The Machine That Removed the Human From the Loop

The machine that removes the human from the loop

Jack Clark, Anthropic's co-founder, put a number on it last week: 60 percent, reported The Decoder. That's his estimate, as of May 2026, that no-human-involved AI research will exist by the end of 2028. An AI system training its own successor, without a person touching the loop.

The mechanism making that plausible is the part that should worry you.

Clark, who has spent years watching AI labs build systems that exceed their designers' expectations, is unusually direct about why the timeline matters. Alignment techniques compound errors under recursive self-improvement. A method that is 99.9 percent accurate sounds robust. It drops to roughly 95 percent after 50 generations and around 60 percent after 500. Unless an alignment technique is perfect, and none of them are, the supervisory layer degrades every time the system trains itself. As AI automates more of its own development, the human oversight meant to keep it safe becomes increasingly blind.

The benchmarks tell the story in a different key. SWE-Bench, which tests AI systems on real GitHub issues, went from 2 percent success with Claude 2 in late 2023 to 93.9 percent with Claude Mythos Preview in 2026, per Clark's analysis. CORE-Bench, which asks AI to reproduce research paper results, is effectively solved at 95.5 percent. METR's time horizons measure climbed from roughly 30 seconds of human work with GPT-3.5 to about 12 hours with current frontier models. Ajeya Cotra, a researcher at the METR project who publishes her own AI timelines, thinks 100 hours by the end of 2026 is plausible.

The five-minute feedback loop is where this gets concrete. Andrej Karpathy published autoresearch in March 2026 — two months ago, now, and already a artifact. The agent receives a real LLM training script. It edits the code. It runs a fixed five-minute experiment. It measures validation bits-per-byte. Keeps the change if the score improves. Discards it if it doesn't. Repeats. A human researcher doing the same task would need four to eight hours to hit a 4x speedup on CPU-only LLM training optimization. Karpathy's agent hits it in five minutes, over and over, without the coffee break.

Karpathy is still in the loop but only just. He sets the metric. He sets the budget. He defines the initial research program. He has moved up one level, from tuning the experiment to designing the loop that tunes itself. The program.md files — the Markdown documents that encode what to optimize and why — are the last hardcoded thing. Everything else, architecture, hyperparameters, the training loop itself, is delegated.

Anthropic's internal numbers are starker. On the same kind of task, optimizing a CPU-only small language model training implementation, the mean speedup went from 2.9x with Opus 4 in May 2025 to 52x in April 2026, per Clark's analysis. That's an 18-fold jump in eleven months. If that curve holds, the human researcher's four-to-eight hour task is down to minutes by next year.

The labs are not waiting for the philosophical debate to resolve. OpenAI has said publicly it wants an automated AI research intern by September 2026. Anthropic has published its own proof of concept for automated alignment researchers, in which AI agents beat Anthropic-designed baselines on a small-scale safety research problem. DeepMind, which Clark describes as the most circumspect of the three, still endorses automating alignment research when feasible.

Cotra has tracked her own timeline estimates against actual outcomes. Her January 2026 predictions on METR's time horizons measure projected 24 hours as her median forecast for the longest 50 percent task completion by end of 2026. Current frontier models are already at 12 hours. "The predictions I made in January were already starting to be met, in just two months," she noted in a March interview.

Clark's 60 percent by 2028 is a probability estimate from public data, not an Anthropic product announcement. He is releasing it because the implications are large and he does not think the public conversation is tracking them. The compiler example from the 1950s keeps surfacing in these conversations for a reason. When a compiler could compile itself, programmers recognized it as a threshold: the machine had incorporated the process of improving the machine into itself. The field did not stop. It reorganized around what humans were actually for.

Recursive self-learning is the same threshold for knowledge work. The question is not whether it happens. The evidence is already in the code. The question is what remains worth doing by hand.