Architecting
Human–AI Synergy

The stack has changed — and openness has to reach the bottom of it. The agentic transition is not a software-only event. Open weights are nearly commodity; the remaining frontier is the open compute substrate beneath them — compiler, kernel language (e.g. Triton), operator library, and cross-vendor enablement (the FlagOS-class porting that brought a flagship open-weight model to more than ten AI chipsets in two or three days). Open source must extend across the four layers of what we call the CLAW stack — Compute, LLMs, Agents, Workflow — or sovereignty over AI will be impossible regardless of which flag flies over the data center.

Trust is engineered, not declared. “We are no longer managing just hallucination. We are managing trust.” Agents already outnumber humans, by the panel's measure, “exponentially.” Verifiability has to move from logging to tamper-resistant, machine-to-machine evidence at every boundary. Sovereign-AI claims without cryptographic verification are incomplete: “we never talk about sovereign AI again without talking about verification in the same sentence.”

To make it concrete: Imagine a bank's AI agent denies a mortgage. Six months later a regulator audits the decision. Today, the bank shows logs the bank itself wrote — an honour system. In an engineered-trust model, the agent emits cryptographically signed traces at every step (its identity, the model version, the chip it ran on, the data it touched, the output it produced) to a tamper-evident ledger. The regulator can replay and verify the decision without trusting the bank.

Education has to lead, not follow. “Ninety-five percent of different experiments have failed, or only maybe from five to ten” in the recent enterprise AI wave — not because the models are weak, but because the cognitive design around them is missing. The classroom is the place where AI either widens the gap between top and median learners or finally closes it. “Whatever you use a tool for, you are going to get worse in that skill” — productive struggle has to be designed in.

Tokens are infrastructure. “In the AI coding bureau, code is cheap. Show me the data.” The unit of strategic resource is shifting from code to tokens. Open-source maintainers are under pressure from AI-generated contributions; corporate sponsorship alone is not the answer. “It doesn't matter how many tokens you give me. My community will still think that I'm the grinch that stole Christmas because it will never be enough to redistribute fairly to absolutely everyone.”

Human–AI synergy is a contract. “Synergy” was repeatedly reframed throughout the day as something to be engineered: explicit roles, verifiable identities, auditable behaviour, and friction by design where learning, judgement, and meaning are at stake.