OpenAI emphasized that the brand new model ‘o3’ was released just three months after the launch of ‘o1’. The model upgrade cycle has been significantly reduced because of latest inference technology. It was emphasized that inference based on ‘test-time compute’ is a far more effective upscaling method than existing pre-training.
Open AI research scientist Noam Brown said on the twenty first (local time) through “There’s,” he said.
OpenAI researcher Jason Wei also said, “O3 has excellent performance, but what’s more necessary is that it only took three months to develop from o1 to o3.”
He continued, “This shows how quickly progress will be made in the brand new paradigm of reinforcement learning (RL), which expands inferential calculations in chains of ideas,” and added, “It is way faster than the prevailing pre-training paradigm for models that took 1 to 2 years.”
OpenAI released ‘o1-Preview’ and ‘o-1 Mini’ on September twelfth, and released ‘o1’ full version and ‘o1-Pro’ on the primary day of Shipms on the fifth. o3 was released on the twentieth, the last day.
On the time of o3’s release, attention was paid to the indisputable fact that its performance was near that of artificial general intelligence (AGI), and there was no give attention to the discharge interval. Then, the staff who participated in the event directly emphasized this.
This contrasts with the indisputable fact that GPT-5, generally known as ‘Orion’, didn’t show significant performance improvement despite two rounds of preliminary training. It has been reported that OpenAI has been developing GPT-5 for 18 months now.
Alternatively, the inference model showed a dramatic improvement in performance in only three months. Although the particular specifications of the model weren’t revealed, benchmarks showed that clear performance improvements were achieved.
Subsequently, it has been proven that the inference model is a ‘latest scaling law’ that can replace solid training. That is what SSI founder Ilya Sutskever, Microsoft CEO Satya Natella, NVIDIA CEO Jensen Huang, and Professor Yoshua Bengio of the University of Montréal have all announced.
As well as, the evaluation that researcher Brown said that “the trajectory will proceed” signifies that a brand new model may emerge early next yr. There was also speculation that ‘o4’ may very well be Orion.
Meanwhile, OpenAI researcher John Holman also left a powerful comment that day. “When Sam and our researchers say AGI is coming, they’re not attempting to sell $2,000 subscriptions or idiot investors,” he said. “(AGI) is definitely coming.”
Likewise, AGI was mentioned twice this month by OpenAI employees. On the seventh, OpenAI technical staff member Bahit Kazemi posted for the primary time on
o3 has not yet been released to the general public, and red team testing will begin next month for a limited number of individuals. Subsequently, it appears that evidently it is going to take more time for cases of experiencing AGI-level performance through o3 to emerge.
Reporter Lim Da-jun ydj@aitimes.com
