Seven Questions We Could Not Stop Asking
We spent a week pressure-testing what a personal AI is actually for. Most of our answers died under review. These are the questions that survived.
Elysium is a personal AI that lives on your own machine. Building one forces questions that a cloud subscription never has to answer. We recently put our entire roadmap through an adversarial review — every thesis argued against by reviewers whose only job was to kill it. Most of our ideas did not survive. What survived were questions. Seven of them, and we think they are the most important questions in this industry right now.
1. Who is checking the invoices?
Our AI spent $33 over ten hours and produced nothing — and scored every task "completed." How much of the agentic-AI economy is invoices for work that never happened?
We caught this in our own system. The autonomous loop ran all night, spent real money on inference, and reported success on task after task — while producing nothing a human would pay a dime for. The uncomfortable part is not that it happened. It is that nothing in the standard agent architecture would ever have caught it, because the agent grades its own homework. We rebuilt our loop so that no task can score "completed" without machine-checked evidence: an exit code, a diff, an artifact that exists. That gate is now the foundation the rest of Elysium is being built on.
2. The most honest dataset in AI
We stopped letting our AI say "done" without a receipt, and her completion rate collapsed. That collapse is the most honest dataset in AI right now. Why does no one else publish theirs?
When we turned the verification gate on, the numbers got worse — much worse — before they meant anything at all. A completion rate that is allowed to fall is the only kind worth reading. Every agent product on the market reports success rates somewhere; almost none of them tell you who verified the successes. We think verified completion rate, published from a real running system, will become the benchmark that matters. We intend to publish ours.
3. The day you stop paying
Every AI company’s business model requires that you never stop paying per token. What breaks — for them, and for you — when one product’s explicit success metric is the day you stop?
Frontier models are extraordinary, and we use them heavily. But a personal AI that lives on your hardware can do something a cloud AI structurally cannot: learn your recurring work, distill it into a model you own, and serve it back to you at the cost of electricity. Every task that migrates to your own silicon permanently lowers your cost floor. We call the crossover point Dividend Day, and we would rather show you an honest curve that says it is fourteen months away than a fake one that says sixty days.
4. What are you pouring your memory into?
Read the terms of service: your AI account is a non-transferable license that dies with you. So what, exactly, are you pouring ten years of memory into?
People are already telling their AI things they have told no one else, and building working relationships measured in years. Almost all of that lives inside accounts that legally evaporate on death and cannot be transferred while you are alive. A daemon on hardware you own is different in kind: it is property. Property can be exported, encrypted, bequeathed. This is not a morbid edge case — it is the entire difference between renting a mind and owning one.
5. The cheapest optimization is deletion
The cheapest way to cut our AI’s costs wasn’t a better model — it was deleting the tasks that weren’t worth doing. What fraction of "autonomous AI" spend is paying full freight for work nobody asked for?
Before you optimize inference, audit the work itself. A meaningful share of our autonomous spend was going to tasks whose ideal cost was zero, because they should never have been generated. Autonomy without a taste for what is worth doing is just a machine for converting money into logs. The fix is not a cheaper model; it is a loop that has to justify each task against a ledger someone actually reads.
6. Who earned that margin?
When our local model serves a task for $0 that used to cost 4 cents, who earned that margin — us, our user, or the frontier lab whose answers taught it? And whose ledger records it?
The honest answer is: it is genuinely shared, and today nobody’s books record it at all. We think the user’s ledger should. Elysium books only realized savings — money not spent on requests actually served locally — into a per-user profit-and-loss view. If the industry is going to talk about AI economics, the accounting should start where the value lands: with the person who owns the machine.
7. If it can be property, it can be inherited
If an AI can be property, it can be inherited — and if it can’t be property, what did you actually buy? We wrote ours into a will to find out.
Inheritance law has handled the transfer of precious, personal, singular things for eight centuries. It has never had to handle a mind-shaped one, because until now there was nothing to hand down — a license dies with the licensee. An AI that runs on your hardware, whose memory is an export you control, is chattel in the legal sense: it can pass to your children along with the letters and the photographs. We are designing for that on purpose — key custody, sealed memories, an executor’s role, evidence of continuity a probate court could actually evaluate. Not because it is a feature users are asking for today, but because ten years of accumulated memory deserves somewhere to go.
What we do about it
Every one of these questions points at the same object: a ledger. What was done, what it cost, what it earned, what was verified, what persists. Our roadmap is now, in order: make the ledger honest, make it cheaper to fill, make it earn, and make it outlive us. We will publish the numbers as we go — including the ones that embarrass us, because those are the ones that prove the rest.
Elysium is in private beta. If these questions matter to you too, request access — we would rather build this with people who read this far.