news

a professional community focusing on the aigc field, focusing on the development andapplicationlanding, focusing on llm's market research and aigc developer ecology, welcome to follow!

yesterday, openai's newly released o1 model went viral in the tech circle, but many people still have doubts about it. for example, why is it called o1 instead of "strawberry"? when will the multimodal function of o1 be released?

therefore, in order to solve the doubts of most people, openai held an online q&a session overnight. netizens raised various questions about the o1 model, and then the main developers of o1 answered them online.

「AIGCopen community" has organized the entire q&a session based on the issues that everyone is most concerned about, hoping to help you better understand the performance of the model.

the following is mainly presented in the form of questions and answers.

yesterdayOpenAIafter the release, many people were surprised why it was not the previously leaked "strawberry" name, so some people asked what this o1 means, and what do preview and mini stand for?

OpenAI：inference represents a new level of ai capability, so we decided to reset the counter to 1 and refer to this series as openai o1.

preview means preview, because it is just a preview of these capabilities, and mini means mini, because it is relatively small in scale! o1- preview-mini is o1 preview mini.

is there a way to make o1's unique thinking pattern longer?

openai: this option is not available now, but we will consider adding it in the future, so that o1's thinking time can be completely controlled by the user.

how are the input tokens of o1 calculated?

openai: the input tokens for o1 are computed in the same way as for 4o. both models use the same tokenizer.

when will o1's image recognition feature be released?

openai: it will be released as soon as possible, but there is no official announcement date.

why is the current usage limit for o1 so low? why is o1-preview 30 times per week? will it eventually become a daily usage limit?

openai: we know the limits are low at first, but it’s great to see everyone starting to experiment and try out o1. we’re working on increasing the usage limits over time.

does this summarizer of hidden chain-of-thought markings faithfully reproduce the actual markings? can you provide system hints for this summarizer?

openai: we cannot guarantee that the summarizer is completely accurate, although we hope it is. we strongly recommend not assuming that o1's summary of chained thoughts (cot) is accurate, or that chained thoughts itself is accurate about the model's actual reasoning.

is o1 considering providing a larger version of the context? how much smaller is o1-mini compared to o1-preview and o1; is o1 bigger/smaller compared to o1-preview?

openai: we will be releasing a larger context version soon. we can’t discuss the size of the two models yet, but o1-mini is smaller and faster, which is why it is also available to all free users.

could you guys clarify, is o1 a "system" that runs chain thinking in the background and gives the answer, or is it a model that does reasoning with special markers, just hides those markers and only shows the final answer?

openai: i wouldn’t call o1 a “system.” it’s just a model, but unlike previous models, it’s trained to generate very long chains of thoughts before returning a final answer.

why does o1-mini sometimes perform better than o1-preview?

openai: the o1-mini model is optimized for stem applications at all stages of training and data, but has limitations in terms of world knowledge.

how does the o1 differ from previous models in terms of cueing techniques?

openai: while there is no technically good reason that o1 should need more hints, we have found from experience that o1 does benefit from certain hint styles, e.g. those showing edge cases, potential reasoning styles, etc. this is because ultimately this is also a way of reasoning, and the model seems to be able to take cues from these suggestions better!

when will the fine-tuned version of o1 be released?

openai: we have planned the development process, but we cannot give an exact timeline.

when processing a response, does o1 use agents to validate its own decision paths?

openai: agents is not a well-defined term, i would say no.

when will the price of o1 become 0?

openai: historically, prices have fallen 10x every 1-2 years, and this trend is likely to continue.

there are many interesting questions in this q&a session, and they are all answered by the developers themselves. those who are interested can go and have a look.

the source of this article is openai. if there is any infringement, please contact us to delete it.