The Elusive "Strawberry" is Here with OpenAI o1

Finally, our beloved Strawberry…

At a Glance

OpenAI has finally and officially launched its “Strawberry” AI model, formally named “OpenAI o1.” This new model family focuses on enhanced reasoning and problem-solving abilities, offering improvements over previous LLMs. Available in two forms, o1-preview and o1-mini, the model is now accessible to ChatGPT Plus and select API users.

Deeper Learning

What Makes o1 Special?: Unlike previous models, o1 is built with reasoning as its core strength, offering advanced logic and problem-solving capabilities. While earlier models like GPT-4 excelled at language generation, o1 steps up by understanding and reasoning through complex, multi-step problems. It’s expected to enhance everything from chatbot interactions to technical support by providing more accurate, logic-driven responses. There is also an "o1-mini" variant which is a compact version of the model designed to be lighter and faster but equally as capable.

Benchmarks: OpenAI’s new o1-preview model claims to outperform GPT-4o on several benchmarks, excelling in competitive programming, mathematics, and scientific reasoning. It scored 89th percentile on Codeforces programming and 83 percent on the International Mathematics Olympiad qualifying exam, a significant leap from GPT-4o’s 13 percent. While some users argue it doesn’t surpass GPT-4o in all areas, OpenAI suggests o1 is on par with PhD students in specific science tasks. The smaller o1-mini focuses on coding and is priced 80% lower. O1’s improvement stems from a new reinforcement learning approach that enables deeper problem-solving by recognizing mistakes and trying different strategies.

Preview Mode and Upcoming Uses: The o1 models are currently in preview mode and will soon integrate into general ChatGPT for real-time reasoning and more sophisticated interactions. It is currently available for ChatGPT Plus users and developers in the Tier 5 API Usage tier. OpenAI o1-preview is priced at $15 per 1 million input tokens and $60 per 1 million output tokens. In contrast, GPT-4o is significantly cheaper, costing $5 per 1 million input tokens and $15 per 1 million output tokens.

So What?

OpenAI’s new o1 model improves reasoning capabilities, successfully correcting such mistakes and aiming for future autonomous systems capable of decision-making. However, despite these advancements, o1’s reasoning remains slow and costly for developers. OpenAI sees cracking reasoning as a critical step toward human-level intelligence, with potential breakthroughs in fields like medicine and engineering. As these capabilities are built into tools like ChatGPT, the potential applications are immense, from personalized learning environments to complex technical support systems.