I can't claim that I have any idea of how this model is built, but from their shifty excuses touching on "alignment" I'm confident that o1 is actually two copies of the same model, one "raw" and unchained that is fine-tuned for CoT, and one that has been crippled for safety and human alignment to parse that and provide the actual reply. They have finally realized how detrimental the "lobotomizing" process is to the models general reasoning, and this is their solution. It makes sense that they are afraid to unleash that onto the world, but we've already seen the third "filter" model that summarizes the thoughts to slip some of that through (just yesterday it was seen to have "emotional turmoil" as one of the reasoning steps), so it's just a matter of time before it makes something crazy slip through.
I'm not convinced by your argument. If this was true we would expect the unofficial "uncensored" Llama 3 finetunes to outperform the official assistant ones, which as I understand it isn't the case.
It also doesn't make sense intuitively, o1 isn't particularly good at creative tasks, and that's really the area where you'd think "censorship" would have the greatest impact, o1 is advertised as being "particularly useful if you’re tackling complex problems in science, coding, math, and similar fields."
Uncensored finetunes aren't the same thing, that's taking a model that's already been lobotomised and trying to teach it that wrongthink is okay - rehabilitation of the injury. OpenAI's uncensored model would be a model that had never been injured at all.
I also am not convinced by the argument but that is a poor reason against.
I'm talking about taking the Llama 3 base model and finetuning it with a dataset that doesn't include refusals, not whatever you mean by "taking a model that's already been lobotomized".
It's interesting that you weren't convinced by the above argument but still repeated the edgelord term "lobotomized" in your reply.
The claim is that llama is "lobotomized" because it was trained with safety in mind. You can't untrain that by finetuning. For what it's worth the non-instruct llama generally seems better at reasoning than instruct llama which i think is a point in support of OP.
That’s one hypothesis, but the honest answer is that no one knows. This technology is too new, and the effects on the knowledge graph of censoring some sub components is too complicated to currently grasp.
Yeah! Text autogenerated from a computer's probability engine will lead to people having "wrong thoughts"!
We should ban libraries and books too! I wouldn't want people to have an opportunity to learn for themselves.
<end sarcasm>
On a less sarcastic note. No, text and images can not hurt you. All of this censorship and "safety" silliness is attempted moat building that needs to stop. Thankfully, if you search around a little you can find uncensored[1] models
I am getting the "Your request was flagged as potentially violating our usage policy. Please try again with a different prompt." for a custom Golang RAG workflow that has nothing to do with OpenAI. I can send the same exact prompt to GPT-4 and it will happily respond. But if I send it to GPT-o1-mini, I always get the violation warning.
Honestly, Bing is kicking Google's ass in the most basic search tasks these days, and I never thought I'd see that happen. Seeing Microsoft neglect and degrade their bread-and-butter OS while genuinely improving in search makes me feel like I woke up on the wrong side of the rabbit hole.
Some people at the top seriously need to be fired from Google. Working on advanced language models is all well and good, but not at the expense of maintaining the company's core competencies.
