I want a computer whose building blocks can be composed for the individual and recomposed tomorrow if tomorrow requires something different. Machine intelligence gives us a reason to reconsider how much of personal computing must arrive as a fixed arrangement. Its speed and ability to interpret what a person is trying to do could make continual composition practical. That is the position of this paper, and the foundation for the three papers that follow.
Decomposition means making capabilities separately usable: finding information, transforming a document, keeping a working collection, scheduling work, displaying a result. These capabilities need accessible interfaces and data that can move between them. A person should be able to use an arrangement suited to the work at hand without becoming responsible for every connection underneath it. The difficult part is maintaining those connections as the work changes.
The proposed role of machine intelligence is to take on more of that interpretation and integration. Speed matters because an arrangement that takes weeks to prepare cannot follow an individual’s changing day. Understanding matters because producing connections quickly is of little use when they serve the wrong purpose. Here, understanding has an observable meaning: interpreting an intention, choosing suitable components, carrying a correction forward and recognizing when the result no longer fits. These abilities must be tested together, including the time spent checking and repairing what the machine produces.
Consider a person preparing for a journey. Today’s arrangement gathers the selected project documents, checks that they open offline and presents a small working index. Tomorrow, the same collection may support drafting a report. Search, storage, conversion and editing remain available building blocks; their composition changes with the undertaking. Daily recomposition is a capability to use when needed. Familiar controls and useful arrangements should persist while they still fit.
Composable systems and intelligent assistance have substantial precedents. Exokernel research separated resource protection from management to permit application-specific operating-system arrangements [1]. Horvitz’s mixed-initiative work examined how automated services and direct manipulation could work together [2]. This position brings those concerns to the continuing composition of an individual’s environment. It proposes reducing the human cost of maintaining that composition. Existing components still need compatible interfaces; intelligence cannot make an inaccessible capability composable by wishing at it.
The operating system supplies the protected foundation: processes, storage, devices and permissions. Recomposition can begin with tools and services above it, then extend into replaceable system components where their interfaces permit. Changes need to preserve the person’s data, respect existing authority and offer a way back. A machine that can rearrange the environment quickly also needs a dependable way to recognize an arrangement worth keeping.
Energy efficiency is a welcome possible consequence. An environment composed for the work actually being done could avoid duplicate indexing, unnecessary background services and repeated movement of the same data. Intelligence could prepare an arrangement that ordinary efficient code then executes many times. My expectation is that this can make our operating environments leaner, with less computation spent supporting capabilities the person is not using.
That expectation has a bill to settle. Interpretation, generation, checking and repair consume energy too. Luccioni and colleagues found substantial differences in inference energy across models and tasks, including higher costs for general-purpose generative systems than task-specific systems in their comparisons [3]. The proposed savings therefore depend on avoided work exceeding the cost of adaptation. Decomposition alone does not establish that result.
The three accompanying papers address conditions for making this proposal credible. When Does Context Earn Its Keep? asks whether selected context earns its cost through accepted outcomes. Finished Is Not Accepted asks what evidence establishes that the intended result happened. Personal Computing Across Changes and Interruptions develops how the undertaking can survive a changed instruction or a restarted process. Together, they turn this position into questions about useful adaptation, dependable completion and continuity.
An evaluation should compare a conventional setup, a competently configured static composition and an adaptive composition on the same undertakings as requirements change. It should measure accepted results, human setup and repair time, adaptation latency and total operational energy over a declared period. Energy accounting must include local and remote computation, failures and background activity; unavailable measurements remain unknown. Report both energy per accepted outcome and the period’s total, since cheaper work may encourage more work. The current papers do not yet establish these gains. The position is a research commitment: make the computer particular to its person, and test whether that particularity reduces the work required from both.
References
- Dawson R. Engler, M. Frans Kaashoek and James O’Toole Jr. 1995. Exokernel: An Operating System Architecture for Application-Level Resource Management. SOSP.
- Eric Horvitz. 1999. Principles of Mixed-Initiative User Interfaces. CHI.
- Alexandra Sasha Luccioni, Yacine Jernite and Emma Strubell. 2024. Power Hungry Processing: Watts Driving the Cost of AI Deployment?. FAccT.