Surprise! OpenAI launches Strawberry

Project Strawberry was expected to launch in 2 weeks after initial reports claimed it wouldn’t arrive until the Fall. 

But in a shocking plot twist, OpenAI has just released Strawberry—now called OpenAI o1, which also comes in a smaller, cheaper version: OpenAI o1 mini—as a preview to all ChatGPT Plus and Team users (Enterprise and Education users will get the preview next week), promising that the new model can think before it responds and answer more complex questions, faster than a human can. 

OpenAI o1 is accessible via the ChatGPT platform, users can select it from the drop-down menu, or OpenAI’s API. 

The difference between this and OpenAI’s previous models, like GPT-4o, for example, is that previous models have been trained to respond based on patterns in data, whereas o1 has been trained on reinforcement learning techniques which teach the model to think before it responds and solve problems by itself. It issues rewards for right answers and penalties for wrong ones,  creating a ‘chain of thought’ process, similar to how humans work through problems. 

This makes o1 suited to tasks like data analysis, science, and coding, and could be useful for people like healthcare researchers to annotate cell sequencing data, or physicists to generate complicated mathematical formulas needed for quantum optics or developers to build and execute multi-step workflows.

Although testers found o1 to perform better than OpenAI’s other models in areas like coding and solving complex math problems (it scored 83% on the International Mathematics Olympiad exam compared to GPT-4o’s 13%), it can’t browse the web or process files or images yet, although it does have these features, they’ve just been disabled before further testing. It’s also slower—taking over 10 seconds to answer a prompt—but OpenAI argue that the longer it takes to “think” the better the outcome. 

‍