This resonated a lot, especially the framing around specification being the work rather than execution.
The Python analogy feels exactly right to me. Each abstraction jump made individual contributors more powerful, but also raised the premium on people who could think clearly about systems, constraints, and failure modes. Syntax was never the real bottleneck. Precision of thought was. If anything, higher abstraction tends to expose that gap more brutally.
I also appreciate the spectrum from well-specified to ambiguous work. That maps cleanly to what I see in practice: LLMs shine when the success criteria are already implicit or externally enforced, and struggle when the task requires knowing what should be asked in the first place. In those cases, the model can only mirror plausibility, not judgement.
The tension you point out in the pessimist narrative is one I rarely see addressed clearly. You can’t simultaneously argue for broad labour displacement and permanently expensive human-intensive services without assuming a very selective, almost magical, form of automation. Either labour becomes cheaper across the board, or the disruption is more contained. Both can’t hold indefinitely.
One thing I keep coming back to is your point about the compiler being pedagogical. If ambiguity is silently resolved by a model rather than surfaced as an error, we may end up with more output but weaker mental models underneath. That doesn’t matter for all use cases, but it matters a lot for the ones that fail expensively.
Overall, this feels like a strong case that the most likely outcome is not collapse or stasis, but redistribution of effort: fewer people doing pure execution, more value placed on framing, judgement, and responsibility for consequences. Which is less dramatic, less viral, and probably closer to how technological change usually unfolds.
Curious how others see this playing out in roles where the spec has traditionally been learned implicitly, rather than written down.
This is a good perspective to have. I think the levels of abstraction concept is interesting. Makes me wonder if that logic still holds as there are still many C programmers out there. I think there is always going to be a space for hand writing code. I don't know for sure, but I find it unlikely that a bunch of Cobol banking software is going to be vibe-coded any time soon.
One thing I disagree on is this : "these sectors are expensive precisely because they’re labour-intensive."
Let's take housing for instance, it's not expensive because of this reason, but because there is not enough regulation and a single person can own up to 200+ houses or condos
Those sectors could usually come cheaper, and in some countries they are, if the right political choice are made.
Even if AI does not replace that much worker and only in a few sectors, if it does, we will need to make the right choices as societies
Housing is expensive because there is not enough regulation? Is that a joke? They can’t even rebuild after the fires in California because excessive regulation makes rebuilding prohibitively expensive and slow. Your argument is categorically false based on the historical record.
This resonated a lot, especially the framing around specification being the work rather than execution.
The Python analogy feels exactly right to me. Each abstraction jump made individual contributors more powerful, but also raised the premium on people who could think clearly about systems, constraints, and failure modes. Syntax was never the real bottleneck. Precision of thought was. If anything, higher abstraction tends to expose that gap more brutally.
I also appreciate the spectrum from well-specified to ambiguous work. That maps cleanly to what I see in practice: LLMs shine when the success criteria are already implicit or externally enforced, and struggle when the task requires knowing what should be asked in the first place. In those cases, the model can only mirror plausibility, not judgement.
The tension you point out in the pessimist narrative is one I rarely see addressed clearly. You can’t simultaneously argue for broad labour displacement and permanently expensive human-intensive services without assuming a very selective, almost magical, form of automation. Either labour becomes cheaper across the board, or the disruption is more contained. Both can’t hold indefinitely.
One thing I keep coming back to is your point about the compiler being pedagogical. If ambiguity is silently resolved by a model rather than surfaced as an error, we may end up with more output but weaker mental models underneath. That doesn’t matter for all use cases, but it matters a lot for the ones that fail expensively.
Overall, this feels like a strong case that the most likely outcome is not collapse or stasis, but redistribution of effort: fewer people doing pure execution, more value placed on framing, judgement, and responsibility for consequences. Which is less dramatic, less viral, and probably closer to how technological change usually unfolds.
Curious how others see this playing out in roles where the spec has traditionally been learned implicitly, rather than written down.
This is a good perspective to have. I think the levels of abstraction concept is interesting. Makes me wonder if that logic still holds as there are still many C programmers out there. I think there is always going to be a space for hand writing code. I don't know for sure, but I find it unlikely that a bunch of Cobol banking software is going to be vibe-coded any time soon.
I agree from deep in my gut. It's so boring I've never fed it a prompt.
AI as an acronym can find a lot of dismissive referents.
It started with Augmented Inference.
I've come to rest on Average Intelligence although I see this as overly generous.
I see AI as median intelligence. It will never reflect the genius end of the spectrum.
Very interesting article !
One thing I disagree on is this : "these sectors are expensive precisely because they’re labour-intensive."
Let's take housing for instance, it's not expensive because of this reason, but because there is not enough regulation and a single person can own up to 200+ houses or condos
Those sectors could usually come cheaper, and in some countries they are, if the right political choice are made.
Even if AI does not replace that much worker and only in a few sectors, if it does, we will need to make the right choices as societies
Cheers
Housing is expensive because there is not enough regulation? Is that a joke? They can’t even rebuild after the fires in California because excessive regulation makes rebuilding prohibitively expensive and slow. Your argument is categorically false based on the historical record.