In May, AI WEEK in Milan attracted about 29,000 visitors from more than 70 countries. The two-day program featured nearly 700 speakers across 17 stages; organizers provided nine and a half days of recordings so attendees could catch up on parallel sessions they missed.
Two focuses: big names and practical demos
The first day featured high-profile figures — EU commissioners, ministers and star speakers. The second day leaned toward practical content: tools, live demonstrations and deployable solutions aimed at answering how organizations should work with AI back at the office.
AI agents were among the most discussed topics: autonomous systems, multi-agent setups, vibe coding (AI-assisted natural-language programming), self-healing systems and world models. Industry-specific, open-source solutions and on-premises architectures also gained attention.
Sovereignty: who owns the future?
AI sovereignty emerged as a prominent concern. Recent geopolitical developments, speakers argued, have exposed vulnerabilities: European countries and companies can be dependent on external providers. According to the Stanford AI Index, the US private sector invested $286 billion in AI last year — 14 times the amount invested by Europe and 23 times that of China (state funding in China is significant but not reliably quantified).
The only clear threat to US dominance highlighted at the event was a talent shortfall: last year the number of AI researchers and developers relocating to the US fell by 80% in a single year.
Star talks: risks and alternatives
Karen Hao summarized the darker sides of AI: exploitation linked to content moderation, the formation of knowledge monopolies, environmental strain from data centers and concentration of power without democratic oversight. She noted that data labeling for AI training is now the fourth fastest-growing occupation in the US, increasingly performed by displaced, degree-holding white-collar workers — a trend that can deepen inequalities. Hao also pointed to shocking environmental figures, including comparisons about energy use at large data centers.
At the same time, she highlighted open-source models (such as Llama, DeepSeek, Mistral) as positive examples of competitive performance with lower resource demands, and warned about the risk of model collapse if high-quality human-generated data become scarce.
Replit: stories matter more than tokens
Michele Catasta, president and head of AI at Replit, argued that success should be measured by human stories rather than token consumption. He told of a mother who built a spelling app on Replit for her dyslexic child. Replit is targeting not only professional developers but also people who cannot code — a market that represents roughly 99% of the population and has been rapidly shaped by vibe coding players.
On AGI, Catasta was cautious: he suggested the AGI narrative can be used to keep the accelerator pressed, motivating companies to build ever larger models.
Llion Jones: transformers’ origin and limits
Llion Jones, founder of Sakana AI and co-author of the 2017 "Attention Is All You Need" paper, recounted how the transformer architecture enabled parallel processing of tokens via attention mechanisms — a breakthrough that matched the capabilities of then-available hardware. The paper has become one of the most cited in the field, with about 250,000 citations referenced in the event coverage.
Jones warned that the transformer’s success can be misleading: it suited particular hardware and was not necessarily the smartest theoretical path to future breakthroughs. Working in Tokyo on post-transformer approaches (including The AI Scientist and new image algorithms), he urged a shift from simply reworking transformer variants: "Stop. Nothing will come of it," he said, arguing for new directions.
Pascale Fung: cognitive world models and partnerable AI
Pascale Fung, co-founder and research lead at AMI Labs, argued that AI can become a full partner if it understands physical reality and learns from experience — not just solve equations but grasp what it means to fall. She described cognitive world models that integrate physical context (space, laws of nature) with mental models (intentions, goals, norms).
Fung presented JEPA (Joint Embedding Predictive Architecture), an architecture that observes human activity (for example, in a kitchen) and intervenes only when needed. Unlike an LLM that parrots a recipe, JEPA watches the real world, interprets cause and effect and tracks human intention and social norms. Importantly, Fung emphasized that safety must be embedded from the start, not added as an after-the-fact guardrail.
Nataliya Kosmyna: tools change how we think
MIT researcher Nataliya Kosmyna shared experiments measuring students’ brain activity while writing essays under three conditions: using ChatGPT, using Google search (without AI features), or without tools. Results showed that students who worked without tools activated more brain regions; the Google group also performed well because they processed information themselves. The ChatGPT users scored worse on several measures: their content was flatter and more career-focused, and 83% of the ChatGPT group could not accurately cite the sources they submitted; 16% felt the essay had nothing to do with them.
Kosmyna dispelled several myths: Socratic prompting does not automatically save us, because humans prefer ready-made answers when tired or rushed. Nor is it always beneficial to outsource shallow tasks (like email or calendar management), because such tasks can act as restorative buffers between deep work tasks. She also noted the productivity paradox: people spend 19% more time verifying LLM outputs and accept only 48% of the responses.
Creative AI: filmmakers and marketing teams
Alex Mashrabov, founder of Higgsfield AI, presented the full-length film "Hell Grind," created in two weeks without physical sets or human actors and with a budget of about $500,000. He highlighted the financing challenges filmmakers face — in many countries it is nearly impossible to raise a million euros for a film — and argued that AI solutions could revive many dormant screenplays.
Higgsfield’s "Supercomputer" creative AI agent targets advertising agencies and corporate marketers: it autonomously analyzes trends, generates campaign concepts and assembles video cuts, reducing tasks that used to take weeks to hours so creative teams can focus on higher-level work.
Atmosphere and organisation
AI WEEK defied a single label: it was simultaneously a conference, trade show, startup fair, hackathon and networking event. Attendees from diverse backgrounds — corporate leaders, small-business owners, business users and IT engineers — reported a productive environment, aided by an event app for filtering programs and scheduling meetings. Some participants arranged dozens of business meetings during the two days.
Logistics were smooth: most volunteers spoke good English, AI tools were used for live translation and subtitling, and program elements ranged from rock bands and DJs to art installations like an "AI confessional booth." Humanoid robots and gaming corners were present but drew less enthusiasm than in previous years.
Takeaways and practical next steps
Although the in-person event lasted two days, recordings and case studies will continue to influence practitioners for months. Tickets for next year were already available: at the time of reporting the cheapest 2027 "Business Pass" cost €159.
The article was written by Szurop István, curator of AI Talks. AI WEEK’s official Hungarian media partner is HVG Kiadó Zrt.; the coverage was produced within the AI Talks and HVG event framework with HVG BrandLab’s cooperation. The HVG weekly and hvg.hu editorial teams were not involved in production or editing.
