Dialogue

Frontier Thought · Greece’s next AI questions

FRONTIER THOUGHT ON AI · 18–20 November 2026

The discussion continues

Four questions, connected sources and specific requests for evidence in Greece.

Who gets a better life when AI gets better?

Amodei separates the speed of model improvement from the speed of adoption. Nadella goes further into what adoption entails: identities, tools, records and an operating environment for agents. Read together, these positions move the question beyond buying access to intelligence. An organisation has to decide which work changes, who retains responsibility and what a useful result looks like. Mitsotakis adds a test of public legitimacy: people need to experience personal benefits. The PHAROS discussion brings the same challenge to smaller Greek firms, which need practical support as well as computing resources. The shared question is therefore distribution. A faster service may free a worker to help difficult cases, increase the organisation’s output, or simply increase the workload expected of everyone. Those outcomes are different political and managerial choices. Our next inquiry should follow one service from the worker’s desk to the person who receives it. Measure waiting, mistakes and recovery alongside time saved. Then ask who received the gain, who absorbed extra work and which obstacle remains after the model is available.

The next question

In one Greek service, who actually receives the time that AI saves, and what would prove that their life or work improved?

Explore the panel and prepare the follow-up

What must humans learn when answers become abundant?

Khan’s proposal is powerful because it changes the unit of support: a student could receive attention and feedback that a busy teacher cannot provide continuously. Hassabis’s Athens conversation points toward a deeper objective, the ability to keep learning as technologies change. Bletsas gives that promise its hardest test. If students outsource the thinking, successful completion can conceal weaker understanding. Li’s emphasis on human dignity and a healthy research ecosystem also asks who gets to shape educational tools and judge their effects. These positions converge on a more demanding definition of learning. A useful tutor should help a student explain a mistake, form a question and solve something unfamiliar after the tutor has gone. Greece’s school pilot gives the conference a concrete setting for asking whether that happens. The inquiry should compare independent work over time, not count logins or accept polished homework as a result. It should also examine who receives teacher support, whether Greek-language explanations are sound and whether checking model mistakes creates more work than the tool removes.

The next question

What can students explain and solve independently a month after using an AI tutor, and which children benefit least?

Explore the panel and prepare the follow-up

Can society stay in control of systems it increasingly depends on?

Hassabis argues for cautious optimism because the potential benefits and risks are large while key uncertainties remain unresolved. Bengio asks whether useful intelligence must be given autonomous goals and proposes a safer direction for system design. Hinton focuses attention on competitive incentives and the difficulty of controlling systems that become more capable than their operators. In her latest Hinton interview, Hoover directly revisits Li’s earlier concern that extreme narratives can distort governance. That source connection gives the conference a concrete disagreement to examine: how should institutions allocate attention between harms already occurring and risks that could become much larger? Bletsas adds the Greek operational perspective. Attacks and disruption require priorities, rehearsed response and restoration. Our synthesis is that resilience needs both a theory of risk and a practical ability to intervene. Follow one consequential service through an outage or a harmful AI action. Ask who notices, who can stop it, what people can still do without it and how recovery is demonstrated. The answers turn a general demand for trust into responsibilities that can be tested.

The next question

If an AI system fails in a Greek hospital or municipality tomorrow, who can stop it and keep the essential service running?

Explore the panel and prepare the follow-up

Will Greece help shape AI, or live with choices made elsewhere?

Mensch’s recent conversation puts Europe’s productive capability and dependence on concentrated suppliers at the centre of sovereignty. Nadella’s argument for infrastructure that supports several model families introduces a practical complement: organisations need the ability to adapt when a supplier or architecture changes. Li adds a public dimension. Universities and public research help society judge the systems it uses rather than depend entirely on private assurances. Daskalakis brings the dependency question directly to Europe and Greece. Regulation needs to be accompanied by technological capability, competition and attention to labour. PHAROS and DAEDALUS provide a concrete place to examine whether those capabilities become accessible. Their existence alone does not settle who can use them or at what cost. The strategic choice is which layers Greece must be able to evaluate, operate or replace, and which can be shared through European partnerships. Our next inquiry should follow one Greek team with a real use case through access, evaluation and deployment. Ask what happens if its current provider withdraws, and identify the capability that would let the service continue.

The next question

What must Greece be able to build, evaluate or replace so that essential services remain a choice it can make for itself?

Explore the panel and prepare the follow-up