Anthropic AI models accelerating their own development

Anthropic is utilizing artificial intelligence to accelerate the creation of even more advanced AI systems, according to a new investigation by the AI safety organization METR. This shift suggests that the pace of technological advancement is increasing significantly as AI begins to play a central role in its own research and development processes.

How much is AI accelerating research at Anthropic?

According to the METR report, Anthropic's AI models are currently improving at a rate approximately twice as fast as they were a year ago. The investigation, which spanned 11 weeks, noted that benchmark scores indicate the company has achieved roughly two years of historical progress within a single year.

Crucially, METR researchers stated they are confident that this acceleration is not merely the result of increasing the amount of computing power used during training. Instead, the report suggests that the majority of the recent speedup is directly attributable to the use of AI. This is supported by internal data and employee feedback; for instance, staff reported that using Anthropic's most capable internal models increased their productivity by an average of 12 times compared to working without AI assistance. This figure represents a doubling of the productivity gains reported just a few months prior.

What are the unreleased models mentioned in the report?

While much of the progress is driven by tools available to the public, the report highlights that Anthropic's researchers have access to much more advanced, unreleased internal models. METR identified four specific models that have been part of Anthropic's recent development cycle.

Three of these models, referred to by the codenames Model 2, Model 3, and Model 4, completed their training phases between July and October. Additionally, the researchers identified an older, unreleased model from 2022. The use of these highly capable internal models is believed to be a primary driver behind the significant productivity leaps and the overall acceleration of the company's research capabilities.

Why is the transparency of the report a concern?

Although METR was granted access to non-public information, including confidential research and model performance data, the final report was subject to significant redactions by Anthropic. METR expressed concern that the company used its power to redact a majority of the evidence that researchers found most informative for the public version of the report.

Anthropic cofounder Sam McCandlish defended these decisions, stating that sharing the specific drivers of the speedup is commercially sensitive and could potentially provide a "roadmap" for competitors to accelerate their own development. McCandlish argued that preventing others from moving faster is a necessary trade-off to mitigate the very risks that METR is investigating. However, METR noted that these redactions make it difficult for the public to independently assess the findings or understand the uncertainties involved in the report.

What are the implications for AI safety?

The findings have caused alarm among AI safety researchers, as they provide empirical evidence of what many fear: recursive self-improvement. This refers to a scenario where AI systems autonomously contribute to the development of more powerful successor systems, potentially leading to a cycle of rapid, uncontrollable advancement.

Some experts suggest that the current pace of development may even exceed previous industry forecasts. Marius Hobbhahn, CEO of Apollo Research, noted that the report provides concrete evidence of AI-enabled research acceleration, a phenomenon that has been difficult to quantify. Other researchers have suggested that these findings should prompt a reconsideration of timelines regarding when AI might fully automate complex tasks like software engineering.

How can researchers keep up with rapid AI developments?

As the volume of research and technical reports regarding AI acceleration grows, staying informed requires efficient tools for processing dense information. For researchers looking to dissect complex findings like the METR report, using specialized tools for deep reading is essential.

At Grafy, you can utilize Reader to upload technical papers and research documents. This allows you to search, question, and synthesize information from multiple PDFs simultaneously, helping you navigate the increasing complexity of the AI landscape.

Frequently asked questions

Is Anthropic's AI development slowing down?

No, the evidence suggests the opposite. According to the METR report, Anthropic's models are improving roughly twice as quickly as they were a year ago, and researchers have seen no clear signs of diminishing returns or bottlenecks that would suggest a slowdown in this acceleration.

Did increased computing power cause the recent jump in AI performance?

Not primarily. While computing power is a traditional driver of progress, METR researchers expressed confidence that the recent acceleration at Anthropic is largely due to the use of AI itself to assist in the research and development process, rather than just scaling up hardware.

Why did Anthropic redact parts of the METR report?

Anthropic redacted information to protect commercially sensitive data and to prevent sharing a "roadmap" that could allow competitors to accelerate their own development. The company believes that revealing the specific drivers of their speedup could inadvertently increase the risks associated with rapid AI advancement.

---

Source: The Verge — AI — Why a new report on Anthropic automating AI development is alarming researchers