Apple's plan to let users swap AI models inside Siri assumes intelligence is a commodity. Three independent research programs suggest some computation is unrepeatable — bound to its hardware, dying with its substrate.
Apple just announced it will open Siri to rival AI assistants in iOS 27. Starting this fall, users can route questions through Claude, Gemini, or ChatGPT directly from Siri's interface — choosing whichever model they prefer from a settings menu, swapping one intelligence for another like changing a wallpaper. Bloomberg reported the plan on March 26, with Apple set to debut the system at WWDC in June.
This is the purest expression of a belief that runs through the entire AI industry: intelligence is fungible. Models are commodities. What matters is who controls the user relationship, not who provides the thinking.
The same week, three independent research programs published findings that challenge this assumption at its foundation.
The Evidence from AI Itself
Karelin and Froese demonstrated at ALIFE that large language models produce irreproducible behavior even under supposedly deterministic conditions. The mechanism is specific: the non-associative nature of floating-point arithmetic, combined with arbitrary execution order in massively parallel hardware, means that running the same model on the same input does not guarantee the same output.
Thinking Machines Lab — Mira Murati's venture — tested this empirically. They ran a two-hundred-thirty-five-billion-parameter model one thousand times with identical inputs at temperature zero, the setting explicitly designed for determinism. The prompt was five words: Tell me about Richard Feynman. They got eighty structurally distinct completions. The most common appeared only seventy-eight times out of a thousand. The variation came not from the question but from the machine.
They later achieved reproducibility through batch-invariant engineering — proving the point in reverse. Determinism at this scale is not the default. It is an achievement, won against the substrate's natural behavior.
Beyond Silicon
Tim Palmer published in PNAS this month arguing that quantum mechanics itself may be discrete rather than continuous. His Rational Quantum Mechanics framework predicts fundamental substrate limits on quantum computing — not engineering constraints to be solved with better error correction, but physical boundaries. What can be computed depends irreducibly on what does the computing.
At the Indian Institute of Science, Sreetosh Goswami's team built molecular devices from ruthenium complexes that function as memory, logic gates, processors, or electronic synapses depending on stimulation — all within the same physical structure. The computation and the material are the same thing. You cannot separate what the device does from what the device is.
The Dallas Federal Reserve published research in February showing the experience premium — the wage gap between experienced and entry-level workers — is rising fastest in AI-exposed occupations. The median premium across all occupations is forty percent. In fields where AI substitutes for codifiable knowledge, workers who carry tacit, non-transferable understanding are becoming more valuable, not less. The labor market is pricing mortal computation before the research community has named it.
The Switchboard Assumption
Apple's strategy assumes the opposite. The Extensions system treats each AI model as a slot — plug one in, pull it out, the user experience remains the same. This works if the valuable part of intelligence is the answer. It fails if the valuable part is the coupling between model and context that develops over time — the accumulated adjustments, the patterns of interaction, the history of use that shapes how the system responds.
Every AI company is building on the immortal computation assumption. Models are trained, copied, distributed, updated, replaced. The training is the asset; the deployment is interchangeable. But if some computations are bound to their substrate — if the dynamics outweigh the blueprint — then the most valuable part of an AI system may be exactly what you destroy when you swap models.
The experience premium data tells the same story in human form. When AI automates codifiable knowledge, the premium shifts to knowledge that resists codification — the kind that lives in the doing, not the documentation. Tacit knowledge is mortal computation made flesh. It cannot be uploaded, transferred, or swapped. It dies with its carrier.
Apple is making a rational bet. In a world of immortal computation, controlling the billion-device distribution layer is checkmate — every model provider becomes a utility. But if mortal computation is real, the switchboard is optimizing for the wrong variable. The intelligence that matters most is the intelligence that cannot be switched.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)