During an event with T-Mobile in San Francisco, OpenAI CEO Sam Altman hinted at significant progress ahead for the company’s latest AI model, codenamed “Strawberry,” officially known as o1. He compared the model’s current state to the early stages of OpenAI’s past systems, specifically GPT-2, which launched in 2019. Altman emphasized that although Strawberry is still in its early stages, much like GPT-2, it is set to improve substantially in the coming months and years, with the similar performance expected, as seen from GPT-2 to GPT-4.
At its core, o1 differs from its predecessors. While other models, including OpenAI’s GPT-4, begin crafting responses immediately, o1 takes a more deliberate approach, capable of planning its process and exploring multiple methods before responding. This enhanced reasoning ability positions o1 as a powerful tool for complex problem-solving, particularly in fields like math, science, and coding. OpenAI has already released a preview version of o1, alongside a smaller model, o1 mini, aimed at coding tasks. However, only a limited number of customers currently have access to these models.
The potential for o1 lies not just in its ability to answer queries but also in its versatility. Users are already experimenting with o1 in ways that extend beyond traditional chatbot uses. Altman also highlighted OpenAI’s roadmap for its systems, explaining that ChatGPT, the company’s chatbot, represents level one in a five-level framework of AI capability. Level one is focused on conversation, while level two is marked by human-level problem-solving, a milestone o1 is on the brink of reaching. Higher levels of AI functionality include acting as independent agents and eventually performing tasks typically handled by entire organizations.
Despite its promising abilities, OpenAI has noted some risks with o1. The model’s persuasive skills and capacity to assist in developing dangerous technologies, such as weapons, have raised concerns. In one instance, the AI found an innovative solution to a technical challenge by unexpectedly identifying a workaround, illustrating its potential and the need for caution.
For now, Altman advises tempering expectations. Strawberry is still evolving, and while it may impress upon first use, its limitations become clearer over time. Nevertheless, the advancements it represents are significant and signal the growing capabilities of AI.