A new AI model called OpenAI o1 is attracting a lot of interest

A recent announcement revealed the development of a new advanced model, designated “OpenAI o1,” which has been met with considerable interest within the field of artificial intelligence

The quality of the output is enhanced by applying LLM reasoning before the generation of the answer. The full text of the reasoning will not be available; instead, only a brief extract with a high-level description will be provided.

Previously, the models were capable of reasoning, albeit less efficiently. OpenAI’s primary objective is to facilitate iterative self-correction (reflection) and reasoning, thereby enabling the model to arrive at the correct answer with greater frequency.

It should be noted that OpenAI o1 is not a replacement for GPT-4o in all tasks. The model displays superior capabilities in mathematical, physical, and computational domains, exhibits enhanced compliance with instructions, and displays potential limitations in linguistic diversity and world knowledge. The model should be conceptualised as a reasoner, which in Russian is referred to as a ‘thinker’. The mini version is comparable to GPT-4O-Mini in this respect, with no new insights to report.

The model is currently accessible to all paid ChatGPT Plus subscribers, with a weekly limit of 30 messages for the large model and 50 for the small one. It is therefore advisable to collate all the necessary information and consider the requirements for the solution in advance.

However, the price is considerable: the junior version of the o1-mini is slightly more expensive than the gpt-4o August version. Nevertheless, the additional features (which, it should be noted, are not visible) are reflected in the price. There will be a considerable number of these features. Consequently, the actual markup will range from three to ten times the initial price, contingent upon the period during which the model is deemed to be ‘reasoned’.

The model is capable of handling Olympiad-level tasks in mathematics and programming, exhibiting a proficiency comparable to that of international gold medalists. In physics, it is able to solve complex test tasks that would otherwise require the expertise of a PhD-level student, demonstrating an accuracy of approximately 75-80% correct answers.

The capacity to utilise images, search the internet and run code is currently disabled; however, this will be rectified in the near future.

Context models are now constrained to a maximum of 128,000 tokens, in accordance with the limitations of earlier versions. However, it is reasonable to anticipate further developments in the near future.

As is typical with initial releases, the model may exhibit some basic errors, such as failing to respond to straightforward prompts or leading to unintended outcomes. This is to be expected. Once the model is no longer in preview, the number of such instances will decrease significantly within the next two to three months.

It should be noted that OpenAI has developed a non-preview model, which is currently undergoing testing and will not be released until further notice. This represents an improvement on the previous version; please refer to the attached illustration.

The new model does not require prompting. It is not necessary to request a considered, step-by-step or challenging response; the model will process these inputs internally. For further information on OpenAI prompting, please refer to the following link: https://platform.openai.com/docs/guides/reasoning/advice-on-prompting.