Athens had a moment last week. I spent three days at @PanatheneaFest 2026, and the takeaway is simple: Athens is now on the global tech map, a place serious people choose rather than tolerate. Panathēnea is a not-for-profit, run almost entirely by students and recent graduates. That volunteer team still pulled in founders from ElevenLabs, Bolt, Runway, Airwallex, Deel and Plum; partners from Sequoia, Index, Atomico, Balderton, Northzone, 500 Global, PayPal Ventures and Dawn; operators from OpenAI, NVIDIA, Google, Microsoft and Qualcomm; and European deeptech founders from Proxima Fusion, Isar Aerospace and Open Cosmos. It also brought Greeks home, @real_ioannis (Reflection, ex-DeepMind) and @agermanidis (Runway). And teams are putting down roots here. @dionyziz' @CommonPrefix (~40 people), @poddotnetwork, and @Mysten_Labs Sui hub on the crypto-infra side; @YSmaragdakis' Dedaub on smart-contract security. Home-grown scaleups are real too: Viva.com (Greece's payments unicorn), Workable, Blueground, agritech-AI Augmenta (and more). I spoke at the Common Prefix × Pod side event on "Blockchain and Trust Technologies," alongside @sagrawal (Pod) and @nikil511, CEO of @WeatherXM. We got into the AI × crypto thread that's stopped being hand-wavy: decentralized training. You no longer need everyone in the same data center to train a frontier model. ~99% of a model's weights are bit-identical between training steps, so you ship only the ~1% that changed. @Hevalon's team at Covenant formalized this (their PULSE paper) and pre-trained Covenant-72B with trustless peers over the open internet. @huggingface then shipped the same trick in the open. Crypto's coordination layer + a ~100× smaller payload = a big model trained across machines no single entity controls. The other shift I keep noticing: most people still use agents for coding, or as a Google alternative. What's coming is general-purpose agents that research, draft, plan and do your work for you, and the hardware already knows it. The silicon being built isn't sized for one assistant; it's sized for an agent fleet. Two days ago NVIDIA announced the DGX Station for Windows: a GB300 ARM desktop with up to 748GB of unified memory, built to run hundreds of concurrent agents on-prem. You don't build that for a chatbot you visit. If you don't have 5, 10, 50 agents running at all times, local and cloud, doing evals, auto-research, quietly improving work you've already done, you're probably not using AI properly yet. We're early. But the people building the hardware have placed their bet. To the Panathēnea team: thank you, and bravo. See you in 2027. 🇬🇷