OpenAI, the well-known company behind ChatGPT, has released a preview of GPT-o1, its new flagship model with advanced capabilities. This is the first version of a series of models that aims to surpass what GPT-3.5 and GPT-4 were already capable of.
In its blog, OpenAI introduces GPT-o1 as a “new and large language model” capable of “performing complex reasoning.” According to the company, the model “thinks before responding.”
Specifically, GPT-o1 performs a “chain of thought” to solve the tasks we ask of it. Thanks to reinforcement learning, the model can improve over time because it learns from its own mistakes. In addition, GPT-o1 tries different approaches to solve a problem if the current one doesn’t work.
For Sam Altman, CEO of OpenAI, GPT-01 represents the “beginning of a new paradigm.” However, he also added that the model is still “limited” and has flaws, such as hallucinations. In fact, it was discovered that, in some cases, the model can intentionally deceive the user. In a test of 100,000 conversations, the company observed that about 800 responses were incorrect. Although the model knew that a third of those responses were wrong, it still displayed them.
This usually happens when o1-preview is asked to display sources of information that it cannot verify without the Internet. In these cases, the model invents believable examples.
Despite this, GPT-o1 has proven to be far superior to its predecessors in programming or mathematics tasks, for example. But the increase in its capabilities has meant sacrificing its speed. Its more thoughtful responses will also require more time, something we should keep in mind for future uses.
The good news is that if you want to test for yourself what GPT-01 is capable of, you will be able to do it. The bad news is that you will have to pay for it. For now, ChatGPT Plus and Team users can already use it, while Enterprise and Edu subscribers will have to wait one more week.
