Returning to his senior year of high school, he went all in and made billions by buying at the botto

Chapter 328 Basic Model

Chen Mo looked at the screen, making no attempt to hide the satisfaction in his eyes.

OpenAI is a relatively new company.

The team's rapid progress is partly due to its unreserved investment in computing power and talent.

More importantly, they did not follow the most popular route of 2015, building separate models for each task.

From the very beginning, Chen Mo proposed that a large amount of unlabeled data, self-supervised training, unified representation, and cross-task transfer are the keys to achieving stronger artificial intelligence.

Although he can't write code, he can help the team avoid detours that have proven to be of limited value in the coming years.

Musk leaned back in his chair, his eyes brimming with pride: "It turns out that the right path works far better than hiring a thousand more Google engineers."

"It only gives us a temporary lead." Chen Mo wasn't blinded by the demonstration results. "Once Google confirms that this route is effective, they can quickly mobilize more computing power and personnel than we can."

"To turn a lead into a real advantage, the next step is to consolidate scattered research."

He stood up and walked towards the whiteboard.

Currently, OpenAI still considers language, image, games, and robotics as several relatively independent research directions.

The team has started sharing some underlying methods, but there is no unified long-term definition.

Chen Mo picked up a pen and wrote four characters in the center of the whiteboard.

Basic model.

Most of the people at the scene were seeing this word for the first time.

"The most important artificial intelligence of the future will not be specialized programs that can only play chess, recognize faces, or answer a certain type of question."

"It should first learn the most basic structures of language, images, and the real world through massive amounts of data."

"Then, it can adapt to different tasks with only a small amount of data, instructions or environmental feedback."

"It is closer to a platform that carries intelligent capabilities."

Sam stared at the whiteboard: "If we follow this logic completely, the current two thousand graphics cards won't even get close to the threshold of the next-generation architecture, and the computing power cost will increase several times every year."

"This is the second problem we need to solve today." Chen Mo sat down again.

The computing power expansion plan was projected onto the screen.

OpenAI currently has over two thousand GPUs and has established partnerships with several cloud computing companies.

This scale would be considered a luxury in a typical research institution.

Faced with a basic model, this is only the first step.

The technical team hopes to quadruple the available computing power in the next eighteen months, while also building its own distributed training system.

The total investment is expected to exceed $300 million.

Musk was the first to speak out: "I support it. If we slow down because we're worried about spending money, Google won't give us a second chance."

Lu Jingyi opened the budget: "Money can solve the problem. But three hundred million US dollars is not a one-time expense."

She looked at Sam.

"OpenAI relies on long-term donations from its founders, which may sustain it for three years, but it may not be able to sustain it for ten years."

"If the basic model truly becomes the platform you define, commercialization will happen sooner or later."

This is also why there is a third solution on the desktop.

Relying entirely on donations, OpenAI will find it difficult to compete with companies like Google and Microsoft in terms of computing power in the long run.

Premature commercialization could lead to research directions being controlled by income and investors.

Yang Zhilin recalled the issues raised at the Tsinghua University dinner party.

"If commercialization is needed in the future, which achievements will remain available?"

"Paper, code, training data, or the complete model?"

"It's completely open; any company can take away the results of OpenAI's billions of dollars of training."

"If it becomes completely closed off, OpenAI will just become another ordinary technology company."

Sam did not answer immediately.

This is precisely the contradiction he has been unable to completely resolve despite submitting organizational governance plans.

Liang Wenfeng spoke for the first time: "The problem may not lie in whether it is open or closed."

Everyone looked at him.

"It depends on how we view artificial intelligence."

Liang Wenfeng pointed to the basic model on the whiteboard.

"If it's just software, the simplest way is to sell licenses."

"If it does become a platform, then model capabilities, computing power, and data will together form a new infrastructure."

"Infrastructure can provide services to external parties, but it cannot hand over all underlying control to a few customers."

He had seen similar situations in the quantitative market.

A single strategy can be sold.

Once a complete trading system is handed over, the recipient no longer receives a product, but rather the ability to continuously develop strategies.

Chen Mo made the definitive statement: "The papers, toolchains, and security specifications are all open source, while core parameters and high-risk capabilities are under closed-door control."

"OpenAI is not a company; it is the nerve center that brings together computing power, universities, and industry."

"We don't need to turn all the best people into employees."

"It can become a research network."

"Universities, laboratories, and enterprises retain their own data and intellectual property rights, and only collaborate on basic methodologies, computing tools, and security research."

The meeting lasted from the afternoon until evening.

The whiteboard is filled with more and more content.

Basic model, long-term memory, cross-task transfer, human feedback, tool usage, model safety.

In 2015, most of these terms had not yet become the common language of the artificial intelligence industry.

Today, however, it was included in the same route map for the first time.

Ultimately, the initiating committee approved the next phase of the computing power expansion plan.

The $300 million will not be disbursed all at once.

The first phase will be used to expand the GPU cluster and distributed training system, with subsequent funding released gradually based on model capabilities, training efficiency, and security assessments.

The OpenAI-Carnegie Mellon Language Intelligence Joint Project has officially entered the contract phase.

Deep Exploration has been incorporated into a collaborative research network, establishing an independent AI research node in the country.

As for the most sensitive organizational issues, no one forced a conclusion within a day.

Chen Mo only managed to get a provisional principle passed.

Research papers, general-purpose tools, and low-risk code remain open.

Complete model parameters, large-scale training data, and the potential for misuse require joint evaluation by technical, security, and governance stakeholders.

No single initiator, donor, or commercial entity can unilaterally decide to release or seal a core achievement.

Whether a business entity should be established, how to limit capital returns, and where ultimate control belongs will be discussed in the next stage.

Musk looked at the provisional guidelines and did not object.

Sam also signed his name.

But Chen Mo knew that the disagreement had not disappeared.

As the models become more powerful and the investment increases, the problems that we temporarily set aside today will eventually come back to everyone's attention.

The meeting ended close to 7 p.m.

The others left the meeting room one after another.

Ilya did not get up immediately.

He replayed the language model demonstration from the afternoon.

"List."

"There's something that's not appropriate to put too explicitly in a formal report."

Chen Mo stopped in his tracks.

"you say."

"Google's overall strength still far surpasses ours."

Ilya stared at the answer on the screen.

"But just looking at unified pre-training and cross-task transfer, this prototype already surpasses the internal projects I saw before I left."

"As long as the next generation of training is successful, Google will no longer be the absolute leader."