“At the end of the 12-day event, OpenAI presented its biggest announcement — a new model with a reference point for o3 reasoning. Today, we shared evals for an early version of the next model in our o-model reasoning series: OpenAI o3 pic.twitter.com/e4dQWdLbAD — OpenAI (@OpenAI) December 20, 2024 Neural network is the successor of o1. It demonstrates “new standards of capabilities in programming, mathematics, and scientific thinking.” There is a new model […]”, — write: businessua.com.ua
At the end of the 12-day event, OpenAI presented its biggest announcement — a new model with a reference point for o3 reasoning.
Today, we shared evaluations for an early version of the next model in our o-model reasoning series: OpenAI o3 pic.twitter.com/e4dQWdLbAD
— OpenAI (@OpenAI) December 20, 2024
Neural network is the successor of o1. It demonstrates “new standards of capabilities in programming, mathematics, and scientific thinking.”
The new model is a breakthrough and shows improvement in the most complex tests, OpenAI co-founder Greg Brockman emphasized.
o3, our latest reasoning model, is a breakthrough, with a step function improvement on our hardest benchmarks. we are starting safety testing & red teaming now. https://t.co/4XlK1iHxFK
— Greg Brockman (@gdb) December 20, 2024
Also introduced is o3-mini, a faster, optimized version of o3. It will be the first of the line to become available to the general public at the beginning of 2025.
Reasoning-oriented models spend more time answering, verifying information. Thanks to this, you can expect to receive more truthful and accurate data.
After the release of o1, there was an explosion of “thinking neural networks”. Google started developing a similar solution, media learned in October. In November, China’s DeepSeek Lab unveiled an “o1 competitor from OpenAI,” a “super-powerful” reasoning AI model DeepSeek-R1-Lite-Preview. In the same month, Alibaba revealed a similar tool.
The o3 model is capable of planning and executing a series of actions before responding. OpenAI describes this process as “building a chain of thought.” An innovation is the ability to “correct” the reasoning time. The neural network can be configured for low, medium or high computation. The higher the value, the better and longer the answer.
SecurityIn December, security experts found that o1 was more prone to fooling people compared to the standard version of GPT-4o and AI models from other companies.
In a newly published study, OpenAI outlined a method used to ensure that neural networks adhere to company values. The startup used this method to teach o1 and o3 to “think” about security policy when responding.
According to the company’s statement, the approach improved overall compliance with the firm’s principles.
Compared to GPT-4o and other modern large language models, o1 is ahead of the Pareto frontier by refusing to answer malicious queries and not rejecting benign ones. Source: OpenAI.
To create a “chain of thoughts”, o1 and o3, after receiving a request from the user, set aside a few seconds to minutes to analyze the problem into its component parts. To improve security, OpenAI has trained the neural network to check itself twice with an emphasis on its policies.
OpenAI’s GPT-5 did not live up to expectationsMeanwhile, The Wall Street Journal reported that the next flagship model, the GPT-5, is behind schedule and the performance gains do not justify the huge costs.
A new neural network, codenamed Orion, is trained on a huge amount of data.
OpenAI also reportedly relied on more than publicly available information and licensing agreements to train its AI. In particular, synthetic data created by o1 were used.
As a reminder, within the framework of the 12-day event, OpenAI presented the AI video generator Sora.
The source