There's little that's "inevitable" about AI

What it claims

The essay scores LLMs against its four criteria one by one. LLMs pass criterion 1: they are not markedly inferior to most human work, because "most human work… is not done at a good standard". They get a mixed result on criterion 2: they replace only the least valuable parts of scarce work, and the more important the work, the less useful they are. They clearly fail criterion 3 (overwhelming advantage) and criterion 4 (cannot be cheaply neutralised). The author's proposed neutralisers are social opprobrium, regulation and economics.

The forecast depends on criterion 4. If LLMs cannot be neutralised, use persists for contingent reasons and everyone carries a "filtering burden" against LLM output. If they can be, skilled non-users gain a "crushing competitive advantage" once belief in LLM productivity erodes. The author puts that erosion at "a few years, or even as long as a decade".

Evidence

The essay cites no data. Its arguments are:

• Speed-up. An illustrative figure of "20% and 50% more PowerBI dashboards" from an LLM-assisted workflow. It has no source or setting. The author adds that gains shrink the closer the work gets to a business's core bottlenecks.
• A falsifiability test. If the advantage were overwhelming, at least one tech company would be shipping 5–10× as many products as its rivals and forcing them out of the market. The author says this has not happened.
• A contrary claim. Companies going "all in" on AI are said to be struggling more than others. No examples or source are given.
• Observation of practitioners. Among practitioners the author considers serious, LLM use is described as different in pattern and lower in magnitude. This is anecdotal.
• The chemical-weapons analogy (from Devereaux). Lethality does not translate into utility for strong armies, for three reasons: the weapons deny ground to both sides, full protection may cost less than a rifle, and delivery is unreliable. In the Tokyo subway attack, five litres of sarin killed twelve people. The mapping to LLMs is asserted, not tested.

Limits

• This is an opinion essay. Every empirical claim is unsourced or anecdotal, and none of it can be used as a productivity benchmark.
• The author sells a competing product. The essay ends by promoting Arca, the author's early-access, paid platform for maintaining skills "to outclass the LLM users".
• The essay admits its criterion-4 verdict has little evidence behind it either way, because "social trends take a while to shake out".
• It treats LLMs as having real misuse value: LLM-backed cyberattacks and astroturfing bots, which the author compares to the terror use of chemical weapons. It does not analyse these further.