While players in the robotics market strive to ensure that “machines will do machines,” developers of artificial intelligence systems are also moving toward a similar goal of self-replication. For example, at Anthropic, up to 26% of activity is in the field of artificial intelligence development Already being implemented Take the leading role of artificial intelligence itself.
Image source: Human
Earlier this year, as stated in an impromptu report on the page Blog selectionno such approach was adopted. If we talk about collaboration with researchers, then 90% of the time artificial intelligence in this sense is used. The startup’s management shared this statistic as a formative example of generally applicable standards for evaluating progress in the development of artificial intelligence systems.
Anthropic also confirmed that it is willing to involve third-party auditors to analyze the security of its system development process: “Artificial intelligence systems are becoming more powerful and are already automating more and more of the tasks they create. As the world considers slowing the pace of advanced artificial intelligence systems, the public needs more information.”.
Anthropic is a proponent of the concept of “recursive self-improvement” in artificial intelligence systems, which means that professional software can continuously improve without extensive human intervention. Some market participants see a future for this method, but others have expressed concerns about its safety.
The company sought to create a way to assess the activity of artificial intelligence agents within its software infrastructure. As of August of this year, there were more than 30,000 AI agents working within the company at any given time, working in both engineering and research. According to the startup, it allocates 6 to 12 percent of its computing resources to security monitoring tasks. According to Anthropic management, disclosing this information will allow other companies to give the public an opportunity to decide what to do with this data.
By analyzing its own activities, Anthropic determined that there are no areas of activity for AI agents to develop and research with complete autonomy. The company called on market participants to create systems that monitor the activities of artificial intelligence agents, which would allow them to intervene in relevant processes when necessary. Another criterion for evaluating the security of AI should be the allocation of computing resources – as mentioned above, a portion of computing resources should be allocated for security monitoring. Finally, all this data needs to be correlated with actual test results of new AI models to understand their true capabilities. Such a set of standards would enable third-party auditors to develop a unified approach to analyzing the security of AI models created by market participants. Anthropic is committed to regularly sharing statistics like this to increase transparency of its activities.
If you find an error, select it with your mouse and press CTRL+ENTER.