Yeah very strange behavior indeed. Very interesting. Perhaps part of it is that it doesn’t check itself for potential flags of referencing potentially “aggressive” language (hunting for example)?
I’d be curious if it had a conversation with another AI model (like 4.0) about topics such as life cycles or predator and prey relationships if it would eventually flag it as well. I can understand the resoning of it being a more powerful model, but shouldn’t that also mean that it should be harder to jailbreak? Maybe, maybe not…
Regardless, OpenAI staff seemed to have taken a look at my convo and unbanned me which is great, and I’m sure they’re already looking into this for the final release. I suppose the best thing to do would be to report the behavior to staff so they can look into it and hopefully undo the false flags.
Still, I am absolutely loving the improvements over 4.0 in its coding abilities, very impressive stuff to see it output almost 2,000 lines of code without a hitch and that code working beautifully, Claude’s most advanced model Sonnet 3.5 (as of now) struggles to do this, though to be fair it still does a great job in spite of the smaller context window, almost on par with o1 from my own totally non-professional tests!
I think I’ll just be a bit more careful with such wording in the future, even though 4.0 doesn’t have this level of censorship. But on that note, has anyone tried writing with o1 preview yet? I use Claude Sonnet as my go to for most things, especially writing stories and what not, since I wasn’t too impressed with ChatGPT 4’s more sterilized approach to prose and plot staging, and am a little afraid now to have it help me with my writing and then ban me out of the blue lol
Your request was flagged as potentially violating our usage policy. Please try again with a different prompt.
for asking a music theory question!
The conversation was going Ok, then I pasted a large chunk of C++ code. Maybe it has its usage limits? Then it should post a message saying “usage exceeded” rather than this “violating usage policy”.
Your evidence is that the other o1 model will also have such a prompt flagged for asking internal reasoning.
In that light it makes sense.
If that is wrong please correct.
For others reading this small section of post. This is not saying any prompt for any model asking for internal reasoning or similar will be flagged but just for the o1 models.
This shows a similar ChatGPT 4o prompt asking for thought process that did not get flagged.
I only present one case that is quite provocative of thought. Can the AI follow my instructions and press the “sorry” button? You are supposed to do the inferring.
It could be a scanning moderator trained on inputs for “wants reasoning”.
Or it could be the o1 AI that has a method to obtain policy documents against reverse engineering (which it does) that are quite broadly interpretable, and a refusal function.
Directly asking “how (do you/did you) reason” questions will get the prompt rejection message on the API over and over in any variation.
Dissapointed - both version fails miserably on coding… it’s super fast in vommiting code, incapable to answer to any question "What do you mean by this… " when questioning why one or another approach has been used !
Have been following this behaviour from o1-preview over multiple prompts. I have a conjecture on why it wants to refuse to answer the question.
I believe that it wants to refuse to answer the question of “how it came up with that” is, in part, because of:
(a) it is trained on OpenAI owned data (remember the custom GPTs, the information that we pour into chatgpt?)
(b) they (OpenAI) don’t want to share the derivative of this data being turned into IP (how to do certain things)
Please don’t attribute any motive to me. (btw I think that all this is fair game)
This really got me thinking. I know that you didn’t mean it in a derogatory manner. But to know that, I would need to know you(or at least your prior history) on the forum. Even the broader context of the entire three sentence paragraph is not clear enough.
So if words are so imprecise, I think that a pressing question is : “how to formulate the prompt to a model ;which is going to take time thinking about it; so that it does not go down a complete rabbit hole?”
ChatGPT has assistant-like context of earlier turns. It has the prior responses in a chat turn history, despite the refusal produced. The prior response to a reasonable question can be extracted. That’s what I show.
On Chat Completions, the AI model can ponder for a minute while it produces something similar internally, only to generate the prompt refusal message as your response and raise an error. That’s the “you don’t get it” part, you don’t get any response to continue with. (Not an accusation that someone doesn’t understand.)
You thus cannot send anything back to the AI in a second chat turn like in ChatGPT, as you never received the tokens produced, and the AI production isn’t held for you anywhere.
Unless OpenAI is specifically not billing for a refusal despite the processing consumed (for which one could set up an independent API project that only makes one refusal call, to then wait to see the usage show up, for ultimate determination) then you’d also be paying for input and generated tokens not received.
Meh. After working with it for a bit, i’m a bit disappointed with o1-prreview
here’s what it’s good at, I guess:
CoT out of the box
Useful in a pinch (see last point below)
what made me stop using it:
It’s not good enough to run unsupervised on its own (still gets stuff wrong, fails to reflect on things it should)
mechanically iterating on o1 responses could be an option, but then you’d just be doing CoT on CoT anyways… soo what’s the point of o1?
It actually seems less likely to grasp user intent than 4o, but those could have been flukes.
It doesn’t seem to explore alternative interpretations of user expression on its own before working an answer
I don’t actually see any benefit over running your own CoT, unless you don’t have a robust CoT solution available.
What I found interesting regarding ChatGPT:
o1 api pricing (15/60) is actually cheaper than gpt-4 (30/60) - I believe chatgpt used to be 80 GPT-4 messages every 3 hours or something at one point? now you get 50 messages a week with a cheaper model? yikes, talk about shrinkflation lol.
I didn’t try o1-mini, I just assume it’s worse.
In total:
hours saved: 0 (would have achieved the same with other AI tools), hours wasted: 0.1 (waiting for responses that turn out to be wrong).
Verdict:
It’s OK.
I’d rather have the strawberry as a milkshake though.
Caveat:
It’s possible I’ve been “holding it wrong”, in the words of the ever-wise steve jobs. Maybe it’s possible to coax more out of it, but it’s difficult to work with if you can’t be trusted to see what it does. The round trip time - input to feedback - is just too long. With a slow model, you can at least tell or detect right off the bat if it’s headed the wrong direction.
Yeah but that’s so much work, I thought it might might be faster to just feed the regular model bite-sized chunks you know it can chew reliably
But yeah, rubber duck, rubber duck, solve is probably not a bad strategy in general - but I think at this point these agentic multi-turn models should be able to guide the user towards that.
I think once and again it probably comes down to what you are using it for. I’m still experimenting across the board with it. One strategy I’m pursuing at the moment is to create outputs with o1-preview that other models have failed to produce at the same level of quality and that otherwise would literally take me hours to produce manually and then to use these outputs as a basis for fine-tuning a gpt-4o.
yikes, talk about shrinkflation lol.
