The second half of the chessboard
When I was young, my neighbour made me an offer: start at 1 and double it for every square on a chessboard. If I made it to the end of the board he’d give me £10.
1, 2, 4, 8, 16, 32, 64, 128, 256, 512... ten doubles in and I was on track. But doubling one sixty-four times doesn’t give you thousands, or millions, or even billions. It gets you over 18 quintillion. 1
I didn’t win the money, but I did learn that exponentials fool you. That they’re familiar at first, and then astonishing.
We are on the chessboard.
In 2019, GPT-2 could handle tasks that took a human a few seconds. The early ChatGPT models could stretch to about a minute. Today, frontier systems can complete tasks that take humans fifteen hours or more.
For over six years, the length of tasks AI can complete autonomously has been doubling roughly every seven months. More recently, closer to every four. 2
At the same time, AI systems are matching or surpassing human performance across more and more domains. And inference costs are collapsing, in some cases by orders of magnitude within a single year. 3
And yet, most organisations look the same. Same structures, same org charts, same processes; just a different set of tools. The doubling has been nothing but handy.
But that’s changing.
At an individual level we can already feel it. Not long ago using AI in serious work carried a stigma. It was easy to dismiss the outputs as “slop”. Then people started using these systems in secret, looking over their shoulder. Now, if you’re not using these systems at all, you’ll soon be irrelevant. If you resist, you’re only kneecapping yourself vs your peers.
Use these tools well and you feel amazement laced with horror. Expertise that took years to build has become a commodity anyone can access for £20 a month.
If we stay on this trend, within just a few years it’s plausible that the cost of intelligence will have fallen so far that maintaining traditional corporate structures will become a competitive liability.
While at first AI is just a tool that helps one employee to do more, eventually, the AI is simply better than the employee.
And although firms may initially hesitate out of loyalty or inertia, market pressures will force action. When intelligence costs nearly nothing, hiring someone for a job a machine does faster, cheaper, and better just doesn’t make sense. Entry-level hiring stops first. Then, as capabilities improve, the pressure moves upward. Hiring freezes become layoffs.
This is the "Intelligence Curse" 4; when intelligence becomes abundant, the economic value of individual human cognition collapses.
Some fields will hold out longer than others. Some roles will prove harder to automate. But we’re all on the same chessboard, just a few squares apart.
And if millions of people can’t sell their cognitive labour anymore, systems relying on mass employment start breaking down.
At best this is destabilising. We can speculate what worse outcomes could be: social mobility destroyed, wage collapse, tax income plummets, public services in ruin, public unrest, political chaos…
Maybe that feels hard to imagine right now. It always is, at 512 on the chessboard I didn’t imagine ending up at 18 quintillion. But a doubling every seven months means we could be just 1-2 years away from AI systems that can carry out work that takes humans days, or even weeks.
Look around in your firm, across all domains and seniority levels; how many doublings can it survive?
The compute coming online in 2026 is staggering in scale; millions of GPUs and TPUs; tens of gigawatts of power; billions of dollars of investment. The most powerful companies in the world are singularly focused on continuing the trend.
This is the most consequential technology in history yet, behind the frontier, boardrooms and ministers debate the risks of rolling out Microsoft CoPilot. They’re focusing on the first half of the board; not enough thought is being put into what might come next.
The second half of the chessboard will bring glory or ruin; it’ll eclipse all that came before it. Like other exponentials, it will feel normal and then astonishing.
For progress there is no cure. 5
References / related reading
I now know this of course is taking inspiration from “the wheat and chessboard problem”. And if curious, 2^64= 18,446,744,073,709,551,616.↩
The Intelligence Curse, Drago & Laine. The entire essay is a harrowing read; the pyramid replacement chapter particularly.↩
Can we survive technology?. It’s worth putting Von Neumann’s quote in full here: “For progress there is no cure. Any attempt to find automatically safe channels for the present explosive variety of progress must lead to frustration. The only safety possible is relative, and it lies in an intelligent exercise of day-to-day judgment.”↩