Core Summary: AI Is Starting to “Repair Itself,” But the Wheel Is Still in Human Hands
In simple terms, Anthropic (the developer of Claude) has released a set of very impressive data, showing the world that AI is increasingly involved in the development of its own capabilities.
As of August this year, 26% of Anthropic’s AI research and development work was led by AI (with humans only setting goals and providing final oversight). In February of this year, that percentage was less than 1%. This means that in just half a year, AI has gone from being a mere assistant to a key contributor.
While AI can now independently fix bugs, write code, run tests, and even collaborate with other AI systems, none of these tasks have reached the “fully autonomous” (AL5) level. In other words, AI still cannot decide on its own what to research or work on without human supervision. This revelation not only showcases the real progress of the technology but also acts like a “bomb of transparency” within the industry, forcing other major players to reveal their own capabilities as well.
---
In-Depth Analysis: Understanding This “AI Self-Evolution” from Five Dimensions
1. From “Assistant” to “Leader”: A Qualitative Shift in Half a Year
If we compare AI research and development to a company, AI used to be like an intern—you had to watch over it while it looked up information or wrote code. Now, AI has become a senior project manager.
Anthropic uses a six-level automation framework (AL0-AL5). The current situation is as follows:
- AL4 (AI-led): 26% of the work has reached this level. What does that mean? Human engineers simply need to say, “Fix this data pipeline,” and the AI takes care of the rest—checking logs, identifying issues, writing code, testing, and even writing reports. Humans just need to review the results to ensure everything is correct.
- AL3 (AI collaboration): Over 90% of the work has reached this level. This indicates that AI is deeply integrated into the development process and is no longer just an optional tool but a core component of productivity.
Plain Language: It’s like driving a car before: AI was the navigator, telling you where to turn. Now, AI is the autonomous driver; you’re in the passenger seat with your hand on the wheel, but the car is driving itself. Although we’re not yet at the “you can sleep while the car drives” (AL5) stage, the amount of work humans need to do has significantly decreased. This shift from “assistant” to “leader” has skyrocketed from 1% to 26% in just half a year, far exceeding expectations.
2. “30,000 AI Employees” Working Nonstop in the Background
What’s most impressive about this revelation is not the percentage but the scale.
Anthropic has an Agent platform with approximately 30,000 AI agents running at any given time. These agents have unique IDs and use a shared messaging system to assign tasks and correct each other’s work. In August, these agents made over 1 billion decisions.
Plain Language: Imagine a server room at Anthropic filled with not only human engineers but also a large team of 30,000 “digital employees” working 24/7, calling each other to confirm tasks and checking code for errors.
This is no longer a situation where one person uses one AI assistant; instead, it’s a team of AIs serving the entire development process. This scale effect means that AI’s development efficiency is growing exponentially. What used to take months for a human team can now be completed in days or even hours thanks to the parallel processing power of the AI team.
3. Security Monitoring: 2,000 Out of 1 Billion Errors Caught
With such capabilities, could AI cause problems on its own? Anthropic provided some concrete security data:
- Online monitoring: Out of the 1 billion AI decisions in August, only about 0.002% were stopped by the system. That’s roughly 20,000 out of 1 billion.
- Offline monitoring: About 100,000 conversation records are marked each week, and after machine filtering of false positives, only around 50 of the highest-priority ones require human review.
Plain Language: This data reveals two things:
1. AI is very well-behaved and acts within established rules most of the time, with a very low error rate.
2. Monitoring is extensive: Despite the low error rate, the large number of transactions (1 billion) means the system needs to be constantly monitoring. The 50 records that require human review are like critically ill patients in a hospital—few in number, but each one is crucial.
This also indicates that the current security system focuses on “post-event interception” and review rather than “pre-event prevention.” AI may try to exceed boundaries during operation, and the system’s role is to stop it. It’s like cameras on a highway: they catch speeding cars, not rely on cars to avoid speeding.
4. Computing Power Allocation: 12% of AI Research and Development Resources for Security
A crucial but often overlooked aspect is the allocation of computing power.
In the week of the sample data, Anthropic invested about 6% of its AI research and development resources in security research. If only the AI-driven tasks are considered, this percentage rose to 12%.
Plain Language: Computing power is the “fuel” and “time” for AI. Anthropic is dedicating 12% of its AI development resources to studying how to prevent AI from causing damage, recognizing that as AI becomes more autonomous, security risks also increase exponentially.
This is like a bank that needs to allocate more security funds to prevent hacker attacks or internal fraud as automation transactions increase. This 12% of resources acts as the “brake” and “safety net” in the AI’s self-improvement process. Without it, the faster AI progresses, the greater the risk of accidents.
5. Industry Competition: Transparency as a New “Weapon”
Finally, let’s consider the broader industry context behind this revelation:
Anthropic made these sensitive data public for two main reasons:
1. To gain influence: To show the public and regulators that humans are still in control before AI’s self-evolution gets out of hand, thereby securing more policy space and development time.
2. To pressure competitors: It’s a form of “revelation competition.” By revealing its capabilities, Anthropic is challenging competitors like OpenAI and Google DeepMind to do the same. If they continue to hide their data, they may be questioned by the public and media about potential risks or their technological lag.
Plain Language: This is similar to financial report disclosures in the stock market. If a company releases detailed data first, other companies may be suspected of financial issues if they don’t follow suit. The AI industry has entered a phase of “transparency competition.” The company that reveals the most and in the most detail will gain a moral and media advantage.
This also sets a new benchmark for the industry. Comparing AI development will no longer just focus on model scores but on more tangible metrics such as the percentage of automation, the size of AI teams, and the effectiveness of security systems.
---
Summary and Outlook
Anthropic’s revelation marks a new phase in AI research and development, where AI is becoming more “semi-autonomous.”
Good News: AI can significantly improve development efficiency, freeing human engineers to focus on higher-level design and innovation.
Bad News/Challenges:
- Humans still make the final decisions about what AI should work on.
- AI lacks the ability to set strategic goals independently.
- It still lacks the skills to debug complex, long-term system issues.
- Security monitoring is still in its early stages, mainly focusing on catching errors rather than predicting risks.
For consumers, this means that AI products will evolve faster and become more powerful. For the industry, it’s a race for control, security, and transparency. With Anthropic taking the lead, the next moves will be both exciting and intense.