Daniel Erasmus keynoted the World AI Summit in San Francisco by presenting the team’s recent work building ClimateGPT 2, and ongoing efforts toward #3. The talk touched on the challenges of building small, outsized capable Large Language Models SLMs . These models democratise access, use less CO2 use per equivalent query and by focussing on a specific domain and enable fine grained corpora curation, accelerated development and task definition.
In the case of ClimateGPT 1 and 2 this pioneering approach lead to a significant reduction in model size resulting in 12x less CO2 per response whilst maintaining knowledge across a range of benchmarks equivalent to models ten times the size.
12x reduction in CO2 per query for equivalent results.
Daniel discussed the partnership between Club of Rome & ClimateGPT and the importance in grounding the model in the 5 interventions proposed by the Earth4All report, the 17 Sustainable Development Goals, forward looking technical & systemic breakthroughs, a decade of the effects of extreme weather events and core of climate change research.
These are not meteorological models.
We are in the anthropocene and must get good at understanding how we, not just the weather- will change.
ClimateGPT begins to describe the effects of climate change on human systems. Food insecurity, water stress, costly access to capital, migration, institutional failure, political upheaval, etc.
View the Keynote or a Short Interview in the margins of the conference below:
World AI Summit San Francisco: Short Interview (2.30min):
World AI Summit San Francisco: Keynote (20min):
Leave a Reply