The step from AGI to superintelligence could be short

Once AI can do almost everything people can, including AI research, the step to Superintelligence could take months or a few years rather than decades. The reason is not one dramatic trick. Digital minds have many ordinary advantages over human ones, and once they work on AI research they all compound at the same time.

Advantages human researchers don’t have

• Copies. Training a human expert takes decades; copying a trained model takes minutes. The day AGI exists, the number of AGI-level researchers is limited by chips, not by schools. The computers that trained it can run very many copies at once; Situational Awareness and AI 2027 both imagine hundreds of thousands or more.
• Speed. Copies can think faster than people and work around the clock, with no sleep, holidays or burnout.
• Parallel work. A thousand copies can try a thousand ideas at once and keep the ones that work.
• Shared learning. When one copy learns something, the update can be passed to every other copy. Human knowledge spreads through slow, lossy teaching, and retires with the expert.
• No coordination tax. Copies share training, background and goals. Much of what slows large human organizations, such as onboarding, misunderstanding, turnover and politics, shrinks.
• Only the best version. A human lab is limited by the people it can hire. An AI lab can run its best model only, and test many variants to find a better one.
• Working on its own foundations. An AGI doing AI research improves the very things that make it capable: algorithms, training data, the code that runs it, and the design of the chips it runs on. Each gain raises the next. See Recursive self-improvement.

Why they compound

Each advantage alone is a speed-up. Together they multiply: more researchers, each faster, better coordinated, sharing everything they learn, working on the thing that makes them smarter. Even if each factor is modest, the product is large. Progress that took the field a decade could take a year. That is the core of the Intelligence explosion argument, and the reason AI 2027 puts only months between a superhuman coder and superintelligence.

The comparison with humans makes the point. Human intelligence has barely changed in 10,000 years because we cannot copy, edit, speed up or merge our minds. AI can do all four.

The strongest objection

A feedback loop is not a speed estimate makes the best case against: a feedback loop says nothing about its speed. Experiments need compute, chips take years to build, power plants take longer, and removing one bottleneck exposes the next.

That shows the speed is uncertain, not that it is slow:

• Many bottlenecks are themselves research problems. Getting more out of the same compute, running small experiments that predict large ones, and writing faster code are exactly the tasks the new researchers would work on.
• Software alone has gone a long way. Better algorithms have made the same results several times cheaper to reach each year, on existing hardware. Automated researchers could push that rate up before any new chip factory is built.
• The hardware already exists. By the time AGI arrives, the world will have built enormous datacenters for it. Superintelligence may need better use of them more than more of them.

Other objections deserve answers too:

• Ideas get harder to find. True, and it is why human research slows down. But a jump of many times in the number and speed of researchers more than offsets a gradual rise in difficulty, at least for a while.
• Copies share blind spots. Copies of one model may all make the same mistakes. But variants can be trained differently, and experiments, not opinions, decide which ideas survive.
• The physical world is slow. Building robots, factories and power plants takes time. That limits how fast superintelligence changes the world, but much less how fast it is reached, since AI research mostly happens on computers.

What would change this

The case weakens if automating coding turns out to speed up AI research only a little, or if progress depends mainly on ever-larger experiments that no amount of cleverness can shrink. It strengthens if AI systems already make AI researchers clearly faster before they can replace them. That speed-up is worth watching as the clearest early signal. For how soon AGI itself might come, see AGI timelines.

Related: AI takeoff, Intelligence Explosion Evidence and Import.