AI models will quickly automate the majority of expert work, but they will struggle with the final, most complex 25%. For a long time, human expertise will be essential for this 'last mile,' making it the ultimate bottleneck and source of economic value.

Productivity models often wrongly assume time saved by AI is redeployed into other work. In reality, many employees use efficiency gains to finish early. This 'human slack' factor dampens macro-level productivity gains, except in highly driven fields like tech, where workers use it to work even more.

AI's primary impact will be augmenting and increasing productivity across entire organizations, not just automating lower-level tasks. The technology can handle a fraction of almost everyone's job, freeing up humans to focus on strategic, creative, and interpersonal work that models cannot perform.

AI's value is overestimated because experts view complex jobs as simple, solvable tasks. The real bottleneck is the unproductive effort required to build a custom training pipeline for every company-specific micro-task. Human workers are valuable precisely because they avoid this “schleppy training loop” by learning on the job, a capability current AI lacks.

In the near future, companies will leverage AI to demand exponentially higher productivity. Individuals unable to produce the output currently done by a team of ten will struggle to find or keep jobs. This is the real meaning of 'productivity gains'.

AI's economic impact is far more benign if it automates a small fraction of tasks across many professions rather than entire jobs. If AI handles 10% of everyone's workload, it results in a direct 10% productivity increase for the whole economy, making society wealthier with virtually no job displacement.

Even if AI accelerates parts of a workflow like coding, overall progress might stall due to Amdahl's Law. The system's speed is limited by its slowest component, meaning human-dependent tasks like strategic thinking could become the new rate-limiting step.

The AI productivity boom is confined to tech because developers have fewer adoption hurdles. Coding is a text-only medium with self-contained context in a codebase. In contrast, roles like marketing or law require complex data setup and workflow re-engineering, slowing down the productivity gains seen in macro-economic data.

The perceived time-saving benefits of using AI for lesson planning may be misleading. Similar to coders who must fix AI-generated mistakes, educators may spend so much time correcting flawed outputs that the net efficiency gain is zero or even negative, a factor often overlooked in a rush to adopt new tools.