OpenAI soft-launched AGI through o3 and o3 Mini models, which achieved nearly 90% on the ARC-AGI benchmark, surpassing human performance and exceeding expectations of Francois Chollet, creator of the benchmark.
The upgraded ARC-AGI benchmark 2 will be available to researchers for public safety testing in the coming year.
While these new models pointing towards AGI, the creator of the ARC-AGI benchmark believes that there are still some easy tasks that these models can't solve.
O3 Mini releases in January 2025.
Reinforcement Learning (RL) architecture is OpenAI's bet for scaling reasoning capabilities.
OpenAI's o3 model beat ARC-AGI benchmark on a record high score of 87.5%.
OpenAI's o3 model also achieved 71.7% accuracy on SWE Bench Verified and a 25% accuracy on the toughest mathematical test available, Epic AI Frontier Math Benchmark.
OpenAI's o3 Mini offers API features like remaining battery life, percentage probability distribution, and so on.
OpenAI is emphasizing even more on safety testing as it is proceeding towards AGI. It introduced the concept of deliberative alignment to this end.
Incubators like Y Combinator are funding startups that solve for a post-AGI world, including space tech, energy-efficient computing, government software and so on.