“If all Germans produce 5 pages with GPT-4o every day for a year, the equivalent energy of two nuclear power plants would be needed” - see https://lnkd.in/eKG58ptQ This was just a small excerpt from an amazingly engaging Stockholm AI event that took place on Tuesday 5 November and featured Pruna AI 🍇 co-founder and chief scientist Bertrand Charpentier. The event was focused on green AI 🍀, and the following are some of the key takeaways: 🔬 AI efficiency methods include pruning, quantization and distillation, and some of these can be applied during training as well as post-hoc. 📈 The choice or combination of efficiency techniques should be based on final goals - whether that is the size/footprint or the speed - and weighed against the model performance, i.e. a smaller model may actually be slower than the bigger model, but maybe that is fine if the memory budget needs to be satisfied (e.g. on-device) ✨ Each technique requires deep expertise on its own, but Pruna AI democratizes this with a single smash() function - backed by years of research by Bertrand and his team, with peer-reviewed papers in ICLR, NeurIPS and others. We would like to thank Bertrand for an enthralling presentation, as well as EQT Ventures Carl Svantesson Mats Magdesjö for sponsoring and hosting this particular event. We would also like to Telavox for being Stockholm AI’s primary sponsor and supporter; and our wonderful volunteers Hussein Chukair and Antony P. for helping with the event, and Johanna Granström for the beautiful photography 📷 that captured this event. The full presentation is available here: https://lnkd.in/dn--A9JX #GreenAI #LLMs #DeepLearning #AI #ClimateChange #StockholmAI #EQTVentures #PrunaAI
-
+1