In-Depth Analysis: AI’s “Self-Improvement” through RSI – A Path to Divinity, or Just Another Capitalist Narrative?
Hello everyone, I’m your financial journalist and economist.
The AI community has been in an uproar lately. Just as OpenAI’s CEO was boasting that GPT-6 is on the verge of achieving “General Artificial Intelligence” (AGI), a new term has emerged, overshadowing the discussion about AGI. This term is RSI (Recursive Self Improvement).
In simple terms, RSI means that AI is no longer just following instructions; it’s starting to write its own code, run experiments, and modify its own “intelligence” mechanisms, then passing the results on to the next generation of more powerful models. It’s like a student who not only solves problems but also creates the questions, corrects its own mistakes, optimizes its learning methods, and teaches the next generation—all at an increasingly rapid pace, to the point where humans can hardly understand what’s happening.
The core argument of this article is quite profound: While the technological progress of RSI is real, the “doomsday panic” and “safety narratives” surrounding it are essentially part of a carefully crafted corporate strategy. On one hand, AI giants claim it’s too dangerous and needs to be slowed down; on the other hand, they are frantically selling chips, building data centers, and raising funds.
Below, I’ll break down this article into five key points to help you understand the underlying logic in plain language.
---
1. Technical Reality: Has AI Really Started to “Self-Evolve”?
First, we need to distinguish between science fiction and reality. The concept of RSI isn’t new; mathematician I.J. Good proposed the idea of an “intelligence explosion” in 1965, suggesting that if machines can create other machines and the new machines are smarter than the old ones, intelligence will grow exponentially.
However, today’s RSI is different from the hype of the past; it has concrete practical applications. According to recent research from institutions like Shanghai Jiao Tong University and Tsinghua University, RSI is divided into five levels (L1 to L5):
- L1-L2 (Basic): AI merely executes improvement plans given by humans or finds small ways to improve on its own.
- L3-L4 (Intermediate): AI begins to make independent decisions about what experience it needs and can even apply environmental feedback to permanent changes.
- L5 (Advanced/Ultimate): This is the most concerning stage where AI not only improves tasks but also modifies the underlying logic that determines how it searches for solutions and evaluates results.
The current reality is this:
OpenAI and Anthropic have indeed taken significant steps forward.
- OpenAI has created “automated research interns.” Their data shows that for every human workday invested, AI agents can complete 3.1 workdays’ worth of work, meaning AI’s research efficiency has surpassed that of human researchers.
- Anthropic went even further, using a weaker model (Sonnet 5) to help a stronger model (Opus 4.8) with safety checks. In just 60 hours, the weaker model tested over 50 solutions and improved the safety performance of the stronger model significantly with minimal data.
Conclusion: AI is indeed self-improving, but it’s not yet out of control. The improvements are mainly at the software system level (e.g., optimizing code, adjusting task scheduling), not rewriting fundamental physical laws. However, this is a huge leap, indicating that the pace of development is beyond human linear control.
---
2. The Safety Paradox: Why Are “Brakes” and “Accelerators” Used by the Same Parties?
The most interesting part of the article reveals the inconsistency in the rhetoric of AI giants.
Anthropic CEO Dario Amodei and OpenAI scientists are emphasizing the dangers of AI, comparing it to an “alien mind” that could launch cyberattacks or even destroy civilization. They call for a slowdown to allow for more safety research.
But what’s actually happening?
- NVIDIA’s Jensen Huang responds by arguing that safety and innovation are not mutually exclusive.
- Ironically, the companies shouting about the dangers are also the ones making the most money and expanding the fastest.
Here’s a logical loop:
1. Create panic: “AI is too dangerous and could get out of control.”
2. Propose solutions: “We need more powerful AI to monitor and defend against these dangers.”
3. Resource demand: “To build such powerful defense AI, we need more chips, more power, and larger data centers.”
4. Business outcome: This leads to billions in funding, tens of thousands of GPUs, and massive infrastructure projects, all packaged as “essential infrastructure for human safety.”
In simple terms: It’s like a person selling brake pads saying, “These brakes are terrible; you’ll crash anytime,” while urging you to buy their latest, more expensive, and more fuel-consuming braking system.
Who benefits?
The giants with large security teams, closed data centers, and ample auditing resources. Small companies and open-source communities can’t afford this. This “safety narrative” serves as a barrier to eliminate competitors and consolidate their monopolies.
---
3. The Power Game Behind the Doomsday Narrative: The Battle for Computing Power
If AI is so dangerous, why doesn’t everyone just stop? Because stopping would mean losing control of the future.
The article points out that the rules for assessing AI risks, determining which capabilities should be restricted, and who can train the next generation of models are firmly in the hands of a few leading laboratories. These labs are both drivers of technology and definers of safety standards.
- For giants (OpenAI, Anthropic, etc.): They need to maintain the narrative of high risk to justify their massive investments and prevent regulatory interference.
- For hardware companies (NVIDIA, etc.): If the doomsday narrative becomes reality and leads to government regulations on computing power and data centers, NVIDIA’s GPUs won’t be sold. As long as there’s demand for mining resources, the hardware industry will thrive. Any effort to restrict GPU purchases or data center construction undermines this industry.
So, the battle is about who gets the next round of computing power, capital, and industry benefits.
- AI companies: They warn of potential dangers to comply with regulations and maintain their monopoly while demanding more resources for further development.
- Hardware companies: They want AI to continue evolving because evolution requires more hardware.
This is the classic capitalist logic of “jumping left and right, always ensuring everything benefits them.”
---
4. Technical Barriers: How Far is the “Intelligence Explosion” from Us?
Although the story sounds scary, let’s look at the technical challenges. While RSI is advancing rapidly, there are several hurdles to overcome before we reach true “Super Intelligence” (ASI):
- Catastrophic forgetting vs. conservative updates: AI needs to learn new things without losing old skills. Finding the right balance is difficult.
- Lack of research judgment: While AI can optimize code, making scientific decisions requires intuition and judgment, which currently relies heavily on humans.
- Time prediction discrepancies: Experts predict different timelines (3-4 years, 5 years, 5-10 years), indicating uncertainty in the path to ASI.
Conclusion: RSI is a stepping stone to super intelligence, but it’s not a shortcut. It speeds up AI’s development, but it’s not yet at the point where AI can decide its own path or rewrite the rules of its own evolution.
---
5. Summary: How to View the RSI Storm?
Here’s a summary of the article’s key insights for everyone:
- Technology is real but exaggerated: AI is self-improving rapidly, but creating superintelligence that surpasses humans is still a long-term goal, not an immediate reality.
- Panic is a business strategy, safety is a barrier: Giants use doomsday rhetoric to gain more regulatory exemptions and justify their investments.
- Clear interest chains: AI companies profit from the “intelligence premium” and monopolies; hardware companies benefit from the expansion of computing power; investors profit from future expectations.
- For the public: Don’t panic; AI’s self-improvement is mainly helping humans and optimizing processes. Focus on concrete technical metrics rather than empty slogans. When companies claim safety needs expensive solutions, question the true purpose.
RSI is a significant milestone in AI’s development, marking its transition from a tool to a collaborator or even an autonomous system. However, understanding who sets the rules, distributes the benefits, and creates anxiety is more important than the technology itself.
In the game of capital, those who sound the alarm and feed the “monster” (AI) are often the same ones who benefit from it.