4 min readNew DelhiUpdated: Aug 30, 2026 11:19 AM IST
The ability of AI models to train other AI models has repeatedly been held up as a key indicator that artificial general intelligence (AGI) – a hypothetical level of intelligence at which automated systems outperform humans on most tasks – has been reached.
A new study by Anthropic has now found early evidence suggesting that AI models may be moving closer to that milestone.
In a paper titled ‘Automated Researchers Can Reliably Mitigate Alignment Failures’ published on Friday, August 28, Chen Yueh-Han, a researcher in Anthropic’s fellows programme, detailed how AI systems could reliably improve a model’s performance on a set of alignment benchmarks.
The study involved developing automated alignment researchers (AARs) with Claude Opus 4.8 that mitigate one alignment failure at a time. The AARs were then evaluated on 10 benchmarks for specific misaligned behaviours, and the study found that the automated systems were able to improve performance on every single one without degrading overall performance.
The findings of the study could be seen as a step toward recursive self-improvement, which many consider to be the next significant step in AI progress.
It comes at a time when Anthropic rival OpenAI has also reportedly made significant internal progress in reaching AGI, with CEO Sam Altman stating that it could happen by the end of this year, according to a report by Time. Mark Chen, OpenAI’s chief research officer, also reportedly estimated that it is “80 per cent of the way” there. For context, as per OpenAI’s charter, AGI is defined as “highly autonomous systems that outperform humans at most economically valuable work.”
Key findings of Anthropic’s study
As mentioned above, Anthropic’s experiment involved building Claude Opus 4.8-powered AARs. Each AAR was tasked with searching literature, proposing a training method, and training the target AI model using the method for about 30 minutes on a Nvidia H200 GPU, gradually increasing safety benchmarks such as MMLU over several iterations.
Story continues below this ad
The most effective training methods were preserved while discarding the ineffective ones. This allowed the automated systems to operate quicker and at a greater scale.
The study also compared the performance of AARs in training models compared to human AI researchers. It found that the best AAR method beat what experienced human AI researchers proposed, on average within six hours. Additionally, AARs could also come out to be the cheaper alternative as they “cost roughly $4 per hour in API inference against the $150 per hour we pay our human researchers.”
“Overall, these results provide early evidence that automated alignment post-training could become practical in the near term,” the paper read.
However, the paper also discussed a few key limitations based on its results. For instance, AARs may not make human AI researchers obsolete anytime soon as the former only works insofar as the benchmarks reflect the actual alignment goals. There is significant work to be done in establishing and maintaining those benchmarks.
Story continues below this ad
Expanding AI research literature that the automated researchers draw from will also require their human counterparts.
Leave a Reply