OpenAI officially announced the launch of "next-generation cutting-edge model" training! It is expected that the training parameters will be further improved, or the "Wensheng video" model Sora will be integrated

National Business Daily

2024-05-29 14:02Published on the official account of Sichuan Daily Economic News

Reporter: Cai Ding Editor: Lan Suying

On Tuesday (May 28), local time, OpenAI announced the establishment of a safety and security committee by the board of directors. In the announcement, the company also announced that it has begun training its "next frontier model" and expects the resulting system to take us to a higher level of capability on the road to AGI.

In the blog post, OpenAI said, "While we pride ourselves on building and releasing models that lead the industry in both capabilities and security, we welcome the intense discussion around this issue at this important time." The company said it had set up the new committee to assess any risks that could arise from the "new frontier model."

Although OpenAI's so-called "next-generation cutting-edge model" is widely expected to be GPT-5, the company has not made any official statement so far. And according to a speech by OpenAI's head of developer experience, Romain Huet, in Paris last week, OpenAI's next-generation model may be named "GPT-Next". However, some foreign media believe that this is actually Omni-2 - a more refined, more training parameters, and stronger enhanced version of GPT-4o.

It is not expected to be named with GPT-5 

OpenAI said the top priority of the newly formed "Safety and Security Committee" "will be to evaluate processes and safeguards for further development of OpenAI over the next 90 days." Subsequently, OpenAI will publicly share an update on the "adoption of recommendations" following a review by the board. This means that OpenAI's board of directors should receive the committee's recommendations by August 26 of this year at the latest.

Image source: OpenAI

If calculated according to such a timeline, OpenAI could release what it calls a "next-generation cutting-edge model" as soon as this fall. However, the New York Times reported that the training of AI models could take months or even years. Once the training is complete, the AI company typically spends a few more months testing the model and fine-tuning it before it can be made available to the public. This means that OpenAI's next-generation model may not be released in 90 days.

In a recent "Security Update" released during the International AI Summit in Seoul, OpenAI also said that the company would spend more time evaluating the capabilities of the new model before it was released, which could explain why there is no specific release date.

OpenAI did not explicitly mention the name of the model under training, but the outside world immediately speculated that this "cutting-edge model" was GPT-5.

Previously, OpenAI's underlying model had followed a clear naming order, i.e., GPT changes in the form of incrementing numbers. However, it should be noted that the model released by OpenAI this month is named GPT-4o.

The reporter of "Daily Economic News" found that when demonstrating ChatGPT Voice at the VivaTech conference held in Paris on May 22 local time, Roman Way, director of developer experience at OpenAI, showed a slide showing the potential growth trend and naming of artificial intelligence models in the next few years, and there was no GPT-5 among them. 

Roman Way showed the GPT-3 era, the GPT-4 era, GPT-Next, and the "model of the future", and "Today" is between the GPT-4 era and GPT-Next. This seems to mean that the next-generation model launched by OpenAI may be named GPT-Next, but some foreign media believe that this is actually Omni-2 - a more refined, more training parameters, and stronger enhanced version of GPT-4o.

Moreover, Roman Wei also said in his speech that in the coming months or this year, people will see more types of OpenAI models, not just ChatGPT models.

Image source: OpenAI

The iterative trend heralds the integration of Sora with next-generation models

Forbes reports that OpenAI's so-called "next-generation cutting-edge model" could replace its leading base model, GPT-4, as well as powerful generative AI tools such as image generators, virtual assistants, search engines, and flagship chatbot ChatGPT.

TECH MEDIA ZDNET REPORTED THAT FOLLOWING PAST TRENDS, THE OUTSIDE WORLD CAN EXPECT THE "NEXT-GENERATION CUTTING-EDGE MODEL" TO RESPOND MORE ACCURATELY BECAUSE IT WILL BE TRAINED ON MORE DATA. The more data a model can train on, the better the model will be able to generate coherent, accurate content, and therefore better performance.

For example, some studies say that GPT-3.5 is trained on 175 billion parameters, while GPT-4 has already reached 1 trillion training parameters. According to the report, if OpenAI's "next-generation cutting-edge model" is released, we may see a bigger leap in the model's training parameters.

Looking back at OpenAI's past basic models, from GPT-3.5 to the latest GPT-4o, it can be seen that with the upgrade of the model, the iteration of the product has made it more intelligent, including price, response speed, context length, etc. You know, GPT-3.5 can only input and output text, and by the time of GPT-4 Turbo, text output can already be obtained through the input of text and images. GPT-4o has demonstrated its ability to not only process text, but also create image content, and answer questions and respond to instructions in a highly conversational manner.

ZDNET reported that according to the iterative trend of OpenAI's basic model, the "next-generation cutting-edge model" may have the ability to output video. In February, OpenAI released Sora, a "Wensheng video" model, which could be integrated into a "next-generation cutting-edge model" to output video.

OpenAI's goal is to advance AI technology faster while outpacing its competitors, while also plaguing critics who believe the technology is increasingly dangerous, fueling the spread of disinformation, displacing jobs, and even threatening human existence. While OpenAI said in a blog post on Tuesday that it expects its "new frontier model" to take us to a higher level of competence on the road to AGI, a form of AI capable of understanding, learning, and performing any intellectual task that a human can perform. However, experts disagree on the feasibility of AGI and when it may be realized. Geoffrey Hinton, known as the godfather of AI, estimates that it could take 20 years to develop AGI. Nvidia CEO Jensen Huang predicts that AI will pass human testing within five years, and AGI will come soon.

National Business Daily

View original image 138K

• 

•