🧠 OpenAI explains model hallucinations 🧠
Groundbreaking research reveals training methods reward confident guessing over admitting uncertainty.

OpenAI explains model hallucinations through groundbreaking research showing that standard training methods create a fundamental conflict where models learn to always guess even when uncertain. Current evaluation methods give full points for lucky guesses but zero for saying 'I don't know,' rewarding confident guessing over admitting uncertainty. The research found that advanced models like o3 hallucinate 33% of the time—double its predecessor o1. The solution involves redesigning evaluation metrics to penalize confident errors more than expressions of uncertainty, which could significantly improve AI reliability and trustworthiness across applications. 

⚖️ Anthropic pays $1.5B copyright settlement ⚖️
First major AI company payout for using copyrighted content sets industry precedent.

Anthropic pays $1.5B copyright settlement to authors, agreeing to pay approximately $3,000 per work for allegedly using their books to train its AI models. The settlement covers around 500,000 books downloaded from shadow libraries like LibGen, marking the first major payout from an AI company for using copyrighted content. A federal judge ruled it was legal for Anthropic to train on copyrighted materials but illegal to acquire them from pirate sites. The company must also destroy the original files, though the settlement only covers past use, setting a significant legal precedent for the AI industry's relationship with copyrighted training data.

🇨🇳 Alibaba releases 1 trillion parameter model 🇨🇳
Qwen3-Max-Preview demonstrates superior performance with massive context window capabilities.

Alibaba releases Qwen3-Max-Preview with 1 trillion parameters now accessible via Qwen Chat and Alibaba Cloud. The model demonstrates superior performance compared to Claude Opus 4, Kimi K2, and DeepSeek-V3.1 on reasoning and coding tasks, while supporting an impressive 262K token context window. Competitive pricing starts at $0.86 per million tokens for shorter prompts, strengthening China's position in the global AI race. The release marks a significant milestone in large language model development, showcasing China's growing capabilities in frontier AI research and deployment.

INTERESTING TO KNOW

📱 Google launches EmbeddingGemma for mobile AI 📱

Google releases EmbeddingGemma, a 308M parameter embedding model designed specifically for on-device AI applications. The model delivers best-in-class performance for its size, enabling Retrieval Augmented Generation (RAG) and semantic search that run directly on mobile devices without internet connectivity. With sub-200MB RAM usage and 15ms inference time, EmbeddingGemma supports 100+ languages and offers customizable output dimensions from 768 to 128 via Matryoshka representation learning.

💸 OpenAI projects $115B cash burn through 2029 💸

OpenAI projects $115B cash burn through 2029, representing an $80 billion increase from previous expectations. The company will burn more than $8 billion this year alone, escalating to $17 billion in 2026 and $45 billion by 2028, driven by soaring data center, talent, and compute infrastructure costs. To offset these massive expenses, OpenAI plans to develop its own data center server chips through partnerships with Broadcom.

H-FARM EVENTS ON AI

Club AI @ H-FARM Campus - September 18, 6:00 PM

Quick heads-up (and tiny shameless plug): we’re highlighting an H-FARM event in this newsletter - because we genuinely think it might be interesting if you fancy a trip in our wonderful campus just outside Venice (for free! 🙂) discussing AI.

H-FARM’s AI Club is a monthly community meetup that brings up to 500 people together to cut through the AI noise with real news, real cases, real debate.

On stage this edition:

• Diego Pizzocaro (CEO, H-FARM AI): the latest AI news but mixed with our comments, perspectives and your questions.

• Luca Pezzullo (President, Order of Psychologists of Veneto Region): “AI & Consciousness: Illusion, Reality, and Risks.”

• Enterprise Voice Agents (CISCO): what really works in large organizations - architecture, rollout, ROI with AI Voice tech.

Language: English (full session)