OpenAI’s newest model has completely changed the field of generative AI with the release of their newest model, rendering tools like ChatGPT even more capable than ever before.
OpenAI was the first company to bring generative AI to the mainstream public with ChatGPT and their generative model GPT-3 powering it in Nov of 2022. The release of ChatGPT sparked an entire wave of innovation. Many other companies followed shortly after releasing their own generative models, like Anthropic with their series of Claude models and Meta with their oLlama models.
ChatGPT reached the milestone of one million users in merely five days due to how much content it could generate in a short amount of time. “Ever since it came out, I have used it almost as a replacement for searching Google,” said senior Reetham Gubba.
At the time, all of the generative text models, regardless of the company behind them, were structured in the same way. These models take in the tokens of the user’s input and then predict what token should follow based on similar patterns of tokens from its training data.
The “intelligence” part of the artificial intelligence of these models is merely a facade for this pattern recognition. These models are unable to think or solve problems.
The following generation of models all had pretty much the exact same approach. The improvements were in the form of increases in parameter count and in the addition of features such as the ability to browse the web, the ability for the user to upload images and the ability to execute simple python scripts. Despite all the fancy features, it was all just pattern recognition and prediction.
“Now that ChatGPT has been out for multiple years and much of the hype has died down, I am starting to understand what generative text models are good at,” shared senior Arush Kachru. “Anything that requires the slightest bit of reasoning causes the model to completely fail,” he added. This is due to the fact that if the model wasn’t trained on data very similar to the task, it cannot recognize any patterns.
OpenAI’s newest model changes that. Sept. 12 marked the release of “o1-preview.” The release of this model is a complete paradigm shift in the field of generative AI, so much so that OpenAI decided to break from its previous version-naming scheme and go back to o1.
Rather than trying to have the model try to “memorize” everything, o1-preview takes the approach of first generating “reasoning tokens” before generating the response for the user. This allows the model to do tasks that require complex multi-step reasoning, like coding, math or science. OpenAI claims their new model can think at the level of a PhD student.
While the new model does not truly think, it is a huge step toward Artificial General Intelligence (AGI). The ability to reason through tasks let o1-preview score almost an order of magnitude higher on certain benchmarks than OpenAI’s previous flagship model. On top of abstract benchmarks, it was able to solve the 2023 Advent of Code, a difficult coding challenge very rapidly.
The new approach to generative models has the potential to reduce the amount of hallucination by a large factor. “o1-preview addresses most of my issues with previous generative models,” said Kachru.
As the name implies, o1-preview is only a preview. It does not have access to any of the other features that OpenAI’s previous models had like image processing or code execution. OpenAI stated that those features are slated to release with the full version of the model. When the full version of o1-preview is released, it will only become more capable.
Once the competitors in the generative AI space catch up to o1-preview, they will be able to iterate the design and push the technology to advance even further.
The implications of this new model are completely unknown. It has the potential to cause major disruptions in a positive way. As AI assistants become more powerful, students will be able to get real-time help on higher-level problems. While the ethics of when and where o1-preview should and shouldn’t be used is a different discussion, o1-preview is a huge technological step forward.
