On the road of artificial intelligence's evolution, Anthropic has once again dropped a major bomb. At a recent developer conference, Anthropic introduced a breakthrough feature called "Dreaming" (Dream) to its intelligent agent platform, Claude. The core concept of this technology is highly sci-fi: allowing AI to enter a state similar to human "sleep" during task intervals, simulating rapid eye movement (REM) cycles to organize and optimize its memory library.
For a long time, AI agents have faced the problem of "memory overload" when handling long-term, complex tasks. Information often accumulates in fragmented forms, and as interactions deepen, repetitive, outdated, or even contradictory data can lead to a decline in AI efficiency. The introduction of the "Dreaming" feature is like providing AI with an asynchronous processing "background brain." It can simultaneously retrieve up to 100 historical conversation records, merge duplicates, update outdated information, and extract macro-level patterns, transforming short-term, trivial memories into long-term, high-value knowledge bases.
Strict Risk Control: Generated Memories Do Not Directly Overwrite
To ensure the safety and controllability of this evolutionary process, Anthropic did not let AI "blindly" update. The optimized memory library generated by Dreaming does not directly overwrite the original data, and developers have veto power, allowing them to review the optimization results before deciding whether to apply them. Meanwhile, the platform also launched a real-time monitoring system and custom instruction features, enabling developers to observe each step of how AI organizes its memories, just like examining experimental specimens, and guide it to focus on specific topics.
In addition, to address the pain point of fluctuating AI output quality, the Outcomes automated scoring system was also launched. This system accurately scores AI-generated results using independent evaluators. If defects are found, it immediately instructs the AI to re-optimize until it meets preset standards.
Impressive Test Results: Task Success Rate Achieves a Leap Forward
The quality of a technology ultimately speaks for itself. Internal test results show that the Dreaming feature performs well in multiple application scenarios, with the success rate of document generation tasks generally increasing by 8% to 10%. In the legal tech field, partner Harvey's practice is particularly impressive: after applying this technology, its task completion rate soared by about six times.
In more complex collaborative scenarios, the multi-agent orchestration system demonstrated strong collaborative capabilities. By breaking down large tasks into multiple specialized subtasks processed in parallel and summarizing the results in a shared file system, the system's reliability saw a qualitative leap. In a simulation experiment of lunar mining drone landing, this system raised the safety score from failing 67% to a perfect 100% standard.
Powerful Computing: SpaceX Supercomputing Center Provides Support
Such intensive memory organization and simulation computing would not be possible without powerful computing support. Anthropic announced a deep strategic partnership with SpaceX, fully leasing the Colossus1 data center located in Texas. With the support of 220,000 high-performance GPUs, Claude not only easily handles the data flood generated by "dreaming," but also provides tangible benefits to regular users: Claude Code usage limits double, access restrictions for Pro/MAX versions are significantly relaxed, and API call speeds are notably optimized.
AI learning to "sleep" may be an important milestone in its journey toward a higher degree of intelligence.
