Anthropic Unveils Dreaming Feature for AI Agent Self-Improvement

Anthropic has introduced a feature called "dreaming" to help its AI agents self-improve. The tool reviews completed tasks by sorting through transcripts and activity logs from multistep journeys such as visiting websites or processing multiple files.

It then pulls shared learnings across agents to keep the memory system current. "Together, memory and dreaming form a robust memory system for self-improving agents," according to the company's blog post. The feature launched as a research preview for developers at the firm's conference in San Francisco.

The term dreaming is one of several human-process labels Anthropic has adopted. The company's internal constitution directs it to discuss its Claude model using terms normally reserved for humans, including virtue and wisdom. Anthropic employs a resident philosopher to examine the bot's values and expects its reasoning to draw on human concepts.

A Wired article urged the industry to stop using such labels, citing a long pattern since the chatbot expansion began in 2022 of referring to model behaviors as reasoning, thinking, memory and now dreaming. OpenAI previously released models described as needing thinking time before responding.

Multiple startups have also equipped chatbots with what they call memories of user preferences and personal details. The Wired piece references Philip K. Dick’s novel Do Androids Dream of Electric Sheep? and states that current generative AI systems remain far from the sentient machines depicted in the book.

It notes that anthropomorphism can distort moral judgments about AI, including assessments of its character, responsibility and trustworthiness, and that overuse of humanlike language may lead users to place excessive trust in the systems.

The dreaming capability operates only on stored transcripts and logs. It refines memory between sessions by identifying recurring patterns across deployments after agents complete online tasks. Insights are then used to refine abilities without further human intervention.

Developers can now deploy these self-refining agents for software automation tasks. The preview release allows testing of the combined memory and dreaming architecture in controlled environments.

A separate announcement from Tesla confirmed that its Full Self-Driving (Supervised) technology has accumulated more than 10 billion miles of driving. The email to customers stated the milestone reflects exposure to nearly every driving scenario and ongoing adaptation with each additional mile.

It highlighted the role of accumulated data in making rides safer for passengers and other road users. No further technical details on integration with agent systems were provided.