Last week, Aravind Srinivas posted “Well said” on X in response to a thread arguing that computer science is gradually returning to the domain of physicists, mathematicians, and electrical engineers.
As AI automates most of what we currently call software engineering. The post got nearly a million views. Dario Amodei has said something similar, suggesting we are six to twelve months away from AI handling most software engineering end to end. Replit’s CEO put it more bluntly: the traditional software engineering job could “sort of disappear.”
The optimistic read of all this, and it is the one getting most of the attention, is that something good is happening. That the field is returning to its intellectual roots. That engineers will soon spend less time writing boilerplate and more time on systems thinking, mathematical reasoning, architecture, the hard stuff. That we are, in other words, being freed up to level up.
It is a genuinely appealing idea. And it deserves a harder look.
The vision being described, where routine work is automated and humans ascend to higher-order thinking, has a very specific assumption baked into it. It assumes that the people currently doing the routine work will have the time, the resources, the institutional support, and the economic runway to make that transition. That is a large assumption. Capital societies have never historically funded that kind of transition on the way down. They fund it on the way up, when the skills being developed are already generating returns for someone.
Anthropic’s own AI Exposure Index ranks programming as the profession most exposed to AI disruption, with roughly 75% of tasks automatable. Entry-level tech jobs are already shrinking in 2026, in the same cycle where these announcements are being made. The engineers most affected by this shift are not the ones with PhDs in mathematics from Berkeley. They are the ones who learned to code because it was a reliable path into the middle class, because bootcamps told them it was, because the industry spent a decade making that promise.
The question nobody in Srinivas’s comment section is asking is what exactly bridges the person who was writing boilerplate last year to the person doing systems-level reasoning next year. It is not a rhetorical question. It has a very material answer: time, money, and access to education. All three of which are distributed in the same uneven way they have always been.
The machines doing the work do not automatically create the conditions for humans to learn. It creates the conditions for the people who own the machines to capture more of the value the machines produce. Those are different things, and conflating them is how we end up with a very elegant theory of human flourishing that somehow never quite reaches the humans who needed it most.
None of this means the shift Srinivas is describing is wrong. Computer science returning to first principles is probably a genuinely good development for the field. The insight is real. The math and physics will matter more. The people who can think at that level will be valuable in ways that compound.
The uncomfortable follow-on question is: valuable to whom, on whose timeline, and what happens to everyone else while the transition sorts itself out?
The industry is very good at describing the destination. The hard part, the part that does not fit in a viral tweet, is who gets to make the journey.
