OpenAI reveals recursive self-improvement: intelligent agents experience exponential growth, and humanity has become a bottleneck

Nov 19, 2026 21:48:31

Author: ASI Revelation, New Intelligence

OpenAI has announced without warning: "The gears of Recursive Self-Improvement (RSI) have officially begun to turn."

In simple terms, OpenAI has confirmed: AI has started to develop AI!

For a long time, people have speculated about the day when super artificial intelligence (ASI) would arrive. Everyone has been watching the scores of large models and the surge in parameter counts.

But OpenAI clearly tells you: The real turning point towards ASI is not when model scores surpass humans, but when the "AI developing AI" closed loop takes shape.

Computing Power Frenzy, Burning $23,000 a Day for "Super Workers"

Internally at OpenAI, the frequency with which researchers use AI agents has completely spiraled out of control—of course, in an upward direction.

Based on API pricing, the spending on agents in September this year saw the median researcher spending $1,500 per day.

Just a month later, in October, this number skyrocketed to $2,400, an increase of 62%.

But this is just the average level. What about the top 10% of "heavy users"? They are burning up to $23,000 a day on agents, a staggering increase of 105%.

Over $20,000 a day. If you have a sense of API pricing, you will understand the massive amount of code generation, logical reasoning, and concurrent task execution this represents. This is not humans typing on keyboards; this is hundreds and thousands of "silicon brains" operating tirelessly at full speed.

To measure this more intuitively, OpenAI introduced a concept: "Agent Workdays" (each 8 hours of agent use counts as one workday).

Data shows that the average "agent workdays" per researcher has terrifyingly surged from an initial 0.5 to 5.7, an increase of 11 times!

In the core pre-training and post-training teams, this number has reached 8.

This means that a human researcher works 8 hours a day, while their "AI avatars" concurrently complete the equivalent of 64 hours of work in the same time.

By mid-October, over one-tenth of researchers had at least 16 agents simultaneously handling tasks every day.

Human engineers' reliance on these AI tools has reached an almost "addictive" level.

In an anonymous survey conducted internally at OpenAI, researchers were asked a very pointed question: "If in the next three months, you were to lose all your coding agents superior to GPT-5.2, what conditions would you offer to get them back?"

The median response was shocking: They would rather cut 40% of their team or reduce the current team's core computing power by 30%.

In the eyes of the top AI researchers, a few expensive H100 computing clusters, even living colleagues, are no longer as valuable as a cutting-edge coding agent.

The Truly Chilling Detail: AI Begins to Have "Research Taste"

What truly made Silicon Valley elites exclaim "madness" is the development secrets of GPT-6.2 Astra.

In the past, big ideas were always the exclusive domain of humans.

When developing earlier versions of GPT-6 Astra, OpenAI researchers proudly stated that the core scientific ideas were entirely conceived by humans, with the model merely acting as a high-level coder responsible for turning ideas into code.

But by the time of GPT-6.2 Astra, a subtle but fatal change occurred.

In interviews, researchers from the pre-training team admitted that core insights were born from repeated dialogues with the model (Codex).

One researcher explicitly stated that Codex accurately identified key insights while they were collaboratively advancing an idea.

Another core member confessed that their initial idea hit a dead end, "Only after Codex suggested making changes did the entire idea truly start to work."

AI has begun to develop intuition, recognizing blind spots overlooked by humans, and even guiding humans on how to create better AI.

This is referred to as "Research Taste."

Identifying problems is often more challenging than solving them. When AI possesses the ability to discern "what kind of data is better" and "what kind of architecture is more promising," Pandora's box is truly opened.

A developer tweeted to describe this feedback loop:

  1. A researcher wants the AI model to be stronger in logical reasoning.
  2. They summon dozens of agents to write underlying code, configure environments, run massive experiments, and test which changes are truly effective.
  3. Those changes proven successful by the agents are packaged into the next version of the model.
  4. The next version of the model thus becomes smarter, enabling it to help researchers solve code bugs that the previous version could not comprehend.

This cycle continues endlessly.

Each better AI helps humanity build the next stronger AI at an unprecedented speed. This is the true essence of "Recursive Self-Improvement" (RSI).

Once the rotational speed of this flywheel surpasses a certain critical point, the speed of AI development will completely detach from the linear cognition humans are accustomed to, presenting an almost vertical exponential explosion.

Rugged Frontier: The "Closed Loop" is Established, but Humans Remain the Bottleneck

Of course, announcing that AI is completely autonomous right now would be overly sensational.

OpenAI admits that current technology still has significant limitations. They refer to the current stage as the "Jagged Frontier."

What is jagged? It is extreme specialization.

From March to October this year, for complex development tasks estimated to take 2 to 4 hours, the "zero-intervention success rate" (i.e., human completely not intervening) of agents doubled from 23% to 46%. Even for more difficult long-term tasks requiring 4 to 8 hours, the success rate increased from 13% to 35%.

But if you break down the tasks, you will find that AI's capability boundaries are very uneven. The success rate for monitoring training operation issues is 79%. However, for writing underlying reasoning and service stack code, the success rate is only 25%.

Data shows that when tasks involve highly abstract "building from scratch," the success rate of AI plummets.

It excels in communication (69%) and analysis (58%), but still behaves like a reckless apprentice in development (43%).

In internal interviews, human researchers continue to express their frustrations with AI.

To fix a small bug by restructuring an entire module, AI tends to get stuck; it produces "AI garbage" PRs (pull requests) that seem fancy but do not work; it generates unusable data and even makes extremely absurd analytical assumptions.

A researcher involved in post-training evaluation helplessly summarized his new workflow:

"I used to try directly telling AI, 'Just make this task like that task,' but I got burned too many times because it always misunderstood, and I had to clean up after it.

Now the strategy has changed; I say: 'You do one first, and I'll review and modify it. Once I confirm it's correct, you can do the remaining ten thousand.'"

You see, this is the current reality. Humans remain the absolute core in the AI development closed loop; however, the role of humans has shifted from "creator" to "auditor."

Humans provide the initial judgment and error tolerance baseline, while AI is responsible for amplifying that baseline a thousandfold.

But this balance is extremely fragile. As one researcher left a thought-provoking remark in the report: "I truly believe that I currently have better taste in scientific ideas than AI. But I have no confidence in how long this situation can continue to hold."

Is the Endgame Set? OpenAI's Urgent Call and New Order

At the end of the report, OpenAI rarely issued a highly pressing industry call: Mandatory disclosure of RSI metrics.

They believe that all top AI laboratories should be required to track and publicly report the core metrics of "Recursive Self-Improvement." The public and regulatory agencies have the right to know how far the process of AI accelerating the development of future AI systems has progressed.

Why disclose? Because the speed of progress has exceeded traditional intuitive perception.

The report reveals that the OpenAI ECI (Internal Capability Index), aimed at measuring the trend of general AI capability progress, shows that since August 2025, the ECI frontier of OpenAI models has been skyrocketing at a rate of 16.4 points per year, surpassing the industry average of 14 points.

Even more frightening is that such astonishing progress has only utilized 2% of OpenAI's total R&D computing power to deploy internal agents.

Imagine what will happen when this proportion reaches 20%, 50%, or even 80%?

The traditional "human development cycle" is being broken. Previously, iterating a model required six months to collect data, three months for training, and two months for fine-tuning. Now, agents are using concurrent computing power to complete weeks of human work in just a few hours.

As the report warns: Historical trends may no longer predict the future.

In the field of mathematics, cutting-edge models have recently shown sudden and rapid breakthroughs, and a "mutation" in the field of AI could arrive at any moment.

Recent Fundraising

More
$10MNov 19
$1.7MNov 18
$110MNov 18

New Tokens

More
Dec 31
Dec 31
SSaturnSTRN
Dec 31

Latest Updates on 𝕏

More
Nov 18
Nov 18