The artificial intelligence lab said on Thursday that Claude from Anthropic is assisting the company in creating the next, more sophisticated version of the model.
Claude is in charge of 26% of Anthropic’s model research and development, which the company claims allows it to finish the majority of a task “end-to-end from a high-level prompt” under human supervision. The model isn’t operating entirely on its own just yet.
However, 90% of the company’s R&D is carried out in “collaboration” with Claude, which Anthropic claimed allows the model to perform “large chunks of work under close human direction.”
The news was made at a time when some prominent AI leaders, like Anthropic CEO Dario Amodei, are advocating for a delay in the technology’s advancement due to safety concerns.
The corporation stated in a blog post that “we should do everything possible to minimize the gap between what frontier labs know and what the public knows” as leaders think about how quickly AI is developing. ” This entails measuring the advancement of AI more accurately, reporting on it in public, and allowing society to choose how to use this information.”
Models that speed up their own evolution may make it “more challenging for humans to understand or control these systems,” according to the business.
It claimed that disseminating these measurements could increase knowledge of how near top AI laboratories are to achieving recursive self-improvement, or a model’s capacity to create its successor on its own.
Additionally, Anthropic encouraged other AI developers to regularly disclose comparable metrics, promoting the usage of a public methodology so that the figures could be compared over time and possibly between labs.
Although Anthropic’s disclosure did not make it obvious how near the firm thinks it is to attaining recursive self-improvement, it is noteworthy how quickly Claude has increased its contributions to research and development.
In February, Claude “led” or performed a significant share of the task under human supervision. It reached that milestone in August after six months of spearheading a quarter of research and development.
The business also shared information on its agent oversight procedures, noting that as of August, about 30,000 agents were engaged in engineering and research.
According to the organization, oversight procedures are essential to determining how often monitoring systems flag agent misconduct.
Anthropic recently pledged to establish external third-party evaluators who will be integrated within the organization to oversee safety initiatives.
