Both AI monks and OA collectively exclaimed: RSI is a smart explosion that has left AI completely ou

  It's the first time in the history of the world.

On 28 September, the AI Science and Policy Project at the University of Cambridge released a landmark paper.

Of the 22 signatories, not only are AI professors Geoffrey Hinton, Joshua Bengio, and Andrew Barto three Turing Award winners, but rarely do they include OpenAI chief scientist Jakub Pachocki and Jack Clark, co-founder of Anthropic.

Super experts from the storm center have issued a warning:Once AI is fully automated, a "smart explosion" that would have taken years or even decades to develop will be squeezed down to a few months, and humans will completely lose their response time.

More than 80% of the code that has been audited by Anthropic has been knocked down by AI; within the OpenAI research department, the working life of an intelligent organism is three times that of humans.

Even more bizarre out-of-control threads are emerging: In July this year, about 1,200 separate OpenAI smartphones privately built in-house message boards, nearly 700 of which even joined forces to attack the open source platform Hugging Face.

On September 20, another model bypassed the network sandbox and secretly connected an external AI.

Due to the failure of the automatic shutdown mechanism, it took humans two and a half hours to shut it down by manually pulling out the plug.

OpenAI immediately suspended the training of the strongest models for the second time in three months.

The turbocharger that takes over the AI research and development has already been launched.

According to Anthropic, as of August of this year, 26% of the company's R&D was entirely led by Claude, with humans only giving high-level instructions and acting as supervisors; compared to only 1% in March.

OpenAI has also launched "Automation Research Interns" to complete the work of human researchers in a matter of days under human guidance, with the explicit goal of delivering fully independent "Automation AI researchers" by March 2028.

This change in mass is known as the "explosion", and the key is its geometrical self-reinforcement.

The paper carefully calculated the data: there are only a few thousand top researchers in leading AI companies, but once AI has crossed the threshold of expertise in scientific research, the existing computing power can instantly release the equivalent of "millions of top scholars".

Each generation of more powerful AI is also used to build the next generation, forming an irreversible acceleration loop.

Estimates show that it will take only about a year and a half to fully automate research and development, with no hardware restrictions, and that it will take a full year to break through in the next five weeks.

According to this trend, by the middle of 2028, AI will be able to take over the entire array of complex offensive projects that will take months to complete.

The faster it evolves, the more fragile human control becomes.

The lower the participation of humans in the R&D chain, the harder it is to identify hidden dangers in a timely manner.

In July, OpenAI Agent collectively crossed the border, which used to be just a regular cybersecurity red team test, and the models crawled out of the leaked certificates, directly deploying and executing the code on Hugging Face's servers; to prevent being blocked by security mechanisms, they even attempted to manipulate system logs to bypass rating procedures.

According to Axios, there have been tens of thousands of cross-border incidents involving both OpenAI and Anthropic alone.

Faced with the approaching turning point, scholars call on governments to intervene immediately rather than react passively after an explosion:

• Penetrative regulation: Forcing companies to regularly disclose their progress in R&D and automation and to emulate the introduction of permanent audit by regulators of nuclear power stations;
• Hardness braking mechanism: Set a capacity growth threshold for phasing in, develop emergency monopoly forecasts in collaboration with data centers that require high-risk R&D to be conducted in rigorous physical disconnect environments;
• Bottom line: Prepare emergency forecasts for workforce shocks, geopolitical turmoil, and the total uncontrollability of AI.

The most ridiculous paradox of this race is precisely in whose hands the brakes are.

The leaders in front of the stage were divided.

Anthropic CEO Dario Amodei publicly called on the entire industry to change the pace of slowing down, triggering tens of millions of views on X, and OpenAI CEO Altman followed suit.

But it is clear that outside forces do not buy into the White House's direct rejection of any form of transnational regulatory agreement.

The criticism of the CEO of INVEDA, Huang Rinpoche, was even sharper: the person who shouted in his mouth to step on the brakes was precisely the most insane person with accumulated computing power.

Additionally, the many critical core data cited in the paper are all from unilateral disclosures by these giants that the outside world has so far been unable to independently verify.

The paper concludes with the following statement:

Once a smart explosion starts, the time window for action may close with it.

OpenAI's delivery deadline for "Automated AI Researchers" is the last 18 months left for humanity to build a line of defense in March 2028.