Source note
Predictions for the Future of AI
Summary
This essay argues that AI adoption will be limited by inference cost, weak consumer demand, and uneven productivity gains. Its strongest software claim is that coding assistants may make good engineers about 20-30% more productive, while non-programmers will build more custom software with new maintenance and security risks.
Problem
- AI inference may be priced below cost, so current demand may depend on hidden subsidies rather than durable customer value.
- Many consumer AI features solve low-value problems, so even large model gains may fail if users will not pay for them.
- Software teams need to know where LLMs raise productivity and where they create expensive, buggy, hard-to-maintain output.
Approach
- The author compares optimistic and pessimistic token-cost scenarios, then asks which AI use cases survive higher prices.
- The core mechanism is simple: if an AI workflow costs more than the value it creates, users stop using it or teams refactor it to use fewer tokens.
- For software work, the essay separates task-level speedups from whole-job replacement and focuses on substitution: more people can create software, but skilled engineers remain needed for judgment, architecture, security, and maintainability.
- The essay uses examples rather than experiments, including AI email replies, AI-generated commerce images, coding assistants, autonomous browser demos, and small-business custom apps.
Results
- No benchmark or controlled experimental results are provided; this is a prediction essay, not an empirical research paper.
- Inference pricing estimates range from an optimistic $0.60-$0.70 paid per $1 of compute to a pessimistic case below $0.10 paid per $1 after subsidies from hyperscaler capex, private credit, circular transactions, and tax treatment.
- The author expects production AI workflows to face cost controls even if an individual engineer has a roughly $4,000/month Claude bill.
- Consumer willingness to pay is framed as low: a restaurant-booking agent may fail at $200/month or $20/month and may only work near $2/month.
- For coding assistants, the author estimates a good engineer may gain about 20-30% productivity on average, while rejecting 10x or 100x claims as unsupported by research.
- One coding example says an LLM-generated 50,000-line project may include about 48,000 lines that duplicate existing open-source functionality with added bugs, making raw code volume a poor productivity metric.