The best arguments begin
with the hardest objections.

1. 01 Isn't this just reinventing the TPU / NPU? +

   A TPU is an accelerator part: however fast, the machine is still organized around the CPU — which launches it, feeds it, collects from it. The Noeton changes the organizing principle itself: memory becomes the resident center, the CPU retires to housekeeping, the OS swaps its kernel. Parts answer "how to compute faster"; the MNA answers "what the machine revolves around".

   The stack: the organizing principle in full →
2. 02 PIM has failed for thirty years. Why now? +

   PIM historically died of two things: no killer workload, immature process. Both changed — LLM decode is a textbook bandwidth-bound load (GEMV-heavy, static weights) sitting squarely in PIM's sweet spot; 3D stacking is commercial; and crucially, Samsung and SK hynix are jointly standardizing LPDDR6-PIM. Standardization is the exhibit-to-product divide the last wave never crossed.

   Roadmap essay 01, with six papers →
3. 03 MLA and sparse attention shrink the KV cache. Doesn't that undercut memory-centrism? +

   KV compression bends the capacity curve, not the position on the roofline: decode still rereads all weights per token, arithmetic intensity stays constant, still pinned to the bandwidth-bound slope. Meanwhile test-time-compute models shift token budgets massively toward decode — the bandwidth-bound share of inference is growing, not shrinking. This is one of the two refinements recorded in paper v2.

   Verify it yourself: the roofline lab →
4. 04 Network-mandatory is anti-user: a deliberately limited offline machine is unacceptable. +

   The heaviest objection, and we don't dodge it: network-mandatory design trades autonomy, privacy, and equity for capability density, and the paper spends a full section on those costs. Our position is to make the trade an explicit contract — degraded mode is a declared capability floor, not a failure; escalation decisions are visible and auditable. Apple's on-device + Private Cloud Compute deployment at consumer scale shows the market accepts honest tiering.

   Pull the offline switch yourself →
5. 05 What about general-purpose computing? Spreadsheets, compilers, databases still need to run. +

   The Noeton doesn't replace the general-purpose computer any more than the GPU replaced the CPU. The housekeeping core keeps a full conventional stack; under the Model-OS, legacy applications become a class of tools — invoked by intent, executed in sandboxes. What's displaced is not general-purpose computing but the machine form organized around the general-purpose core. Your spreadsheet will run fine — it just stops being the machine's reason to exist.

6. 06 Dataflow compilers will never keep up with model architecture churn. +

   Granted: the compiler is the entirety of L2's difficulty — every architectural turn (MoE, MLA, linear attention) demands a compiler chase. Our hedges: pin the abstraction at an operator-level IR rather than specific models; keep prefill in the mature GPU ecosystem and point the fabric only at decode, the slowest-changing inner loop; and partner with shipping reconfigurable silicon instead of taping out — reconfigurability is itself a hedge against drift.

   Roadmap essay 03: compiler first →
7. 07 How is this different from MemGPT, AIOS, and other "LLM as OS" papers? +

   They validated the same mappings (kernel/process/syscall/filesystem) at the software layer, and we cite them gratefully. The difference is depth: the Noeton sinks the mapping from "a runtime atop an existing OS" into the machine's native organization — in the five-layer reference design, the Model-OS sits on a dataflow fabric and memory substrate built for it. Software proves the semantics; the full machine answers performance and energy. The two corroborate rather than substitute.

   The Model-OS series: five mappings →
8. 08 An LLM as kernel? One hallucination and the whole machine goes mad. +

   Which is why the privilege model is a first-class citizen of the MNA, not a patch: graded side effects (reads run free / writes sandboxed / irreversible acts require a human), full audit trails, memory-snapshot rollback — a kernel that "thought wrong" can be rolled back like a transaction. The deterministic fabric makes behavior reproducible and auditable — the precondition for asking "why did the machine do that". Honestly: engineering alignment into a permission system is open research; we list it as one of the kernel's three hard problems rather than pretending it's solved.

9. 09 Why not wait for NVIDIA to finish the job? Isn't Rubin CPX already doing this? +

   Rubin CPX and Dynamo prove the incumbent is being pulled in our direction — but commercial gravity makes it incremental: the CUDA ecosystem, HBM margins, and rack-level economics are all tied to the current form, so the incumbent's optimal play is to keep patching the von Neumann machine, not to reorder it. Whole-machine reorganizations have always belonged to entrants: the smartphone wasn't built by Nokia, nor the cloud by Sun.

   Evidence wall: every move the incumbents make →
10. 10 A position paper, defining the next computer? +

    RISC began as a position polemic, the IP network as a memo, UNIX's philosophy preceded its market. A position paper's value lies not in proclamation but in falsifiability: every principle here ships with quantitative roofline/energy arguments, hands-on experiments, and a continuously updated evidence wall — and v2 publicly records where the evidence corrected us. Come falsify it; that is what publishing is for.

    Download paper v2 →

Still unconvinced? See how the industry votes →