One reason I'm sympathetic to this view is how far behind pre-AI automation is. Whenever I work with small-businesses and governments (and bigco's too but in different ways) I am amazed by how much make work there actually is. Somebody is doing data entry, but considers it valuable because some fraction of that work is identifying accuracy issues early, or translating between systems and domains. Then somebody is applying basically rote business logic to make a decision and route things to somebody else to do work. All of that could be automated. As an industry we've been automating that kind of thing for 50+ years, but still there are mountains of inefficient processes that could be automated, that are basically just shuffling paper. But the overhead cost of hiring someone to analyze your processes and adopt it to some kind of system, the costs of maintaining that system, to costs of getting it wrong when the requirement didn't account for various edge-cases that weren't in the requirements because of the tacit knowledge embedded in employees, often makes the project not worth it or a failure.
If LLM capabilities just help provide better glue (or even just political/hype motivation) to enable easier integration/automation of basic white-collar workflows there could be significant disruption. There is a lot of low-hanging fruit out there ... every spreadsheet is a new potential SaaS business.
Yes, people keep saying LLM tech is laughable because it can't solve for cold-fusion or some nonsense in a single prompt.
In reality, all it has to do is replace a crap ton of white collar "lgtm and pass it on" jobs that exist with minimal transformation or logic applied. This is a crap ton of people
If LLM capabilities just help provide better glue (or even just political/hype motivation) to enable easier integration/automation of basic white-collar workflows there could be significant disruption. There is a lot of low-hanging fruit out there ... every spreadsheet is a new potential SaaS business.