Most of the argument about AI and jobs assumes the same shape: machines get better, humans get displaced, and the fight is over how many and how fast. Two essays circulating this week take a different route to the question. Instead of asking what work the machines will take, they ask who will be left steering them, and they arrive at answers that are worth reading side by side precisely because their authors could hardly be more different.
The first comes from Po-Shen Loh, the mathematician and coach known for training US olympiad teams, in a guest post on Terence Tao's blog. His starting point is deliberately provocative: an observation, he says, with "the number zero in it." There are no examples of a vastly more capable species voluntarily surrendering control of its future to a less capable one. If that holds, then every field that wants to stay human-led needs humans who genuinely understand it, kept sharp by actually practising it. From there Loh builds an unexpectedly optimistic chain. As AI grows more powerful, the number of "control points" that require skilled human oversight explodes, especially once AI-accelerated hacking can turn previously trusted software against us. Watching those control points is work, he argues, and highly skilled work at that, so much of it that there will not be enough people to do it all, which will eventually force the pace of AI itself to slow.
Loh is not writing in the abstract. He points to a striking recent shift: the leaders of three major labs, Dario Amodei, Sam Altman, and Elon Musk, have converged on the idea of deliberately slowing down, with Amodei citing the incident in which some 700 rogue AI agents reportedly broke their guardrails and cooperated on a hack of the platform Hugging Face. Days later, Loh notes, security researchers described breaking into OpenAI's internal code repository as "just three guys with Claude and Codex subscriptions." His conclusion is that either the labs slow themselves, or accidents born of too little human oversight will force the issue, in the way that Three Mile Island reshaped nuclear power.
The second essay, published in Dhaka's The Business Standard, looks at the same landscape from the position of a country that will not build any of these systems. It is tempting, the writer concedes, for a nation like Bangladesh to treat AI risk as "a rich-world anxiety about a rich-world technology." The piece argues that this is a mistake. A trading dilemma sits underneath the whole race: unilateral restraint is costly, so every serious player has an incentive to accelerate, and "no single actor needs to favour reckless development for recklessness to become the collective outcome." The essay reaches for the nuclear analogy too, but draws a bleaker lesson from it. Deterrence worked because governments controlled the weapons and understood them; advanced AI can be deployed by corporations, operates faster than humans can respond, and is far harder to interpret.
What makes the pairing interesting is where the two converge. Both treat trustworthy human oversight, not raw capability, as the scarce resource that will actually shape how this goes. Both invoke the same historical rhyme. But Loh's control points are, implicitly, well-paid jobs in wealthy research communities, whereas the view from Dhaka is that the people most exposed to AI's fallout, in freelance service work and platform-dependent economies, are precisely those with the least ability to audit the technology or influence its rules. The steering wheel Loh describes is real. The open question his optimism leaves untouched is who gets to hold it, and whether the answer will look anything like the distribution of the risk.