Join 40,000 readers and get the full notes and summary: https://newsletter.aiwithkyle.com/subscribe 🎟️ Kyle recently appeared as the first-ever guest on the "AI for the Rest of Us" podcast, available on Spotify, YouTube, and Apple Podcasts. He's also speaking at their upcoming conference on 15-16th October in London: AI for the Rest of Us conference, London : https://luma.com/AIfortherestofus25 (Use code **IAMWITHKYLE** for a discount). Subscribe and turn on notifications to catch the next live stream: https://www.youtube.com/@iamkylebalmer?sub_confirmation=1 Want to become fast-track becoming an AI Trainer? Apply to join the next AI Workshop Kit cohort: https://newsletter.aiwithkyle.com/c/workshop-kit This video is from the ‘AI with Kyle News and Updates Live Stream’ First aired: 9 September 2025 Watch full edited live stream: https://youtu.be/GzK1U8McvoU —— More Useful Resources —— My other daily live show - 30-day building a business in public challenge: https://www.youtube.com/live/Z0Z_1Eg3HUk (day one) Free AI Entrepreneurship Playbooks: https://aiwithkyle.com/catalog Free Quiz: What AI business should you start? https://aiwithkyle.com/tools Free AI Training Business Planner: https://aiwithkyle.com/tools/workshop-builder Free 10-Week ‘Vibe Coding’ AI Summer Camp: https://aiwithkyle.com/courses/10-week-ai-summer-camp —— Contact —— Best is in YouTube comments. For business enquiries only: https://www.passionfroot.me/iamkylebalmer OpenAI published research revealing why AI models hallucinate - they're overfitting to benchmarks that penalize "I don't know" answers, forcing models to make up plausible-sounding but wrong information instead. The paper admits the industry has been too focused on test scores rather than user utility, and suggests changing evaluation methods to reward honest uncertainty when models don't actually know something.