[NEWS] OpenAI announces "OpenAI o1", an AI that performs inference

OpenAI has announced a new model of AI, "OpenAI o1". It was made available as a ChatGPT Plus subscription on the 13th, claiming that performance and inference capabilities have increased significantly. OpenAI said in its official blog that 

"We are introducing OpenAI o1, a new large-scale language model trained with reinforcement learning and capable of performing complex inference." As a major feature, 

it claimed that "o1 thinks before answering." AI industry observers have been expecting top AI developers to roll out a new "Strawberry" model for several weeks, but the differences between the various models in development have not been made public.

OpenAI has described the new model as a major leap forward, and has changed it from the ChatGPT-3, ChatGPT-3.5, and ChatGPT-4o series to OpenAI o1.

"This is a major advancement for complex reasoning tasks, reaching a new level of AI capability. With this in mind, we have reset the counter and named it OpenAI o1," he said.

The company said these new models "take their time" to think before acting, and use "chain of thought" reasoning to perform complex tasks extremely efficiently.

According to OpenAI's shared AI test benchmarks, especially OpenAI's comparisons on tasks considered to be of doctoral level complexity, even the smallest model in this new lineup outperforms the current top-level GPT-4o in several key areas.

The new models released emphasize what OpenAI calls "deliberative reasoning," in which the system takes time to internally process a response.

This process is intended to produce more thoughtful and consistent answers, especially in reasoning-heavy tasks.

OpenAI also released test results showing significant improvements over GPT-40 in tasks such as coding, calculus, and data analysis.

The company also revealed that Open o1 showed less dramatic improvements in creative tasks such as creative writing.

In the company's tests, OpenAI's products were inferior to Claude AI in these areas. Nevertheless, the test results of the new model received great acclaim in the industry.

The new model's function implements a thought chain AI process during inference. In simple terms, it means that it uses a segmented approach to reason through the problem step by step and provide the final result that the user ultimately sees.

Genius' reasoning approach

That is, it is the same as the genius's reasoning approach.

"The o1 series is trained with large-scale reinforcement learning to reason using thought chains. Training a model to incorporate thought chains before answering can have significant benefits, but at the same time, it also increases the potential risks that arise from high intelligence," he said.

It is said that the true novelty of the model's architecture remains a matter of debate among technical experts.

OpenAI has not revealed how this process differs from token-based generation.

It is unclear whether it is an actual resource allocation for inference, or a hidden thought chain command, or a combination of both techniques.

Previously, an open source AI model called Reflection had attempted a similar inference-focused approach, but it was criticized for its lack of transparency.

Incorporating more guidelines into the thought chain process not only improves the accuracy of the model, but also makes it less susceptible to jailbreak techniques, as it provides more time and steps to catch potentially harmful results when they are generated.

It remains to be seen whether this deliberative reasoning approach can be effectively utilized for real-time applications that require high response times. 

OpenAI, meanwhile, said it plans to expand the model's capabilities, such as improving web search capabilities and multimodal interactions. 

The model will be tuned over time to meet OpenAI's minimum standards for safety, jailbreak resistance, and autonomy.

Reference: Announcement

Image: Announcement

Related articles

OpenAI announces new AI "GPT-4o" with double processing speed

Sakana AI, a Japanese company founded by a former Google researcher, raises 20 billion yen, becoming a unicorn in just one year since its founding