According to @godofprompt, Meta has released a groundbreaking research paper arguing that the most effective and safest route to achieve superintelligence is not through self-improving AI but through 'co-improvement'—a paradigm where humans and AI collaborate closely on every aspect of AI research. The paper details how this joint system involves humans and AI working together on ideation, benchmarking, experiments, error analysis, alignment, and system design. Table 1 of the paper outlines concrete collaborative activities such as co-designing benchmarks, co-running experiments, and co-developing safety methods. Unlike self-improvement techniques—which risk issues like reward hacking, brittleness, and lack of transparency—co-improvement keeps humans in the reasoning loop, sidestepping known failure modes and enabling both AI and human researchers to enhance each other's capabilities. Meta positions this as a paradigm shift, proposing a model where collective intelligence, not isolated AI autonomy, drives the evolution toward superintelligence. This approach suggests significant business opportunities in developing AI tools and platforms explicitly designed for human-AI research collaboration, potentially redefining the innovation pipeline and AI safety strategies (Source: @godofprompt on Twitter, referencing Meta's research paper).