PERSISTENCE
Context that compounds indefinitely without ever centralizing. The operator's memory is the asset that appreciates while everything else depreciates.
Dario Amodei imagined a compressed 21st century — a country of geniuses in a datacenter making 100 years of progress in 10. We borrow his method and point it at a different object: the compounding century — intelligence that, once owned, appreciates for the operator who owns it. His axis is speed. Ours is ownership.
Amodei's first move: refuse the hype-label, enumerate the capability. We do the same. We don't call it AGI. We call it powerful operator intelligence — an owned system that is compounding (every session leaves it sharper), sovereign (memory and orchestration on infrastructure you control), autonomous in loops (plans, acts, self-critiques, halts), and model-agnostic (the edge survives every frontier model swap). Every prediction below asks the marginal question — what does more of this buy the operator, and on what timescale — never the binary "can AI do X."
Context that compounds indefinitely without ever centralizing. The operator's memory is the asset that appreciates while everything else depreciates.
The operator's edge — IP, prompts, codebases, customer graph — must never be hostage to a platform that can reprice, deprecate, or lock them out.
Preservation. Agents, decisions, and work must survive the relentless churn of models and frameworks — or it isn't compounding, it's just a long session.
The virtual-researcher leap. Not a system that retrieves knowledge, but one that directs it — designs the inquiry, runs the synthesis, invents the method.
Operator-owned autonomous actors. The hardest pillar, deliberately last — it contains the question this whole roadmap is a bet on: does the operator edge survive AGI?
Five pillars × four horizons. Every grand claim forced into a checkable cell. The "NOW" column is real and running in PaperChase v0.1 today — this is what makes it a roadmap, not a manifesto.
| NOW · v0.1 | NEAR · 1–3 YR | MID · 3–10 YR | CENTURY · 10 YR+ | |
|---|---|---|---|---|
| 01 PERSISTENCE | vault.db + reflect | owned cross-device sync; semantic recall | portable cognitive asset, encrypted | lifetime memory as durable core |
| 02 SOVEREIGNTY | local-first, MIT core, Tier-3 IP | IP provenance + ownership proofs | 'good enough' open-weight tier | operator IP as asset class |
| 03 LINEAGE | intentional backups, decision records | full reconstructability; ancestry graphs | model-agnostic migration | 100-yr operator genome, inheritable |
| 04 KNOWLEDGE | PaperCodex + deep-research harness | engine proposes the questions | proprietary compounding synthesis | directing intelligence advances the field |
| 05 AUTONOMY | PLAN→ACT→CRITIQUE; risk laws | longer-horizon goals, tiered auth | self-improving trustworthy agents | open: does the edge survive AGI? |
Intelligence is necessary but not sufficient — it is gated. Following Amodei, every prediction names the factor it must beat. A prediction that can't name its limiting factor isn't a prediction. It's a wish. We delete wishes.
You are here. The 100-year roadmap — the thesis everything else gathers around.
Where the thesis becomes conversation, relationships, and shipped tools — Honolulu. Join →
The long-form thesis, argued in public on Substack. Subscribe ↗