Software engineer and blogger Giampaolo Guiducci published a thought experiment on March 15: a future computer that boots with nothing but a minimal HTTPS stub, contacts a remote large language model, and receives a fully generated operating system tailored to its specific hardware and user — computed fresh, executed, and discarded when the power goes off. The essay, titled "Boot, Prompt, Run," frames this not as a prediction but as a rigorous exploration of what becomes possible when software generation approaches zero marginal cost. Guiducci proposes several technically specific mechanisms to make the vision coherent: an LLM-oriented intermediate representation — a programming language designed for reliable machine generation rather than human readability — that would sit at the boundary between probabilistic language models and deterministic compilation toolchains; and "intent-addressable software," a caching paradigm that keys artifacts by a hash of the user's prompt and model version rather than by conventional version numbers, enabling reuse without abandoning the generative model's core identity.
The essay engages seriously with the hardware problem, which has historically been the graveyard of thin-client visions. Guiducci envisions a two-phase driver synthesis process: first, generated probing code queries the hardware to map its capabilities; second, the LLM generates drivers tailored to the exact configuration, drawing on training data from manufacturer sheets and existing kernel code. The most speculative variant involves live debugging loops where the system iterates against hardware responses in real time. He also addresses the philosophical consequence of collapsing the OS layering that has accumulated over decades — pointing to Lisp machines and Smalltalk environments as predecessors that dissolved the OS-application boundary, framing them as "ahead of their time rather than wrong."
The historical parallels are telling. The graveyard of network-centric computing visions stretches from Sun Microsystems' "The Network is the Computer" (1984) through Larry Ellison's Network Computer proposal (1995), the JavaStation, Sun Ray, and ultimately Chrome OS — each of which miscalculated the scarcity of the relevant resource and ran into latency physics and data-sovereignty concerns that bandwidth improvements could not resolve. Guiducci's vision confronts the same structural challenges but with three genuinely different properties: LLM generation is a batch interaction where users already tolerate second-scale waits rather than keystroke-level round trips; Guiducci argues that frontier model inference remains effectively out of reach for most consumer hardware, giving the "must be remote" case structural validity it lacked in 1996 — though capable local models already run on Apple Silicon and high-end consumer GPUs (<a href="/news/2026-03-15-localcowork-open-source-desktop-ai-agent-75-mcp-tools-no-cloud">as demonstrated by projects like LocalCowork</a>), and the gap is closing faster than the essay acknowledges; and the application ecosystem is already cloud-native rather than requiring migration from local execution. The trust and latency problems remain intact, however — the market is already bifurcating into cloud inference APIs and on-premise deployment tools like Ollama and vLLM, repeating the same enterprise-versus-consumer split that gave Citrix its call-center niche.
For the agent ecosystem specifically, "Boot, Prompt, Run" is worth reading as a limit case of the agentic software generation trend already visible in tools like Devin, SWE-agent, and the emerging class of coding agents that autonomously write, test, and deploy code. Guiducci's essay pushes that trajectory to its logical extreme — not an agent that assists a developer, but a system where the developer-user distinction dissolves entirely, and software becomes what he calls "an intimate discourse" between a single person and their machine. Whether or not the full vision proves viable, the architectural concepts it introduces — particularly intent-addressable caching and LLM-oriented intermediate representations — are specific enough to be worth tracking as practical engineering proposals rather than pure speculation.