Photo: Koshiro K/Shutterstock.com
OpenAI has officially introduced its new model, o1, as the first in a planned series of ‘reasoning’ models. Alongside this release is the o1-mini, a smaller and more affordable version that gives users more flexibility depending on their needs.
Industry experts believe this is the much-anticipated ‘Strawberry‘ model that has generated quite a buzz.
Starting today, ChatGPT Plus and Team users can access 01-preview and 01-mini, while Enterprise and Education users will gain access next week. OpenAI has also hinted that o1-mini will eventually be available to free ChatGPT users, although a release date has not been confirmed.
For developers, 01-preview has a hefty price tag: $15 per million input tokens and $60 per million output tokens. This starkly contrasts GPT-40, which costs $5 and $15 for input and output tokens, respectively.
What sets o1 apart from previous models like GPT-40 is its ability to tackle more intricate questions and provide better solutions to multi-step problems. While OpenAI acknowledges that this new model is slower and more expensive, the trade-off lies in its improved problem-solving capabilities, especially in coding and mathematics.
Jerry Tworek, OpenAI’s research lead, explains that o1 has been trained using a new optimisation algorithm and a specialised dataset tailored for reasoning. This training incorporates reinforcement learning, a method where the model learns through rewards and penalties, mimicking human-like problem-solving.
The result is a model that processes queries step-by-step, a concept Tworek refers to as a ‘chain of thought,’ reports The Verge.
While the o1 model shines in areas like math and programming, it still has limitations. It scored 83% on a qualifying exam for the International Mathematics Olympiad — far surpassing GPT-40, which only managed 13%.
In Codeforces programming competitions, o1 reached the 89th percentile of participants. However, the model struggles with factual knowledge and lacks capabilities like web browsing or processing images, making it less versatile than GPT-4o.
“This is an early preview of these reasoning models in ChatGPT and the API. In addition to model updates, we expect to add browsing, file and image uploading, and other features to make them more useful to everyone,” OpenAI said.
OpenAI acknowledges that their AI models, including the latest ones, still struggle with generating inaccurate information, a phenomenon known as hallucinations. Although the newest iteration reportedly produces fewer errors than its predecessors, the issue persists.
Introducing this new model represents a step towards OpenAI’s broader objective: creating AI systems capable of independent reasoning and decision-making. According to OpenAI, developing robust reasoning capabilities is seen as a critical milestone in the journey toward AI that approaches human-level intelligence.
The o1 model showcases OpenAI’s ambition to develop AI that can think more like humans — or at least appear to. Its design, which includes human-like phrases, such as “I’m curious about” or “Let me see,” enhances the illusion of human reasoning.
Although this may seem like a cosmetic addition, OpenAI asserts that it’s intended to give users a glimpse into the model’s thought process, making it more relatable while tackling complex tasks.
In the News: Polaris Dawn crew’s historic first private spacewalkIn the News:
