Import AI flags automated AI research

This reads less like hype and more like a sober extrapolation from measurable capability trends. The real inflection, if it happens, is not "AI thinks like a genius" but "AI can run enough of the research loop to compound progress."

• –SWE-Bench saturation and longer METR horizons suggest models are already handling chunks of real engineering work.
• –Reproducing papers, tuning models, and optimizing kernels are exactly the kind of repeatable tasks AI can absorb first.
• –The near-term risk is a faster feedback loop in AI labs, not instant runaway self-improvement.
• –For AI builders, this implies more automated experimentation, more synthetic colleagues, and higher pressure on evaluation and alignment.