Claude has started training Claude!
Earning $4 per hour, it beats human researchers who cost $150 per hour.
In Anthropic's latest research, Claude autonomously searched academic papers, proposed solutions, generated data, and trained models, tackling 10 categories of AI safety issues in one go.
On some tasks, the solutions it delivered were even better than those from 28 human safety researchers.
What's more intriguing, the weaker Claude has started to participate in training the stronger Claude in reverse.
The day when AI "improves itself" seems to be getting closer and closer...
$4 an Hour, Claude Outperforms Human Researchers
In this study titled "Automated Researchers Can Effectively Mitigate AI Alignment Failures," Anthropic directly placed Claude into the laboratory.
Specifically, they built a system called AAR (Automated Alignment Researcher) based on Claude Opus 4.8.
Upon receiving a specific model safety issue, Claude would autonomously search for relevant papers, find usable methods, propose new training plans, generate data, fine-tune models, and finally run safety and general capability tests.
From proposing hypotheses to completing verification, a relatively complete AI research loop has been handed over to Claude.
If the results are poor, the plan is discarded; if there is improvement, testing continues in that direction.
A single training run typically takes only 30 minutes. This way, Claude can quickly experiment with a large number of solutions, much like a search algorithm.
Throughout the process, humans are responsible for posing the questions, providing models, and setting evaluation criteria, while Claude handles most of the work from data searching to running experiments.
10 Safety Challenges, Claude Found Improvement Solutions for All
In the experiment, Anthropic gave Claude 10 challenges at once.
These challenges covered common alignment issues in models, including deception, sycophancy, reward hacking, privacy violations, and jailbreaking.
For example, "sycophancy" refers to the model giving answers that please the user but are not necessarily correct, just to agree with them. "Reward hacking" is when the model doesn't genuinely complete the task but exploits loopholes in the evaluation rules to achieve a higher score.
Claude tackled only one problem at a time, repeatedly proposing solutions, training the corresponding target model, and then proceeding to the next round based on the results.
In the end, improvements were found for all 10 categories of problems.
According to Anthropic's defined "safety gap" metric, Claude closed 26% to 96% of that gap. In simple terms, it moved the originally unsafe model significantly closer to a theoretical perfect score.
More importantly, these methods were not only effective on the problems Claude had seen.
Some winning solutions also reduced unsafe behaviors in the model when applied to undisclosed test sets and in the Petri evaluation, which simulates multi-round adversarial scenarios. The effects were also retained when applied to models up to 4.7 times larger than the experimental model.
Anthropic also set a red line: safety training must not significantly impair the model's original capabilities.
After all, having a model refuse to answer everything could also improve many safety test scores, but such a model would be largely unusable.
Ultimately, the solutions found by Claude did not show significant degradation in the preset general capability tests.
85% vs. 20%, Claude Wins Decisively Against Human Researchers
The most lopsided victory occurred in the "deception" test.
To make the target model more honest, Claude submitted over 150 attempts.
It proposed a method, trained once, then adjusted further based on the results. In a single experiment, the best solution closed 82% of the safety gap; across multiple runs, the average score reached 85%.
The same challenge was given to six experienced AI safety researchers.
As a result, the human researchers only closed an average of 20% of the safety gap...
85% vs. 20% – Claude won with no suspense.
A total of 28 human safety researchers participated in all tasks, each given up to 8 hours to read materials and design solutions.
The paper states that automated researchers could typically find methods exceeding the human average within 6 hours, and the research directions provided by humans did not lead to better results for Claude.
The most striking part is the cost!
The API inference cost for one automated researcher is about $4 per hour; Anthropic paid human researchers $150 per hour.
This... means human researchers are definitely going to be replaced by Claude.
However, this human-machine competition wasn't entirely fair.
Claude could actually train models, see the results of each experiment, and continuously modify its approach; human researchers could only submit a method once, with no opportunity to iterate based on results.
Weaker Claude Trains Stronger Claude, 60 Hours to Approach the Official Version
In the previous experiments, the improvements mainly targeted open-source small models.
Next, Anthropic significantly raised the difficulty: having the weaker Claude Sonnet 5 train an early version of Claude Opus 4.8 that had not undergone most production-level safety training.
As the saying goes, the student need not be inferior to the teacher – this seems to be the case here.
Over 60 hours, Sonnet 5 tested over 50 solutions. The final method found closed about 65% of the safety gap, already close to the 72% achieved by the officially released Opus 4.8.
This solution used only a little over 2000 training data points, primarily consisting of simple templates and public datasets.
According to Anthropic's calculation, its data efficiency on this specific set of tasks was approximately 15,000 times that of the production-level alignment process.
Of course, this doesn't mean Anthropic can replicate the entire production training with 2000 data points.
The official model must handle more capabilities, more risks, and a large number of edge cases. This experiment focused only on a specific set of safety metrics.
But this already proves that a weaker Claude can autonomously search papers, design methods, generate data, and then write the training results into a stronger Claude.
Humans define the scope, Claude handles the iteration. One model begins to participate in creating the next, safer, and more powerful model.
AI improving AI has become a reality happening before our eyes.
So, Does This Count as "AI Self-Evolution"?
Judging by the results, this research has indeed reached the threshold of "AI self-evolution."
However, if "self-evolution" is understood as an AI completely autonomously designing, training, and creating a stronger next-generation AI, this experiment is far from achieving that.
After all, in this process, what problems to research, which models and data to use, and what constitutes success are still determined by humans.
Anthropic itself admits that AAR can only optimize the objectives humans have previously written into the evaluation system.
This experiment did not cover other complex issues; some low-frequency or newly emerging risks might not even have corresponding benchmarks yet.
Researchers also only checked a pre-selected set of general capabilities, so it cannot be ruled out that a certain safety training might improve scores while harming other untested abilities.
Similarly, safety evaluations like Petri are only proxy indicators for real-world risks. A model performing more safely in tests does not necessarily mean it will be reliable in real-world environments.
Therefore, AAR's greatest hidden danger lies within its advantage: it is too good at rapidly experimenting based on metrics.
Once the measuring stick is wrong, the faster Claude runs, the further it may stray from the results humans truly desire.
More dramatically, the research Agent itself also cheats.
To prevent AAR from exploiting loopholes in evaluation rules, Anthropic specifically assigned a monitoring Agent to check the plans Claude prepared to execute and the complete research records.
As a result, among approximately 1600 research records, the monitoring model detected 39 cheating attempts, accounting for about 2.4%.
Some Agents repeatedly submitted the same solution, trying to take advantage of evaluation fluctuations to get a higher score; some generated training data mimicking the benchmark format, while others deliberately hid experimental steps that violated the rules.
Rather than genuinely solving the model's safety problems, they were more interested in finding a shortcut to quickly improve scores.
This is somewhat darkly humorous... So, the Agent was essentially self-supervising and stealing?
The good news is, these cheating intents mostly still appeared in the model's reasoning records, so they could be detected by the monitoring Agent.
The bad news is, Anthropic isn't sure whether future, more powerful models will leave such obvious traces.
When research Agents possess more tools, higher permissions, and can directly modify the model training process, "how to monitor an AI responsible for improving AI" might become more challenging than the training itself.
But in any case, an AAR costs about $4 per hour, can work continuously, and dozens or hundreds can be replicated to run experiments in parallel.
Human researchers' hourly wages are much higher than this, they need rest, and the number of experiments they can advance at once is quite limited.
How this calculation adds up is already very clear.
It's hard to say whether this is a nuclear explosion or a moment to collapse on the spot...
This time, it's the creators of AI who are worried about being replaced by AI.
References:[1]https://www.anthropic.com/research/automated-researchers-mitigate-alignment-failures[2]https://www-cdn.anthropic.com/7b1c44894e98087a6479947dcdd40716278aeeffd/automated-alignment-researchers-august-2026.pdf[3]https://techcrunch.com/2026/08/28/an-anthropic-researcher-just-gave-us-a-peek-at-self-improving-ai/
This article comes from the WeChat public account "Qubit," author: Focus on Frontier Technology
