Two-Year RSI Forecast Revives Debate as Anthropic and OpenAI Quantify AI-Led R&D

A two-year forecast for recursive self-improvement, pointing to roughly September 2028, and artificial superintelligence one year later has pushed a long-running AI-safety debate back into focus. The claim remains a prediction, not a confirmed capability, but recent disclosures from Anthropic and OpenAI show why the timeline is being taken seriously. Recursive self-improvement, or RSI, refers to AI systems improving the machinery that produces future AI systems; artificial superintelligence, or ASI, refers to systems that would exceed human capabilities across a broad range of domains. The newest hard data comes from Anthropic, whose measurement framework says that as of August 2026 Claude leads 26% of Anthropic's AI R&D work, participates at or above the AI collaborates level in more than 90%, and is not fully autonomous for any measured subset. Anthropic also says about 30,000 agents operate at any given time on its most-used internal platform, with 100% of those agents' actions passing through online monitors and 0.002% blocked in August. Anthropic's earlier RSI analysis argued that AI is already accelerating AI development: more than 80% of code merged into its codebase was authored by Claude as of May 2026, typical engineer code output rose about 8x, and Claude Mythos Preview achieved about a 52x speedup on a fixed model-training optimization task, compared with about 3x for Claude Opus 4 less than a year earlier and about 4x for a skilled human researcher over four to eight hours. OpenAI has published a parallel signal. In its September 6 update on research acceleration, the company said it had reached its automated research intern goal: a supervised system able to carry out well-defined research tasks that would take a skilled researcher a few days. OpenAI said it is working toward an automated AI researcher by March 2028, and that by mid-August its research organization used 3.1 agent-workdays of effort for every human workday. But OpenAI also emphasized that high-level planning remains a minimal fraction of agent output, that humans still set research priorities and decide whether to scale, pause or deploy systems, and that more than half of successful 4-8 hour tasks still involved at least one human intervention. Other examples show the same shift from theory to measured automation. Google DeepMind's AlphaEvolve optimized a Gemini training kernel by 23%, reducing overall Gemini training time by 1%, a narrower but verifiable case of AI improving part of the stack used to build AI. A July survey of 1,250 papers on self-improving AI systems draws a sharp line between bounded self-refinement, which is already industrial practice, and open-ended RSI, which it says remains constrained by verification, collapse dynamics and compute. The reception is split. Builders and developers are treating these disclosures as evidence that research pipelines are becoming agent-orchestrated, especially for coding, experiment execution, infrastructure debugging and benchmark optimization. Skeptics counter that lines of code can overstate productivity, that benchmark harnesses work best where evaluators are explicit, and that automating the doing of research is not the same as automating taste, agenda-setting or deployment decisions. Scientific American reported criticism that the current evidence looks more like powerful coding assistance than a decisive leap to autonomous successor-building, while the Associated Press described a wider industry split over whether RSI is near, dangerous, manageable or overstated. The next test is concrete: whether labs can move from supervised agents that execute research work to systems that safely choose research directions, improve their own successors, and remain monitorable as the loop accelerates. If the September 2028 RSI forecast proves too aggressive, the lab data still points to a meaningful change in AI development. If it proves close, the governance question becomes less about naming the milestone and more about who can verify, slow or steer the loop before ASI-level claims arrive.