Built for, and by, humans
Human values set the requirements and the technology is built to meet them — not the reverse. Safe, secure and private by construction, and judged by the real action taken and the value delivered to a person.
Start from the human, then build the machine. That ordering is the whole position. Safety, security, privacy, dignity, and the right to understand and contest a decision that affects you are not features to be weighed against performance — they are the requirements the architecture exists to satisfy. Much of the industry has been running this the other way around: build the most capable system available, ship it, then ask which values can still be fitted around what was already built. Anything that survives that process survives at the convenience of the architecture, which means it can be traded away the next time a deadline or a benchmark demands it.
Safe and secure has to mean something enforceable. The moment a system can act on its own behalf — call tools, move data, move money, actuate something in the physical world — its safety cannot rest on the good behaviour of a model, or on an instruction we hope it keeps following. The guarantees belong in the construction: isolation and attestation rooted in the silicon, confidential execution so data is protected while it is in use and not only at rest and in flight, a verifiable record of what ran, on whose data and under whose authority, and a way to withdraw that authority afterwards. This is why open roots of trust like OpenTitan and Caliptra matter more than they appear to — they make a safety claim into something a third party can check rather than something a vendor asserts. A safeguard written into a prompt is an intention. A safeguard anchored in hardware is a property you can audit.
The same inversion applies to how we measure. The benchmark culture of the last decade optimised for something measurable but not especially useful: how a model scores on a fixed test, in isolation, with no account of what it then does in the world. That was a reasonable proxy while models were tools a person drove directly. It stops being reasonable the moment systems act over long horizons. An agentic system’s value is the quality of the action it takes and the outcome a person actually gets — harder to measure than a leaderboard position, which is precisely why the industry has been slow to build the measurement. What we choose to measure is what the hardware underneath eventually gets optimised for, so getting the metric wrong is not a reporting problem. It is an architecture problem that compounds for a decade.
Designing for people also means designing for the person operating the system, not only the person consuming its output. Observability, explainability and the ability to attest what informed a decision are part of being built for humans — and they are what make the difference between a system a person can be accountable for and one they can only hope was right.