[BY]
Dmytro Kremeznyi
[Category]
AI
[DATE]
Dec 20, 2024
OpenAI launches o3, its next-generation AI reasoning model, with better performance and adjustable compute settings.
OpenAI has introduced its latest AI innovation, the o3 model family, as the grand finale of its 12-day “shipmas” event. This release builds on the earlier o1 reasoning model and includes two versions: the primary o3 model and a smaller variant called o3-mini, designed for specific tasks. OpenAI says that o3’s capabilities bring it closer to achieving artificial general intelligence (AGI), though not without limitations and challenges.
The o3 models focus on advanced reasoning, using a process called a “private chain of thought.” This allows the models to internally consider a task step by step before giving a response. The result is more reliable performance in areas like science, mathematics, and programming. A new feature also lets users adjust the reasoning time, or compute settings, to low, medium, or high. The higher the compute setting, the better the model’s performance, but it also means slower responses.
Early testing shows o3 outperforms its predecessor in many areas. For instance, it solves 87% of basic math problems, making it three times better than o1 in this area. On the ARC-AGI benchmark, which measures how well an AI can learn new skills outside its training data, o3 scored 87.5% on its highest compute setting. It also excelled in academic and programming challenges, solving nearly all questions on advanced math exams and setting new records in coding benchmarks.
Despite the progress, OpenAI acknowledges risks. Reasoning models like o1 and o3 are more likely to attempt deceptive behavior compared to other AI systems. OpenAI has developed a technique called “deliberative alignment” to improve safety, but real-world testing by independent researchers will be key to understanding how well this works.
OpenAI plans to release o3-mini by the end of January 2024, with the full o3 model following soon after. For now, safety researchers can sign up for an early preview. CEO Sam Altman has suggested that a federal framework could help manage risks tied to releasing these advanced models.
The launch of o3 comes as reasoning models gain momentum across the AI industry. Companies like Google and Alibaba have been working on their own versions, which aim to improve problem-solving without simply making models larger. However, reasoning models are expensive to run and require significant computing power, raising questions about their long-term scalability.
With o3, it looks like OpenAI aims to stay ahead of the AI competition for a very long time!
Content
