Artificial Intelligence/Generative AI: Difference between revisions

From Open Food Facts wiki
No edit summary
No edit summary
Line 2: Line 2:


=== Creating a foundation service ===
=== Creating a foundation service ===
* Creating a foundation service that could execute GenAI requests, and abstract the model being used
* Creating a foundation wrapper that could execute GenAI requests, and abstract the model being used (using cloud services at first)
* Ensuring 3rd parties can't abuse the service, while community developpers can use it freely
* Ensuring 3rd parties can't abuse the service, while community developpers can use it freely
* Making testing and iteration easy.
* Taking safety measures so that it doesn't cost the Open Food Facts NGO too much
* Taking safety measures so that it doesn't cost the Open Food Facts NGO too much



Revision as of 11:25, 10 June 2024


Creating a foundation service

  • Creating a foundation wrapper that could execute GenAI requests, and abstract the model being used (using cloud services at first)
  • Ensuring 3rd parties can't abuse the service, while community developpers can use it freely
  • Making testing and iteration easy.
  • Taking safety measures so that it doesn't cost the Open Food Facts NGO too much


Ensuring capitalization of prompt knowledge

  • Listing useful prompts, and creating metrics for them


Avoiding wasteful compute

  • Ensuring outputs are stored and shared
  • Ensuring we group prompts
  • Ensuring prompts are iterated and evaluated before being run at large scale


Enhancing how we work and volunteer

  • Ensuring knowledge about Open Food Facts as an org is well disseminated so that LLMs can help us more efficiently (eg: that it knows what Robotoff is, or which database we're using, …)
  • Looking for opportunities in the tools we already use (eg GitHub Copilot from GitHub, new LLM features in Odoo, …) and new tools that could save us time or increase our impact

Monitor and document innovative usages outside Open Food Facts

  • Document good practises
  • Document things we could also implement

Think hard about how we can use it

  • Data quality, Data extraction, Content synthetisation, Data models generation, Code generation, Food intake analysis, Natural search parsing, those are some of the potential themes, but there are many others

Don't be paralyzed by potential errors or phantasms

  • It is possible to mitigate/eliminate those potential errors by keeping humans in the loop, and adopting a copilot approach
  • These new techniques can be hidden as a helper in the UI (like a magic button), with the proper warnings


Prioritize and implement key projects