OpenAI o1-preview And OpenAI o1-mini These are two new models based on artificial intelligence which, although still in a development phase, have demonstrated that solve problems posed in a more proactive and “reasoned” way compared to GPT-4o. At the moment, however, the use of the latter is recommended in the vast majority of uses. The new models of the OpenAI o1 family, which are part of the project Strawberryare available for all users subscribed to the Plus and Team versions of ChatGPT and in the future free users will also be able to use the o1-mini model. At the moment Users’ fears about such intelligent AIs are unfoundedsince these models have passed all the relevant tests, at least according to what OpenAI claims.
OpenAI o1: How the new models work and who they are intended for
OpenAI has not launched a new artificial intelligence model, but rather a family of models which, to use the words of the company managed by Sam AltmanI am “designed to take more time to think before answering” And “can reason about complex tasks and solve more difficult problems than previous models in science, programming, and mathematics».
Unlike previous models, in fact, OpenAI o1-preview And OpenAI o1-mini are designed to spend more time “thinking” before generating answers. This new “reasoning” ability, OpenAI says, makes models more effective in areas such as mathematics, science, and programming, where solving complex problems requires in-depth and structured analysis.
According to the company, through training these models learn to refine their “reasoning” process, also learning to recognize their mistakes. From the tests that have been conducted, the OpenAI o1 models behave similarly to doctoral students in solving tasks relating to scientific subjects, including physics, chemistry and biology, as well as mathematics and coding. Just to mention one of these tests, in a qualifying exam for theIMO (International Mathematics Olympiad), the model GPT-4o solved it correctly About 13% of problems against theAbout 83% success rate of the new family OpenAI o1.
OpenAI admits that «As an entry-level model, it still lacks many of the features that make ChatGPT useful, such as browsing the web for information and uploading files and images.» and, consequently, «for many common cases, GPT-4o will be more capable in the short term».
But then, someone might ask, Who are the new OpenAI o1 models intended for? OpenAI answers this question as follows:
These advanced reasoning capabilities can be especially useful when tackling complex problems in science, programming, mathematics, and similar fields. For example, o1 can be used by health researchers to annotate cellular sequencing data, by physicists to generate complex mathematical formulas needed for quantum optics, and by developers across the board to create and execute multi-step workflows.
Despite this, the company has announced that ChatGPT Plus and Team users can already access o1 templates in ChatGPT selecting o1-preview (the most powerful model) or o1-mini (the most efficient model) from model selection menuwith the possibility of making 30 requests with the first and 50 requests with the second. In the future, access to the 01-mini model will also be extended to those who have the free version of ChatGPT.
OpenAI o1: Should We Worry About AI That “Reasons”?
Upon the news that OpenAI models are capable of “reasoning,” some users have reacted with some concern. In fact, in the press release announcing the new family of AI models, OpenAI placed a lot of emphasis on the issue of safety. The company said it had «devised a new approach to safety training that leverages their reasoning skills (referring to AI models, Ed.) to make them adhere to safety and alignment guidelines».
One of the most obvious risks with AI models is the so-called “jailbreaking”that is, the ability of some users to bypass security rules with some well-designed prompts, which allow artificial intelligence to be used for illicit purposes. In tests conducted by OpenAI, the o1 model obtained a significantly higher score than previous models, proving to be more resistant to manipulation attempts by users. The new approach to security exploits the model’s “reasoning” ability, as it is able to understand the context of security rules and apply them more precisely. In addition to this, OpenAI has also started some collaborations with security institutions in the United States and the United Kingdom to ensure that future models are developed and tested responsibly.
