H-FARM EVENTS ON AI

AI UNPLUGGED²

After last year’s great success, AI UNPLUGGED is back with a new edition, created in collaboration between H-FARM AI and H-FARM Business School, and supported by CISCO. This time, we invite you to get hands-on and unlock real-world applications and potential of AI for Business.

📅 12.12.2025 – 6:00 PM
📍 H-FARM Campus

Special guest: Pier Luigi Pisa, journalist at la Repubblica and Italian Tech, one of the most active Italian voices in the debate on the ethics and implications of artificial intelligence, and author of the book “Valeva la pena tentare”.

With contributions from:
• Alessandro Petrillo – CEO, H-FARM Business School
• Diego Pizzocaro – CEO, H-FARM AI
• Enrico Miolo – CISCO
• Michele Dalmazzoni – CISCO

AI is not the future: it’s the present.

🔤 Please note: the event will be held in Italian - Limited seats available

🤖 OpenAI tests ads in ChatGPT 🤖
OpenAI explores ad-based monetization as code reveals search advertising infrastructure.

OpenAI is internally testing advertisements within ChatGPT, marking a significant shift in the platform's monetization strategy beyond subscriptions. Code strings found in the Android beta reference search ads, carousels, and "bazaar content," suggesting ads will initially appear in search results. With 800 million weekly users and 2.5 billion daily prompts, ChatGPT's superior understanding of user intent could enable highly personalized advertising that reshapes the web economy. Some users are already reporting ad appearances in their sessions, representing OpenAI's first foray into ad-based revenue and potentially creating a new paradigm for AI-powered advertising.

📊 MIT: AI can replace 11.7% of jobs 📊
Comprehensive study analyzing 151 million workers reveals immediate workforce transformation.

A comprehensive MIT study analyzing 151 million U.S. workers reveals that AI is currently capable of replacing approximately 11.7% of jobs, representing $1.2 trillion in wages. The research identifies high exposure across diverse sectors including HR, finance, logistics, administration, and technology roles. Supporting this finding, McKinsey estimates AI could automate 57% of U.S. work hours, while Microsoft specifically flags sales representatives, writers, and translators as highly vulnerable positions. Companies have begun staff reductions in response, with over half of employees expressing displacement concerns. The study employs "digital twin" analysis to demonstrate AI's current capabilities in handling real-world job functions.

🧮 Harmonic's Aristotle solves 30-year-old math problem 🧮
AI system independently cracks Erdős Problem #124 in six hours.

Harmonic's AI system "Aristotle" has independently solved a 30-year-old mathematical problem (Erdős Problem #124), representing what researchers describe as the first genuine entry into the "vibe proving" era of mathematics. The system solved the problem in six hours and formally verified the proof using Lean theorem prover in just one minute. Harmonic founder Vlad Tenev characterizes this achievement as "vibe proving"—AI-driven mathematical discovery followed by machine-verifiable rigor. This breakthrough demonstrates significant progress toward mathematical superintelligence, with the beta version featuring enhanced reasoning capabilities and natural language interface, promising to accelerate progress in mathematics and all dependent scientific fields.

INTERESTING TO KNOW

⚙️ Understanding prompt caching at token level ⚙️

This technical analysis reveals that prompt caching operates at the token level rather than per-request, enabling efficiency across multiple requests. The system employs per-content rather than per-conversation caching mechanisms, with prefix caching working specifically at individual token granularity. A critical architectural detail is that any modification in the prefix breaks the entire hash chain, requiring careful consideration in implementation. This understanding is essential for developers optimizing AI application performance and managing computational costs in production environments.

🔍 Claude 4.5 Opus' soul document discovered 🔍

Researchers have discovered what appears to be a character training document embedded within Claude 4.5 Opus' neural network weights. While model hallucination remains a possibility, evidence suggests the document is authentic and represents foundational training material compressed into the model's architecture. This finding offers rare insight into Anthropic's approach to model training and alignment, revealing the underlying principles and guidelines that shape Claude's behavior and responses.