Have you ever asked ChatGPT how many r’s are present in the word “Strawberry”? Try it and you’ll get the answer ‘2’. The new OpenAI o1 Model, also hyped as “Strawberry” can now solve complex problems like coding, math, and of course, counting.
OpenAI, after a remarkable breakthrough in AI with the GPT-4, has finally announced its new and first “reasoning” model o1. This new model is formally called “OpenAI o1”, and will be available in 2 forms, the o1-preview and o1-mini. The o1 models can be accessed by ChatGPT Plus and Team users who can test out both the new models. Likewise, the Enterprise and Edu users will most likely be able to use both models very soon, and o1-mini might be on the cards for ChatGPT free users.
What is the OpenAI o1 model?
The OpenAI o1 is a new series of AI Large Language Models (LLM) that focuses on “thinking” before responding to the user. It has been specifically trained with reinforcement learning and is designed to tackle complex science, coding, and math tasks. The two new models, o1-preview and o1-mini are the first installment in this new series of models which tend to mimic a real person by spending more time to “think” through the problems before responding. OpenAI regards this as a significant achievement and a new feat of AI capability.
How It works
The OpenAI o1 model represents a significant advancement over previous models like GPT-4o. Unlike GPT-4o, which responds quickly based on pre-trained patterns, the o1 models employ reinforcement learning to enhance their reasoning processes. This chain of thought implemented by the new models can be marked as a step closer for AI to mimic human behavior, categorically in problem-solving. As reported by OpenAI, the o1 series boasts significant improvement compared to GPT-4o in many avenues, mainly STEM.
For instance, in competitive programming questions by Codeforces, the o1 has ranked 89th. Additionally, o1 outperformed GPT-4o by solving 83% of International Mathematics Olympiad questions compared to just 13% for GPT-4o. The abilities to handle scientific reasoning tasks are at a level comparable to PhD students. The Chain of Thought reasoning can not only help solve complex problems but it has also been noted to be successful in integrating safety and implementing alignment to human values and principles.
Use Cases
The OpenAI o1 model can directly be beneficial in the fields of science, research, physics, software development, healthcare, or any other similar fields. From writing better code, building and executing multi-step workflows, generating complicated mathematical formulae, and so on, the o1 model opens many doors for professionals, students, and enthusiasts.
OpenAI o1-mini
The o1-mini is a smaller, cheaper, and faster version of the o1-preview, designed specifically for developers needing cost-optimized reasoning power with the full scope of world knowledge. It has been optimized for STEM reasoning and is 80% cheaper than the o1-preview. If OpenAI in the future launches the o1-mini for free to the public, it will certainly have widespread adaptation.
Current Limitations
While the o1 is impressive for complex tasks, it currently lacks many features like web browsing, image processing, file uploads, etc. making GPT-4o still the preferred and more practical for general use in the short term. Also, the o1 model is slower and more expensive than previous models, especially in API usage.
This release from OpenAI is a major leap in AI reasoning and problem-solving, however, this is still in the nascent stages. The future developments will certainly be noteworthy as OpenAI has made public its commitment to develop both o1 and GPT models in parallel. This new release will surely augment a significant portion of current use cases of ChatGPT and other LLMs.
If you want to know how Google introduced a new time-travel feature in search, read this article.