Sharp Tech with Ben Thompson (Preview) Google Starts Dancing, The Winners and Losers of Gemini Week, OpenAI Has an Advertising Problem
Nov 21, 2025
Gemini 3 is shaking up the AI landscape with impressive benchmarks and a unique cost advantage thanks to Google's TPUs. The conversation explores Google's potential in the enterprise market and the challenges facing OpenAI if it can't pivot to an advertising model. There's a fascinating analysis of Anthropic's coding prowess and a dive into why some companies might struggle in this evolving race. The hosts also discuss how narratives shape perceptions in tech and the implications of Gemini's arrival for established players like Nvidia and Amazon.
AI Snips
Chapters
Transcript
Episode notes
Verifiable Tasks Make Benchmarks Valuable
- Benchmarks are most useful in verifiable domains like coding where outputs can be validated and improved via reinforcement.
- Anthropic retained a lead in coding benchmarks, showing niche strengths persist despite overall leaderboard shifts.
You Get What You Measure
- Large companies tend to optimize for measured KPIs, which can lead to building toward benchmarks rather than real-world robustness.
- That drives both genuine progress and risks of models overfitting benchmark tests.
Personal Tests: ChatGPT Beat Gemini
- Ben tested ChatGPT and Gemini on networking and turkey prep and found ChatGPT more practically useful in both cases.
- Gemini missed key details like fridge drying for turkey and QoS for networking, undermining confidence.