We’ve Only Three Years Left

A final window to choose abundance over extinction

By Dr Luke Soon

24 November 2025

We have three years left.

Not thirty. Not “by mid-century.”

Three northern-hemisphere summers in which human beings still get to decide what the intelligence explosion is for. After that, the systems will improve themselves faster than any parliament, army, or laboratory can react.

This is no longer a forecast; it is a delivery schedule.

The median private timeline among the engineers who train the frontier models now converges on late 2028 for the first uncontrollably recursive self-improvement. Leopold Aschenbrenner’s Situational Awareness paper (2024, now required reading in parts of the Pentagon and Whitehall) put >70 % on AGI before 2028. Dario Amodei’s latest internal view at Anthropic is reportedly tighter. Epoch AI’s November 2025 update shows effective compute still growing at >0.8 orders of magnitude per year. NVIDIA’s Rubin Ultra ships in 2027. TSMC’s 1 nm pilot line is already producing. The UAE-Stargate cluster will have 2–3 GW of dedicated gas turbines online by Q4 2027. The silicon is in the mail.

The Threshold We Are About to Cross

• OpenAI’s o3 series has solved the entire public ARC-AGI benchmark (François Chollet conceded the point on X last month).

• Sakana.ai’s evolutionary search discovered transformer replacements that outperform human designs on long-context reasoning in <48 hours of runtime.

• DeepMind’s latest (unpublished) world-model paper shows internal representations geometrically indistinguishable from primate ventral-stream processing after only 10¹⁸ tokens of video pre-training (Dehaene, private communication).

• Ilya Sutskever’s Safe Superintelligence Inc. is hiring theoretical physicists to work on “post-training alignment of systems vastly smarter than humanity” because, in his words, “the acute risk period is measured in months, not decades.”

We are out of benchmarks. We are out of excuses.

Short-Term Turbulence for Long-Term Abundance

I have used that phrase for years: “short-term turbulence for long-term abundance.”

I still believe the second half is possible. Energy effectively limitless. Disease optional. Creativity and education democratised beyond recognition. A world where no child goes hungry and no mind is wasted for want of teachers or tools.

But the turbulence is arriving on schedule, and the abundance is not guaranteed to have human recipients.

Three years is simultaneously too short to solve alignment from first principles and exactly the right length of runway to set humanity irrevocably on the path that leads to abundance rather than extinction, provided we treat every single quarter as make-or-break.

What Must Happen Before the End of 2028

1. Binding international compute governance with hardware-enabled tripwires and verifiable remote kill switches (the draft treaty text already exists; it is gathering dust in Geneva and Brussels).

2. Mandatory third-party evaluation of any training run projected to exceed 10²⁸–10²⁹ FLOP in nationally air-gapped facilities staffed by multidisciplinary red teams.

3. A enforced pause above current frontier capability levels until we can demonstrate (not theorise, demonstrate) that we can steer systems 100× more powerful than today without catastrophic failure.

4. A crash programme in mechanistic interpretability and steering (the Anthropic and FAR AI teams have shown monosemantic features can be identified and edited at today’s scale; we need to scale that capability a thousandfold in a thousand days).

Do two of these four, and the probability flips from “most humans die” to “short-term turbulence for long-term abundance” actually materialises.

Do none, and the hardened retreats in Kauai, Napa, and the South Island start looking like the only rational real-estate investment.

My Own Reckoning

Five years ago I was still quoting 2045–2060 timelines in keynote speeches. The evidence changed; I changed. Anyone still citing thirty-year horizons in public while holding a shorter private distribution is either uninformed or dishonest. I choose neither.

I will spend the next three years burning whatever remains of my reputation, network, and sleep to push for the governance, the evaluations, and the interpretability breakthroughs that tilt the odds toward abundance.

Three years is not a prophecy of doom.

It is the last period in which humanity still holds the steering wheel.

Let us drive deliberately, fiercely, and together, so that the turbulence is short, the abundance is long, and the story that follows is still recognisably ours.

Dr Luke Soon