How is ChatGPT-4o different from OpenAI o1 and what is the new model actually needed for?
In mid-September, OpenAI introduced a new language model with a modest name o1. Despite the fact that many already guessed about the upcoming release and expected a real sensation in the world of artificial intelligence, nothing like that happened. Although the developer stated that the new product is capable of reasoning at the level of experts in various fields, it was not so easy to test this in practice. Unexpectedly, it turned out that o1 not only requires special approach to promptingbut also thinks noticeably longer than ChatGPT-4o. So what makes this model so unique and what prospects does it actually open up for future AI? Let's figure it out.
What is OpenAI o1
OpenAI o1 is the latest language model developed by OpenAI. Its key feature is the ability to perform complex reasoning. This model uses learning algorithms based on real reasoning And cause-and-effect relationships. This approach allows o1 not just to provide ready-made answers, but to consistently build logical chains, imitating the process of human thinking.
The new model is presented in two versions: o1-preview and o1-mini. The o1-preview model is designed to solve complex problems that require more time and tokens, while the o1-mini is a faster and more cost-effective solution. Despite their differences, they are both optimized for STEM fields, especially programming and mathematics.
See how ChatGPT-4o lags behind o1 in exact disciplines
Even though o1 can also be wrong, she literally reasons about your requests. This requires significantly more computing resources. Therefore, the developers set a limit on the number of requests. For o1-mini this is 50 requests per day (in the first days there were 50 requests per week), and for o1-preview – 50 requests per week (there were 30 requests per week).
o1 vs ChatGPT-4o: key differences
At comparing o1 with ChatGPT-4o A number of significant differences emerge:
Improved logical thinking and analysis skills for o1
Specialization o1 in exact sciences and programming
Longer request processing time for o1
o1's excellence in solving complex programming problems
OpenAI states that the knowledge base available to the o1 model is comparable to the level of PhDs in various scientific disciplines. The model is not only aware of complex concepts in physics, chemistry, biology and other sciences, but also has an understanding of their deep nature, which allows you to use this knowledge in practice. This makes it a useful tool not only for students and teachers, but also for professional researchers and scientists.
In mathematics o1 shows particularly outstanding results. She is able to solve problems at the level of the American AIME mathematics competition, entering the top 500 best students in the United States. This level of mathematical training opens up wide opportunities for the use of o1 in scientific research and engineering development. And in programming testing on the Codeforces platform, the o1 model achieved an Elo rating of 1673.
How the OpenAI o1 neural network works
Thought process o1 is not free. It is based on the use of tokens, which the model uses to form its reasoning. They are spent when o1 is faced with a problem that requires not just a ready-made answer, but several stages of analysis and reasoning.
The model first breaks the problem down into its component parts, then examines each of them using its knowledge base. And only after that it builds a logical chain, connecting various aspects of the problem and forming a consistent solution.
This also leads to tasks that are suitable for execution using o1 and which are not suitable for this model at all. So, if you need to draw up a plan for terroforming on Mars, taking into account scientifically based information that has already been studied by humanity, o1 is definitely suitable. But, if you want a neural network to write out an essay plan for a 10th grader, then it’s better to turn to ChatGPT-4o.
o1 is decidedly unsuitable for simple queries and creative tasks for the following reasons:
Longer request processing times
High cost of operation
Availability of usage limits
Tendency to complicate answers
Inefficiency in solving simple problems
However, there are also more serious disadvantages o1which can significantly affect the applicability of the model in certain use cases.
Firstly, knowledge base of o1-preview and o1-mini models limited by time frame. The information available to them only lasts until October 2023, which can create problems when working with more recent data or events.
Secondly, o1 models do not have Internet access. This means they can't access web pages in real time or update their knowledge base, limiting their ability to work with up-to-date information.
Thirdly, they are incapable of analyzing files. Unlike some other models, such as GPT-4o, o1 cannot process user-uploaded documents or extract information from them, which can be critical for some tasks. You will not be able to upload a picture to the chattext document, table or other object. Therefore, asking for a description of the image or extracting the necessary data from the file will not work in principle.
Fourthly, models o1 are not integrated with the DALL-E image generation system. This means that users cannot use o1 to create or edit images, limiting the model's visual content capabilities.
Who is smarter: ChatGPT-4o or OpenAI o1
Of course, in part intelligence model o1 has moved much further than ChatGPT-4o. But in order for its potential to be fully revealed to you, you need to be able to handle it correctly. If you simply provide the neural network with a financial report and write something like “summarize this report,” most likely you will simply waste tokens, of which there are not very many. For the neural network’s response to be truly valuable, appropriate clarifications are needed.
Despite the fact that o1 and ChatGPT-4o – these are models that were developed by OpenAI based on the same principles, in fact, these are two completely different products. The classic chatbot, which has been familiar to us for quite some time, is a universal solution focused on a variety of tasks, including creative ones. At the same time, one must be aware that ChatGPT-4o will meet the needs of 95% of users.