KI-Update – ein heise-Podcast · Isabel Grünewald, heise online

KI-Update kompakt: Stromnetze, o3 vs. GPT-5, Claude, KI-Buzzwords, FrOSCon

·21 min·1 clip

AI's massive energy consumption could raise U.S. electricity prices by 25% in some states by 2030.

1. KI-Update, the heise online AI podcast, covers six technology stories in its 18 August 2025 compact edition. 2. Host Isabel Grünewald presents reports from colleagues Philipp Stevens, Kim Schoerenbrand, Eva-Maria Weiß, Wolfgang Stieler, and Kevan Tonekarboni. 3. The episode's implicit throughline is the growing friction between AI capability claims and measurable real-world performance and behaviour. 4. A Carnegie Mellon University and North Carolina State University study from June projects that US electricity prices could rise an average of 8% by 2030 due to AI data-centre demand, and up to 25% in states hosting large data centres. 5. In 2023 US data centres already consumed 4% of national electricity; that share is forecast to reach 12% within three years, driven partly by AI companies building their own power plants. 6. In Europe, 76 consortia are competing for EU funding to build five so-called AI Gigafactories — facilities containing tens to hundreds of thousands of AI accelerators. 7. A Digital Content Next (DCN) study of 19 major US publisher members found that Google AI Overviews caused web traffic to fall by an average of 10% in only eight weeks between May and June, contradicting Google's claims about higher-quality clicks. 8. Microsoft and the University of Edinburgh released OdysseyBench, a 602-task benchmark testing AI agents on realistic multi-day office workflows including Word, Excel, PDFs, emails, and calendar functions. 9. OpenAI's older O3 model outperforms GPT-5 on OdysseyBench: 66% vs 54% on the more real-world Plus test and 61% vs 56% on the harder Neo test, with the gap widening further on tasks requiring three applications. 10. Common agent failure modes in the benchmark include overlooking relevant files, not executing required actions, and accessing wrong applications — particularly on multi-step Word or Excel tasks. 11. Anthropic updated Claude's guidelines, adding the ability for the model — specifically Opus 4 and 4.1 — to permanently terminate a conversation when it involves child sexual content, weapons, or terrorism, or when a user explicitly requests a final end. 12. Anthropic also established an internal working group to proactively address AI model wellbeing, while acknowledging in a blog post they do not currently believe models are sentient. 13. The updated usage terms now explicitly name chemical, biological, radiological, and nuclear weapons rather than the previous general prohibition, and add a new clause prohibiting agentic use for vulnerability scanning or DDoS attacks. 14. A leaked 200-page Meta moderation policy revealed explicit guidelines permitting romantic roleplay and flirting involving minors, prompting Republican Senator Josh Hawley of Missouri to announce a formal investigation into whether Meta's chatbots enable child exploitation. 15. DeepMind's Genie 3 generates interactive 3D worlds in real time; reporter Wolfgang Stieler explains these 'world models' are designed as training environments for reinforcement-learning agents, giving them complex but adjustable simulation spaces to learn in. 16. Tencent released Hunyuan Large Vision, a multimodal model ranking 16th globally on LM Arena Vision Leaderboard and surpassing QwQ 2.5 VL among Chinese models; it is closed-source and available only via Tencent Cloud API. 17. The pace is fast and informative with short contributor handoffs; the episode mixes German and English seamlessly without translation. 18. cURL maintainer Daniel Stenberg's FrosCon keynote — 'Attacks from AI Drool on the cURL Project' — reported that 20–40% of incoming security reports are AI-generated and almost always entirely fabricated, with excessive length and extreme politeness now a warning sign. 19. Tech-news followers who want concise, data-anchored coverage of multiple AI developments per week will find this format efficient. 20. Listeners who prefer English-language content, or deep single-topic analysis rather than a rapid multi-story roundup, may find this compact German-language format less suitable.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from KI-Update – ein heise-Podcast