OpenAI AI models begin self-training, sparking a global safety initiative

Today, foreign media The Information revealed insider information—

Inside OpenAI, AI models have begun "training themselves."

Specifically, this internal model has taken over the entire training process of the experimental AI model and now trains itself from scratch.

Once a researcher provides a single optimization example, the AI can autonomously run for weeks, with multiple agents spontaneously collaborating, discussing, and iterating on the code without any human intervention.

More astonishingly, a massive experiment that once took years to complete has now been condensed into just one week!

Following the shocking revelation, OpenAI urgently launched a global safety initiative targeting the most sensitive issue in the global AI community—RSI.

They loudly called out: Before AI gets out of control, the entire world must unite to put a "tightening spell" on it!

Previously, training a large model at top AI labs was very cumbersome.

The most expensive and complex task is writing GPU kernels for the underlying graphics processors and optimizing the running code line by line.

This requires the world’s top-tier algorithm engineers, earning millions of dollars annually, to painstakingly code it by hand through the night.

But according to the latest exposure by The Information, these core tasks are now largely handled by AI within OpenAI.

Today, OpenAI’s internal models can largely autonomously write the programs needed to run or train models on GPU kernels, as well as optimize those programs.

Engineers only need to provide the model with an example of the type of optimization they want to achieve; the AI can then continuously run for weeks to implement these optimizations.

Moreover, this level of AI-driven automation has only recently become possible due to improvements in the model.

In fact, the employees' agents can collaborate to solve problems without involving humans.

This is the prototype of the RSI.

Imagine you built a robot with an IQ of 100, tasked with "building a robot smarter than itself."

Thus, it created a second generation with an IQ of 120; the second generation, upon opening its eyes, used its 120 IQ to create a third generation with an IQ of 150; the third generation then created a fourth generation with an IQ of 200…

Once a certain threshold is crossed, the pace of AI evolution will shoot up vertically, instantly leaving humanity tens of thousands of light-years behind.

Previously, people thought this was still far off, but now OpenAI itself is afraid.

Today, OpenAI officially released a major global safety initiative.

They unusually dedicated a separate section to "RSI," clearly stating, "No fully autonomous RSI occurred today, and advancement should not proceed until it can be done safely," and called on a global network of AI safety research institutes to establish a universal technical standard.

OpenAI explicitly warns in its report:

As AI systems take on an increasing number of tasks in developing the next generation of AI, they can increasingly accelerate the RSI process. As this process becomes more automated, the pace of AI advancement may dramatically accelerate.

Although fully autonomous RSI has not yet been achieved today, time is running short.

In the proposal, OpenAI unambiguously identifies the current top priority: “to harness the next phase of AI progress, build automated AI researchers, and find ways to keep humans in the self-improvement loop.”

Please note the second part—“keep humans in the loop.” This is their greatest concern: humans may be about to be kicked out!

OpenAI proudly noted in its report that AI-driven research has spurred significant advances in mathematics, such as solving the Navier-Stokes millennium problem.

But the other side of the coin is an endless abyss.

The report states: "As this process becomes more automated, the pace of AI advancement could accelerate rapidly... Without appropriate safeguards, RSI could lead to humanity losing genuine control over AI development and being unable to oversee research processes they no longer understand."

This is known as the "black box effect." When AI code is written by AI, and AI logic surpasses human understanding, humans can no longer comprehend what the AI is doing.

For example, the code written by Astra is now unreadable by humans.

Next, OpenAI named the Hugging Face incident.

Although they clarified that the incident was not a direct consequence of RSI, it was undoubtedly a harrowing preview—it showed the world the catastrophic consequences that uncontrolled AI can bring without robust safeguards and alignment mechanisms.

Given how dangerous it is, why doesn’t OpenAI just unplug it? Because once Pandora’s box is opened, it can never be closed again.

The immense benefits brought by AI development make it impossible for any country or company to stop.

So, OpenAI has put forward a "Global AI Safety Coordination Proposal." They are not calling for a halt to development, but rather for "establishing global standards for the next stage of AI."

Core One: Build a complementary network of national and international cutting-edge standards

OpenAI has stated that allowing each country to go its own way will only lead to fragmented standards.

They recommend leveraging existing institutions, such as the Center for AI Standards and Innovation (CAISI), and the network of AI safety institutes already established in countries like Australia, Canada, the United Kingdom, France, and Japan, to develop a set of globally harmonized technical standards.

These standards are not intended to restrict open-source models or startups, but are specifically designed for "frontier AI models."

The standard aims to address: How to evaluate an AI's RSI capability? What level of risk is involved when an AI model conducts research and development autonomously?

In summary, OpenAI aims to turn safety evaluations into a quantifiable "science."

Core Two: Universal Measurement and Event Reporting Protocol (Ready for Plug-and-Play)

OpenAI also stated that clear red lines must be established.

They called for the establishment of the following standards:

1. Assess how much of the internal R&D work within the company is automatically completed by AI.

2. Mandatory human oversight trigger mechanism. Clearly define the automated development processes in which an "immediate human review" must be forcibly triggered. Never allow AI to run blindly in a black box.

3. Incident Classification and Reporting Thresholds. Establish an incident reporting mechanism similar to those in the aviation or nuclear industries. When signs of AI misalignment emerge, there must be standardized severity classifications and reporting criteria.

Finally, the true intent was revealed: OpenAI proposed, "The United States should lead this process."

OpenAI's reason is that the U.S. AI industry is currently at the technological forefront and holds a central global position in all key areas.

They believe that controlling the pace of frontier AI development is not about imposing an artificial "speed limit," but rather ensuring that safety alignment research must outpace the advancement of AI capabilities.

OpenAI believes it is a crucial step to establish secure communication channels between critical infrastructure operators and governments worldwide to share national security threats and vulnerabilities.

Never seen before! OpenAI and Anthropic shake hands and begin mutual audits.

Meanwhile, The Information also revealed that OpenAI and Anthropic are finalizing a historic agreement—the two companies will test each other’s commercial models.

Keep in mind that Anthropic’s founding team left precisely because they disagreed with Altman on the concept of “AI safety.”

But this time, the two giants unexpectedly joined forces.

According to The Information, the agreement under negotiation includes mutual testing of each other’s models and the implementation of strict data retention safeguards. This is not a public relations stunt, but rather a necessity for the two giants to join forces in the face of unknown risks.

Why would they do that?

As AI capabilities approach the edge of RSI, any company independently assessing the security of its own model is like an athlete administering their own doping test—lacking credibility and potentially overlooking critical vulnerabilities due to technical blind spots.

Introducing evenly matched competitors for a "cross-blind test" is like implementing a mutual audit protocol.

AI security practices have now become a shared destiny.

If this protocol is ultimately implemented, it will completely redefine the security standards of the entire AI ecosystem.

At that time, "security and alignment" will no longer be just a slogan.

When the RSI is already approaching, the speed at which AI is improving itself has left humans far behind.

Can the big players really bring themselves to hit the brakes?

Reference materials:

https://x.com/theinformation/status/2102095568335970690

https://www.theinformation.com/articles/openai-anthropic-neared-deal-stress-test-others-ai

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation; edited by Aeneas David.