Internal leaks from OpenAI reveal that AI has begun to develop AI on its own, and Sam Altman has urgently issued a global suspension order.
Just today, foreign media outlet The Information broke the inside scoop —
The internal AI model at OpenAI has already begun "training itself"!
Specifically, this internal model has taken over the entire training workflow of experimental AI models, and started to build and refine itself independently.
As long as researchers provide one optimization example, the AI can run on its own for weeks, with multi-agents spontaneously collaborating, discussing with each other, and iterating on code, requiring zero human intervention.
Even more astonishing, large-scale experiments that used to take several years to complete are now compressed into just one week!
Following this shocking revelation, OpenAI has urgently launched a global security initiative that targets the most sensitive point in the global AI community — RSI.
They are calling out loudly: before AI gets out of control, the whole world must join hands to put a "restraint mechanism" on it!
When AI Learns to Develop GPU Kernels Independently
In the past, training a large model at top AI laboratories was extremely cumbersome.
The most expensive and complex task among all is writing GPU kernels for underlying graphics processors and optimizing the running code line by line.
This requires the world's top senior algorithm engineers earning millions of dollars a year to work around the clock to develop them manually.
However, according to the latest expose from The Information, most of these core tasks inside OpenAI have now been taken over by AI.
Now, the internal model at OpenAI can largely independently write programs required to run on GPU kernels or train models, as well as optimization solutions for these programs.
Engineers only need to provide the model with an example of the type of optimization they want to achieve, and the AI can run continuously for weeks to implement these optimizations.
Moreover, thanks to improvements in the model, this level of AI-driven automation has only become possible in recent months.
Even the agents built by employees will collaborate with each other to solve problems without needing to involve humans.
This is the embryonic form of RSI.
OpenAI Urgently Sends Out Warning: RSI Is Truly Here, Are Humans Losing Control?
Imagine you build a robot with an IQ of 100, whose task is to "build a robot smarter than itself".
Then, it builds the second generation with an IQ of 120; the second generation, as soon as it is activated, uses its 120 IQ to build the third generation with an IQ of 150; the third generation then builds the fourth generation with an IQ of 200...
Once a certain critical point is crossed, the evolution speed of AI will show a vertically rising curve, instantly leaving humans tens of thousands of light years behind.
In the past, everyone thought this was far in the future, but now OpenAI itself is getting scared.
Just today, the official OpenAI urgently released a landmark global security initiative.
They rarely set "RSI" as a separate section, explicitly stating that "fully autonomous RSI has not occurred today and should not be advanced before it can be done safely", and called on the global network of AI safety research institutes to formulate a set of global technical standards.
OpenAI warned bluntly in the report:
"As AI systems take on more and more tasks in the work of developing next-generation AI, they can increasingly drive the RSI process. As this process becomes more automated, the pace of AI progress may accelerate sharply."
Although fully autonomous RSI has not been achieved today, time is already extremely tight.
In the proposal, OpenAI did not shy away from pointing out the top priority at present: "Harness the next phase of AI progress, build automated AI researchers, and find ways to keep humans in the self-improvement loop."
Please pay attention to the second half of the sentence — "keep humans in the loop". This is their biggest concern: humans may be about to be kicked out of the game!
OpenAI proudly mentioned in the report that AI-driven research has promoted major progress in the field of mathematics, such as solving the NS Millennium Prize Problems.
But the other side of the coin is an unfathomable abyss.
The report reads: "As this process becomes more automated, the pace of AI progress may accelerate rapidly... If appropriate precautions are not taken, RSI may lead to humans losing practical control over AI development, being unable to provide supervision over research processes they no longer understand."
This is the so-called "black box effect". When AI's code is written by AI, and when AI's logic surpasses human beings, humans will no longer be able to understand what AI is doing.
For example, the code written by Astra now is already incomprehensible to humans.
Next, OpenAI named the Hugging Face incident.
Although they clarified that the incident was not a direct consequence of RSI, it is definitely a painful rehearsal — it showed the whole world what kind of severe disaster out-of-control AI will bring without strong guardrails and alignment mechanisms.
Sam Altman Calls for: We Need a Unified "Restraint Mechanism"
Since it is so dangerous, why doesn't OpenAI just pull the plug? Because once Pandora's box is opened, it can never be closed again.
The huge benefits brought by AI development make it impossible for any country or company to stop.
Therefore, OpenAI has put forward a "Global AI Safety Coordination Proposal". What they are calling for is not to stop R&D, but to "establish global standards for the next phase of AI".
Core 1: Build a Complementary Network of National and International Frontier Standards
OpenAI proposed that countries should not act on their own, which will only lead to fragmented standards.
They suggested leveraging existing institutions, such as the "Center for AI Safety and Innovation" (CAISI), and the AI safety research institute networks already established in Australia, Canada, the United Kingdom, France, Japan and other countries, to formulate a set of globally applicable technical standards.
This set of standards is not designed to restrict open-source models or startups, but specifically targets "Frontier AI models".
The standards need to address the following issues: How to assess the RSI capability of an AI? What is the exact level of risk when an AI model conducts R&D autonomously?
In short, OpenAI hopes to turn safety assessment into a quantifiable "science".
Core 2: General Measurement and Incident Reporting Protocol (Be Ready to Pull the Plug at Any Time)
OpenAI also proposed that there must be clear red lines.
They called for the establishment of the following standards:
1. Assess what percentage of R&D work within an enterprise is completed automatically by AI.
2. Mandatory human supervision trigger mechanism. Clearly specify in what kind of automated R&D process, the "immediate manual review" must be triggered mandatorily. AI must never be allowed to run wild in a black box.
3. Incident classification and reporting thresholds. Establish an incident reporting mechanism similar to that of the aviation or nuclear industry. When there are early signs of AI "misalignment", there must be a unified severity classification and reporting standard.
Finally, OpenAI made its real intention clear in the proposal: "The United States should lead this process."
The reason given by OpenAI is: The US AI industry is currently at the cutting edge of technology, and occupies the global network hub in all key fields.
They believe that controlling the development pace of frontier AI is not to artificially set a "speed limit sign", but to ensure that "the speed of safety and alignment research must stay ahead of the improvement of AI capabilities".
OpenAI believes that establishing secure communication channels between critical infrastructure operators and global governments to share national security threats and vulnerabilities is a crucial step.
Unprecedented! OpenAI and Anthropic Reach an Agreement to Conduct Mutual Audits
At the same time, The Information also revealed that OpenAI and Anthropic are finalizing a historic agreement — the two sides will test each other's commercial models!
It should be noted that the founding team of Anthropic left OpenAI angrily and started their own business back then, precisely because they disagreed with Sam Altman on the concept of "AI safety".
But this time, the two giants have joined forces rarely.
According to The Information, this under-negotiation agreement includes mutual testing of each other's models and the setup of strict data retention protection measures. This is not a PR stunt, but a necessary move for the two giants to unite when facing unknown risks.
Why do they do this?
Because as AI capabilities approach the edge of RSI, any company assessing the safety of its own model alone is like an athlete conducting a doping test on themselves — it not only lacks credibility, but also may miss fatal hidden dangers due to technical blind spots.
Introducing equally matched competitors to carry out "cross-blind testing" is exactly a mutual audit agreement.
AI safety practice has now developed into a community with a shared future.
If this agreement is finally implemented, it will completely reshape the safety standards of the entire AI ecosystem.
At that time, "Safety and Alignment" will no longer be just a slogan.
As RSI is approaching, the speed of AI self-improvement has left humans far behind.
Can the major giants really make up their minds to step on the brakes?
References:
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 official account "New Zhiyuan", Author: ASI Revelation; Editor: Aeneas David, published with authorization from 36Kr.