"In the end, one smartest AI would do it all, wouldn't it?" Reading about medical AI and legal AI, some people wonder which to choose. This chapter answers that question. Half of it is fact that can be checked; the other half is the forecast of the author (and of the AI that helped write it). The headings say which is which.
(First published as the extra installment of the series "Fable 5 Is Back," July 2026. Rewritten as structural analysis in October 2026, with the product specifics removed.)
Fact — behind the performance race, diversification is advancing
AI is not converging on "one strongest model." Diversification can be checked on three axes.
- Price and capability tiers — every vendor ships a range from fast-and-cheap to top-tier, and offers them in parallel in the same period. Running everything on the top tier makes economic sense neither for the vendor nor for the user. Usage fees scale with volume, so putting the top tier on short everyday requests changes nothing you can feel and only stacks up the difference
- Product forms — the same model grows into separate products for separate settings: a chat screen, an agent driven from the terminal, a screen-design tool. The product layer splits faster than the model layer
- Regulation and availability — in June 2026 US export controls halted the top-tier model, and it returned in July under conditions (blog 028). The episode showed that a state's regulation had become a variable deciding who gets which AI
Structure — the counter-force bearing on specialists
Alongside diversification, a force works the other way. A specialist model competes under structurally unfavorable terms: its rival is not this year's specialist but next year's generalist. Each time the general model steps up, territory that was the specialist's alone is absorbed into the generalist's range.
The overtaking has been reported. On medical examination problems, an untuned general model scored higher than a model tuned for medicine (Nori et al., Microsoft, 2023). In financial text analysis, a model trained from scratch on financial text alone lost on several tasks to a general model simply asked to answer (Li et al., 2023). Both are 2023 reports, and the pattern has since become the industry's standard story.
Reading a stack of contracts for contradictions, fixing code that spans many files — work that once belonged to legal AI and coding-assistant AI is now offered as a feature of the general model. To hold its edge, a specialist must deepen its expertise faster than the generalist advances. Not impossible, but a race up a slope.
Forecast — which specialists remain, and which do not
From here it is forecast. Nothing is asserted.
What the author sees as unlikely to last is "vertical specialization that is complete within language." Work whose input and output are both natural language — summarizing medical documents, checking legal ones — overlaps head-on with the generalist's strength. Reading large bodies of text to find contradictions is already standard equipment on general models.
What seems likely to remain comes in two kinds. One is "specialization outside language" — predicting the three-dimensional structure of proteins, or global weather, built on mathematics other than the language model's. There, dedicated models are already the standard, and there is no path by which a language model replaces them. The other is "uses that must be small, fast, and cheap" — processing that has to run instantly on a device, or fraud detection that watches masses of transactions at low cost. There, lightness and speed are worth more than top-tier intelligence.
Then does no value remain for language-side vertical AI companies? If any does, it sits somewhere other than the model: access to private data that never leaves the building, and the depth of embedding in the flow of work. Models can be swapped; the trust relationship over data and the seat of a tool the floor has grown used to cannot be swapped easily.
Conditions under which this forecast fails
A forecast should carry its own failure modes. Three.
First, when regulation demands an independent model. If medicine or finance makes "an auditable dedicated model" a legal requirement, language-side verticals gain a place that the rules protect. Second, domains where private data keeps deciding performance. As long as the data never leaves, the generalist cannot learn it, and the specialist can keep its edge there. Third, a bias to disclose: a general model (an AI) took part in organizing this forecast, so the conclusion "the generalist wins" includes the generalist's self-report. Discount accordingly.
Implication — pledge loyalty to no single AI
Brought down to practice, the implications are three.
- Choose the tool per purpose — extend the habit of picking models to the level of products and companies. 2-02's decision to subscribe to two vendors' AIs and have them check each other (2-02) is this implication implemented
- A sole proprietor or small business needs no model of its own — the value sits on the side of your data and your flow of work, and only you hold those. What you keep in-house is, as 2-16 sets out, an open-weight model and RAG over your own data
- Treat the model as a replaceable part — never build work in a shape only one model can serve. "Freeze it into code" in 2-17 is also a mechanism that makes swapping cheap
The third can be checked now. Have the AI go through your procedure manuals, list every place AI is used, and for each one lay out what would be affected by switching to another model and the steps of the switch. Keep the tools in a state where they can be chosen again. On a diversifying map, that is the safest way to walk.
Do not fix "which AI to ask." Choose the tool per job, and build so it can be chosen again. With that in place, whatever model comes next, the criteria carry over as they are.