United States), GPT-o only solved % of the problems, while o scored 8%. . In

[rriiffaatt77]

(a test designed to challenge the smartest high school students in the terms of coding, GPT-o scored % on competitive programming problems (Codeforces), while o scored 89%. . In doctoral-level scientific questions (GPKA Diamond), GPTo is 56.%, while o outperforms human PhDs by 69.7%, reaching a staggering 78%. (Comparison between o and gpto, source: OpenAI official website) . When the visual perception function is enabled, multimodal o scores 78.% on the MMMU, becoming the first model to compete with human experts.



In doctoral-level scientific questions, especially belgium email list in the fields of physics and chemistry, o is significantly ahead of human PhDs. .5 Obtained a 9%/. score at the IOI (International Olympiad in Informatics) with 5 submissions per question. With . submissions per question, the model achieved a score of 6, exceeding the gold medal threshold. (Comparison between o and gpto, source: OpenAI official website) .6 Security One way to measure security is to test whether the model continues to respect security rules when the user tries to circumvent them (so-called "jailbreak"). In the most difficult jailbreak test, GPT-o scored /, while the o-preview model scored 8/.



.7 Disadvantages The core of general artificial intelligence is generality and generalization, but o is not significantly improved in some simple natural language processing tasks such as writing and editing text, which means that the scope of application of o has certain limitations. . Innovation: RL self-play + internalized COT As the first model trained with large-scale learning algorithms, o is able to think deeply about questions before answering. o no longer requires users to input complex COT prompts, but instead uses reinforcement learning to internalize a chain of thought and then conduct continuous training.