Anthropic: "Dreaming" AIs to Boost Their Efficiency

Anthropic and the "Dream" Innovation for AIs

Anthropic, a company specializing in artificial intelligence development, recently unveiled a major advancement: its AI agents have gained the ability to "dream." This innovative process is designed to enhance the working memory of AIs and minimize the errors they make. This announcement was made during the company's annual developer conference, where new tools were also presented.

A Vision for Autonomous Improvement

Jack Clark, co-founder of Anthropic, expressed his belief that these new tools will enable AIs to progress autonomously. The concept of "dreaming" introduced by Anthropic is an anthropomorphic technique aimed at making AI agents self-improving. This approach was revealed during the developer conference held on Wednesday. The "dream" is part of a broader initiative to revolutionize software engineering and other fields of knowledge through increasingly autonomous tools. This capability will be integrated into Anthropic's developing product, the Claude Managed Agents.

Memory Optimization and Error Reduction

The "dream" technique aims to refine the memory of systems by conducting evaluations between sessions. It analyzes past behaviors to identify recurring patterns, thereby helping agents adopt better working methods and reduce errors. Currently, this process is offered as a research preview, and interested developers must submit a request to access it.

Revenue Growth and Sector Expansion

Anthropic has observed a significant increase in its revenue in recent months, largely due to the adoption of its Claude Code service by software engineers. This service, along with other related offerings, assists developers in deploying agents for long-term coding projects. The startup is looking to extend these capabilities beyond software engineering, targeting sectors such as finance and law.

Towards More Accurate and Productive Agents

The ability of agents to remember and learn from their past experiences could enhance their accuracy and productivity over time, thereby increasing their value for paying clients. This new release also highlights an influential essay published by Jack Clark last Monday. In his Import AI newsletter, Clark estimated that there is a 60% chance that advanced AI models could autonomously train their successors by the end of 2028. He emphasized the current momentum in research and the launch of new products by leading laboratories. Although "dreaming" is not explicitly mentioned, it fits into the company's overall effort to transform its AI models into work agents capable of partially managing their tasks.

Transition to Public Beta

Additionally, on Wednesday, Anthropic transitioned two of its Managed Agents tools from research preview to public beta. One of these tools uses rubric-type results to guide the agents, while the other is capable of delegating tasks to multiple sub-agents.