Teaching rocks to speak

Teaching Rocks to Speak: The Promise of Large Geologic Models and Generative AI. If you are attending the IMAGE conference this month you may wish to pop over to see Andrew Davidoff’s presentation.

In my opinion it’s a non-technical high level thought provoking piece on subsurface Multi-modal Large Language Models (MLLM) from the perspective of someone who is an experienced industry geologist. The opportunity, value, business challenges and the legal structures to be addressed if proprietary and public content are to combined.

There are also some fundamental questions on whether it’s actually worth fine tuning state-of-the-art LLMs as encoder (generative) models at all…

On another note, over the past 18 months there have been, and continue to be, some serious issues around ethics in international facing geology specific LLM Chatbots and Applications.

These issues range from:
– data privacy (of personal data and confidential company content) – lack of informed consent measures, breaking various AI ethical recommendations from UNESCO, and various country and jurisdictional laws.
– intentional censorship of geoscience data outside of international norms.
– misleading and incorrect use of promotional terms like ‘Open science’ and ‘Open source’ when training data and code is not released so technology is not reproducible and models are opaque
– deliberately vague terms of use and what rights you are giving to the proprietor for online AI
– dual use motivations for AI, all use cases not disclosed
– potential exploitation of certain communities
– copyright infringement
– bias and transparency
– meaningful fairness and equity

These may present opportunities to learn from, and move towards more ethical approaches in the subsurface and geosciences.

Disclaimer: Andrew reached out to me sharing his slides this week, which I thought were a thoughtful and positive contribution to the debate. I have not contributed to the slides nor do I endorse them or benefit in any way for this ‘plug’.

Image conference: https://www.imageevent.org