A fluent response, a good demonstration and a benchmark score answer different questions. The business needs to know whether the system works on its own difficult cases, including when it should refuse.
The following are published positions brought together by AIStrategy. They are not answers given to this page’s question or a transcript of a shared panel. Our synthesis and next questions follow below.
THOUGHTLEADERS.GR / THE PERSPECTIVES
Start with what has actually been said.
NCSR Demokritos
NCSR Demokritos · ΕΚΕΦΕ Δημόκριτος
Vangelis Karkaletsis
Research leadership & PHAROS technical coordination
PUBLISHED POSITION / EDITORIAL SUMMARY
Useful public AI requires a path from research to services, with dependable infrastructure and evidence of impact.
Read the source context
Editorial reading of his account of Connect: Proof of Impact, held on 19 May 2026. This is a research-to-service argument, not a claim that every deployment is ready.
More output is not automatically better reasoning: efficiency and accuracy need to be measured together.
Read the source context
Editorial synthesis of the coauthored GFPO paper, August 2025. The results concern specific models and benchmarks; they do not predict a company’s savings.
Build AI safety into the whole system, with independent checks, monitoring and evidence, from the start.
Read the source context
With Qinghua Lu, AI Safety: Not Optional, Not Later, submitted 9 September 2026. The preprint proposes an assurance architecture across models, system controls and verification. It does not certify any deployment.
Follow a published position through its original source, the Greek question, a practical case and a next step. This collection reads the central ThoughtLeaders feed.
Connecting to the public collection
Checking source availability…
Published perspectives will appear here when the source is available.
Independent public statements, connected editorially. Source dates describe the original publication; they do not imply attendance at a common event.
Reliability is a property of a complete service, not a model name. Data, instructions, access, review and human recourse determine what happens when an answer is wrong.
What evidence would move this forward?
Build a small test set with routine, ambiguous and unsafe requests. Keep examples out of tuning, score independently and log both failures and successful escalation.