Want a detailed analysis? Be ready to sweat for it! One of the cool things I did at #SmartCityDay is ride a stationary bike for data science
🔌🚴📉🔥

By extrapolating the (virtual) energy cost of prompts to a GPT-OSS 120B model on a Mac Mini server, the experiment set up by the BFH medical informatics group called #enerKI helps to make the relative energy efficiency of simple and complex prompt responses evident.

Main takeaways from the closing panel:

- #Apertus needs to be used to explain how AI works across the production chain, used as a model of models, and encourage more local “baby Apertus” quantized and embedded versions;
- There is a strong interest in using academia-approved LLM in education and science communications;
- AI is still seen as a curious toy in much of industry, not delivering clear ROI, threatening to pull various carpets out from under feet;
- Research on the training side is needed to establish not just competitiveness, also dataset bias and ethical guardrails;
- Vibe coding is here to stay, but probably not really where we would expect;
- Public AI is a movement to even the playing field that we should embrace.

With thanks to the organizers particularly @barochat for a spectacular #SmartCityDay