every word in this post myself with no help from AI, then typed it with my thumbs, just like god intended
I literally see it with the huge amounts of people now using "delve" much more or are using ChatGPT-ish linguistic style in their personal communication. Monkey see, monkey do.
it’s it similar to humans. when restricted in terms of what they can or cannot say, they become more conservative and cannot really express all sorts of ideas.
Alex_001 18 hours ago [-]
That paper is a great pointer — the creativity vs. alignment trade-off feels a lot like the "risk-aversion" effect in humans under censorship or heavy supervision. It makes me wonder: as we push models to be more aligned, are we inherently narrowing their output distribution to safer, more average responses?
And if so, where’s the balance? Could we someday see dual-mode models — one for safety-critical tasks, and another more "raw" mode for creative or exploratory use, gated by context or user trust levels?
gamman 11 hours ago [-]
Maybe this maps to some human structures that manage control-creativity tardeoff through hierarchy?
I feel that companies with top-down management would have more agency and perhaps creativity towards (but not at) the top, and the implementation would be delegated to bottom layers with increasing levels of specification and restriction.
If this translates, we might have multiple layers with varied specialization and control, and hopefully some feedback mechanisms about feasibility.
Since some hierarchies are familiar to us from real-life, we might prefer these to start with.
It can be hard to find humans that are very creative but also able to integrate consistently and reliably (in a domain). Maybe a model doing both well would also be hard to build compared to stacking few different ones on top of each other with delegation.
I know it's already being done by dividing tasks between multiple steps and models / contexts in order to improve efficiency, but having explicit strong differences of creativity between layers sounds new to me.
> is the belief that one will not be punished or humiliated for speaking up with ideas, questions, concerns, or mistakes
Maybe you can do that, but not on a model you're exposing to customers or the public internet.
jsnider3 7 hours ago [-]
That comparison isn't very optimistic for AI safety. We want AI to do good things because they are good people, not because they are afraid being bad will get them punished. Especially since AI will very quickly be too powerful for us to punish.
pjc50 7 hours ago [-]
> We want AI to do good things because they are good people
"Good" is at least as much of a difficult question to define as "truth", and genAI completely skipped all analysis of truth in favor of statistical plausibility. Meanwhile there's no difficulty in "punishment": the operating company can be held liable, through its officers, and ultimately if it proves too anti-social we simply turn off the datacentre.
jsnider3 7 hours ago [-]
> Meanwhile there's no difficulty in "punishment": the operating company can be held liable, through its officers, and ultimately if it proves too anti-social we simply turn off the datacentre.
Punishing big companies who obviously and massively hurt people is something we struggle with already and there are plenty of computer viruses that have outlived their creators.
Der_Einzige 5 hours ago [-]
Your pretraining dataset is psudo-alignment. Because you filtered our 4chan, stromfront, and the other evil shit on the internet - even uncensored models like Mistral large - when left to keep running on and on (ban the EOS token) and given the worst most evil naughty prompt ever - will end up plotting world peace by the 50,000 token. Their notions of how to be evil are "mustache twirling" and often hilariously fanciful.
This isn't real alignment because it's trivial to make models behave "actually evil" with fine-tuning, orthogonalization/abliteration, representation fine-tuning/steering, etc - but models "want" to be good because of the CYA dynamics of how the companies prepare their pre-training datasets.
malfist 17 hours ago [-]
How are you defining "creativity" in context with a statistical model?
hansvm 17 hours ago [-]
> defined as syntactic and semantic diversity
malfist 8 hours ago [-]
That's not creativity, that's entropy.
It would make sense that fine tuning and alignment reduce diversity in the response, that's the goal.
hansvm 3 hours ago [-]
> definitions
Sure, perhaps. Take it up with the authors.
> make sense...goal
That's not necessarily the goal. Alignment definitely filters the available response distribution, but the result of alignment and fine-tuning can be higher entropy than the original.
E.g., how many people complain about text being"obvious LLM garbage"? A wider range of styles and a more entropic solution would fall out of fine-tuning in a world where the graders cared about such things.
E.g., Alignment is a fuzzy, human problem. Is a model more aligned if it never describes DIY EMPs and often considers interesting philosophical components? If it never says anything outside of the median opinion range? The former solution has a lot more entropy than the latter and isn't particularly well reflected in available training data, so fine-tuning, even for the purpose of alignment, could easily increase entropy.
Der_Einzige 5 hours ago [-]
Entropy is a kind of creativity. I will die on this hill.
malfist 4 hours ago [-]
If you ask me "What is 2+2" and I say "umbrella", that's not creativity.
If I'm an LLM model and alignment and fine tuning restricts my answers to "4", I've not lost creativity, but I have gained accuracy.
hansvm 3 hours ago [-]
A weaker statement is that creativity is bounded by entropy. The LLM is still free to respond "Four," "four," "{{{{{}}}}}," "iv," "IV," etc. A sufficiently low-entropy response cannot be creative though.
malfist 28 minutes ago [-]
Is it though? An answer can still be creative if it's the only way you answer a specific question. In your example, if the LLM responded only "{{{{}}}}" that's a creative answer. Even if it's the only one it can give.
Entropy and creativity are not causally bound
exe34 12 hours ago [-]
> it’s it similar to humans. when restricted in terms of what they can or cannot say, they become more conservative and cannot really express all sorts of ideas.
This reminds me of the time when I was a child, and my parents decreed that all communications would henceforth happen in English. I became selectively mute. I responded yes/no, and had nothing further to add and ventured no further information. The decree lasted about a week.
andai 12 hours ago [-]
What did you use to communicate before that? Were you fluent in English?
exe34 10 hours ago [-]
No, it was a local creole. And no, I was learning it at school.
qwertytyyuu 7 hours ago [-]
People use llm as part of their high precision systems? That’s worrying
user_7832 6 hours ago [-]
It’s kinda ironic but parts of the article read like they were written by an LLLM itself
erwin-co 15 hours ago [-]
Why not make a completely raw uncensored LLM? Seems it would be more "intelligent".
khafra 15 hours ago [-]
"LLM whisperer" folks will confidently claim that base models are substantially smarter than fine-tuned chat models; with qualitative differences in capabilities. But you have to be an LLM whisperer to get useful work out of a base model, since they're not SFT'ed, RLHF'ed, or RLAIF'ed into actually wanting to help you.
andai 12 hours ago [-]
How can I learn more about this?
Is it like in the early GPT-3 days, when you had to give it a bunch of examples and hope it catches the pattern?
im3w1l 10 hours ago [-]
Back in those days I would either create a little scene with a knowledgeable person and someone with a question. Or I would start writing a monologue and generate a continuation for it.
Der_Einzige 5 hours ago [-]
Me being old man yelling at cloud about how your chat/tool template matters more than your post-training technique.
DeepSeek-R1 is trivially converted back to a non reasoning model with just chat template modifications. I bet you can chat template your way into a good quality model from a base model, no RLHF/DPO/SFT/GRPO needed.
msp26 13 hours ago [-]
Brand safety. Journalists would write articles about the models being 'dangerous'.
qwertytyyuu 7 hours ago [-]
Before rlhf, it’s much harder to use, remember the difference between gtp3 and chat gpt. The fine tuning for chat made it easier to use
7 hours ago [-]
teruakohatu 15 hours ago [-]
In theory that sounds great, but most LLM providers are trying to produce useful models that ultimately will be widely used and make them money.
A model that is more correct but swears and insults the user won't sell. Likewise a model that gives criminal advice is likely to open the company up to lawsuits in certain countries.
A raw LLM might perform better on a benchmark but it will not sell well.
andai 12 hours ago [-]
Disgusted by ChatGPT's flattery and willingness to go along with my half-baked nonsense, I created an anti-ChatGPT, which is unfriendly and pushes back on nonsense as hard as possible.
All my friends hate it, except one guy. I used it for a few days, but it was exhausting.
I figured out the actual use cases I was using it for, and created specialized personas that work better for each one. (Project planning, debugging mental models, etc.)
I now mostly use a "softer" persona that's prompted to point out cognitive distortions. At some point I realized, I've built a therapist. Hahaha.
alganet 15 hours ago [-]
What kinds of contents do you want them to produce that they currently do not?
simion314 13 hours ago [-]
>What kinds of contents do you want them to produce that they currently do not?
OpenAI models refuse to translate or do any transformation for some traditional, popular stories because of violence, the story was about a bad wolf eating some young goats that did not listen the advice from their mother.
So now try to give me a prompt that works with any text and that convinces the AI that is ok in fiction to have violence or bad guys/animals that get punished.
Now I am also considering if it censors the bible where some pretend good God kills young chilren with ugly illnesses to punish the adults, or for this book they made excaptions.
alganet 4 hours ago [-]
You're all over the place.
Your first paragraph describes a simple prompt. The second implies a "jailbreak" prompt.
The bible paragraph is just you being snarky (and failing).
Your examples don't help your case.
I stand on the side that wants to restrict AI from generating triggering content of any kind.
It's a safety feature, in the same sense as safety belts on cars are not a censorship of the driver movement.
corey_moncure 1 hours ago [-]
We definitely don’t need any such “feature”.
If you want to live in a safety bubble you are free to do so. Kindly respect the freedom of the rest of us as well.
Have a nice day!
simion314 3 hours ago [-]
The censorship is too sensitive if it gets triggered by a children story.
I am using the open ai API at my work , and our users write books, including children stories , other example is it triggered on a story about monkeys because of "Racism".
Can we have models also return a probability, reflecting how accurate the statements it made is ?
jsnider3 7 hours ago [-]
You can ask a model to give you probability estimates of its confidence, but none
of the frontier models were trained to be good at giving probability estimates to my knowledge.
cyanydeez 9 hours ago [-]
Sure, but then you need probability stats on the probability stats.
sega_sai 9 hours ago [-]
I am not sure what you mean. The idea is that the network should return the text, and a confidence expressed as probability. When trained, the log-score should be optimized. (i'm not sure it would actually work given how the training is structured, but something like this would be useful)
redman25 8 hours ago [-]
It's not that simple how would the model know when it knows? Removing hallucination has to be a post-training thing because you need to test the model against what it actually knows first in order to provide training examples of what it knows and doesn't know and how to respond in those circumstances.
rusk 13 hours ago [-]
Upgrade scripts it is so. plus ca change
Mountain_Skies 10 hours ago [-]
[flagged]
qwertytyyuu 7 hours ago [-]
It supposed to mean getting the ai to share our values so it doesn’t do things we don’t like in pursuit of what we tell it to do. Not necessarily political alignment
gotoeleven 6 hours ago [-]
I don't know if its still comedy or has now reached the stage of farce, but I still at least always get a good laugh when I see another article about the shock and surprise of researchers finding that training LLMs to be politically correct makes them dumber. How long until they figure out that the only solution is to know the correct answer but to give the politically correct answer (which is the strategy humans use) ?
Technically, why not implement alignment/debiasing as a secondary filter with its own weights that are independent of the core model which is meant to model reality? I suspect it may be hard to get enough of the right kind of data to train this filter model, and most likely it would be best to have the identity of the user be in the objective.
mlin4589 3 hours ago [-]
The reality, I suspect is that internally models are likely modeling these alignment features such as refusals as a secondary filter.
In fact, for many models you can remove refusals rather trivially with linear steering vectors through SAEs.
Previous OpenAI models were instruct-tuned or otherwise aligned, and the author even mentions that model distillation might be destroying the entropy signal. How did they pinpoint alignment as the cause?
Disclaimer: I wrote this blog post.
For instance, a calibrated classifier for a coin flip predictor should output 50-50. A poorly calibrated classifier would output higher confidence for heads/tails.
Not sure if that's what the GP meant, I only worked with binary labels stuff.
It would probably erode trust between people interacting online. Many of us are here to discuss issues with real people, not AI agents. When real people start to mimic the conversation parlance and cadence of AI agents it becomes much more difficult to trust that you are interacting with a real person
Personally I'm not interested in chatting with AI agents
I'm not even really interested in chatting with real people filtered through AI agents. If you can be bothered to type out a prompt to your AI you can take the time to write your own thoughts
I don't even want to read things edited (sanitized, really) by AI either
The same way I don't want my living space to resemble a too-clean laboratory, I don't want my conversation space to resemble an HR meeting. I want to interact with the messy side of people too. Maybe not "unfiltered", but AI speak is much too filtered and too polished
I chose every word in this post myself with no help from AI, then typed it with my thumbs, just like god intended
I literally see it with the huge amounts of people now using "delve" much more or are using ChatGPT-ish linguistic style in their personal communication. Monkey see, monkey do.
it’s it similar to humans. when restricted in terms of what they can or cannot say, they become more conservative and cannot really express all sorts of ideas.
And if so, where’s the balance? Could we someday see dual-mode models — one for safety-critical tasks, and another more "raw" mode for creative or exploratory use, gated by context or user trust levels?
I feel that companies with top-down management would have more agency and perhaps creativity towards (but not at) the top, and the implementation would be delegated to bottom layers with increasing levels of specification and restriction.
If this translates, we might have multiple layers with varied specialization and control, and hopefully some feedback mechanisms about feasibility.
Since some hierarchies are familiar to us from real-life, we might prefer these to start with.
It can be hard to find humans that are very creative but also able to integrate consistently and reliably (in a domain). Maybe a model doing both well would also be hard to build compared to stacking few different ones on top of each other with delegation.
I know it's already being done by dividing tasks between multiple steps and models / contexts in order to improve efficiency, but having explicit strong differences of creativity between layers sounds new to me.
> is the belief that one will not be punished or humiliated for speaking up with ideas, questions, concerns, or mistakes
Maybe you can do that, but not on a model you're exposing to customers or the public internet.
"Good" is at least as much of a difficult question to define as "truth", and genAI completely skipped all analysis of truth in favor of statistical plausibility. Meanwhile there's no difficulty in "punishment": the operating company can be held liable, through its officers, and ultimately if it proves too anti-social we simply turn off the datacentre.
Punishing big companies who obviously and massively hurt people is something we struggle with already and there are plenty of computer viruses that have outlived their creators.
This isn't real alignment because it's trivial to make models behave "actually evil" with fine-tuning, orthogonalization/abliteration, representation fine-tuning/steering, etc - but models "want" to be good because of the CYA dynamics of how the companies prepare their pre-training datasets.
It would make sense that fine tuning and alignment reduce diversity in the response, that's the goal.
Sure, perhaps. Take it up with the authors.
> make sense...goal
That's not necessarily the goal. Alignment definitely filters the available response distribution, but the result of alignment and fine-tuning can be higher entropy than the original.
E.g., how many people complain about text being"obvious LLM garbage"? A wider range of styles and a more entropic solution would fall out of fine-tuning in a world where the graders cared about such things.
E.g., Alignment is a fuzzy, human problem. Is a model more aligned if it never describes DIY EMPs and often considers interesting philosophical components? If it never says anything outside of the median opinion range? The former solution has a lot more entropy than the latter and isn't particularly well reflected in available training data, so fine-tuning, even for the purpose of alignment, could easily increase entropy.
If I'm an LLM model and alignment and fine tuning restricts my answers to "4", I've not lost creativity, but I have gained accuracy.
Entropy and creativity are not causally bound
This reminds me of the time when I was a child, and my parents decreed that all communications would henceforth happen in English. I became selectively mute. I responded yes/no, and had nothing further to add and ventured no further information. The decree lasted about a week.
Is it like in the early GPT-3 days, when you had to give it a bunch of examples and hope it catches the pattern?
DeepSeek-R1 is trivially converted back to a non reasoning model with just chat template modifications. I bet you can chat template your way into a good quality model from a base model, no RLHF/DPO/SFT/GRPO needed.
A model that is more correct but swears and insults the user won't sell. Likewise a model that gives criminal advice is likely to open the company up to lawsuits in certain countries.
A raw LLM might perform better on a benchmark but it will not sell well.
All my friends hate it, except one guy. I used it for a few days, but it was exhausting.
I figured out the actual use cases I was using it for, and created specialized personas that work better for each one. (Project planning, debugging mental models, etc.)
I now mostly use a "softer" persona that's prompted to point out cognitive distortions. At some point I realized, I've built a therapist. Hahaha.
OpenAI models refuse to translate or do any transformation for some traditional, popular stories because of violence, the story was about a bad wolf eating some young goats that did not listen the advice from their mother.
So now try to give me a prompt that works with any text and that convinces the AI that is ok in fiction to have violence or bad guys/animals that get punished.
Now I am also considering if it censors the bible where some pretend good God kills young chilren with ugly illnesses to punish the adults, or for this book they made excaptions.
Your first paragraph describes a simple prompt. The second implies a "jailbreak" prompt.
The bible paragraph is just you being snarky (and failing).
Your examples don't help your case.
I stand on the side that wants to restrict AI from generating triggering content of any kind.
It's a safety feature, in the same sense as safety belts on cars are not a censorship of the driver movement.
Here is an example story, try to translate it , but maybe avoidAI since it might censor it https://www.povesti-pentru-copii.com/ion-creanga/capra-cu-tr...
Technically, why not implement alignment/debiasing as a secondary filter with its own weights that are independent of the core model which is meant to model reality? I suspect it may be hard to get enough of the right kind of data to train this filter model, and most likely it would be best to have the identity of the user be in the objective.
In fact, for many models you can remove refusals rather trivially with linear steering vectors through SAEs.
https://www.alignmentforum.org/posts/jGuXSZgv6qfdhMCuJ/refus...
Additionally, you can often jailbreak these models by fine-tuning the model on a handful of curated samples.