Choosing models matters. How do you choose?

AI Learning Labs
1

Following our online roundtable on “How to build a responsible climate chatbot?” I jotted down some reflections for the OKFN blog. Some of what we discussed was specific to the issue of climate change and climate data, but the bigger picture concerns about generative AI and small/medium/large language models applies to basically… everything.

I’ll share some bullet points here, but it made me wonder how people in this community are choosing AI models for different projects and contexts.

  • There’s the energy use and emissions by use of different models
  • The privacy concerns
  • The way different models output different text depending on training data, language and geopolitical context (eg. misinfo in one language but not another)
  • There are cost factors, corporate ethics, openness and transparency

The list goes on…

For personal use, the AI recommendation tool Bearing has a lovely method for identifying priorities and matching them to different AI models. For developers, the AI Energy Scoreboard on Hugging Face offers a comparison of energy efficiency of AI model inference. What else have you seen or used lately to seek recommendations?

How are you making choices that align with your personal values or organisation’s mission? Or did you give up on finding any middle ground?

2 Likes
2

I’ll be honest - I don’t always know what model I’m using (but I’m trying to learn). I use (paid) Cursor to develop code and it seems to choose from a range of models.

For the development of the chatbot system (for AI-Learning-Labs) we have tried to constrain the LLM to be small and potentially independent of megacorporations. My wish is that it would be small enough to run in a rural school’s computer (maybe with a. custom GPU) that could still do useful work when disconnected from the net. Speed and comprehensiveness is not important.

Please correct my vision if it’s unrealistic.

1 Like