Open-Source Maximalism

From the AI Research Compass wiki

Open-Source Maximalism: Ring title for extreme openness scores: weights, data, code and recipes must all be public, or the result does not count as science.

Beliefs

Open-Source Maximalism is a ring title earned by openness scores above 0.6, independent of position on the main chart. Its holders believe the strongest models’ weights should be released, that results on closed models nobody can reproduce should not count as scientific contributions, and that academia can do frontier research with open models alone. Openness is not a nice-to-have but the condition for verification, safety research, fair competition and global access.

Because the title measures a stance on openness rather than method, it sits on the left ring outside the axes: an open-weights language monolith and an open-hardware robot co-op can both earn it. Its neutral shade reflects that openness says nothing about which research paradigm to pursue; its ambition lies elsewhere, beyond “open weights” toward fully open training data, code, logs and intermediate checkpoints, a standard few releases have met.

Arguments

  • Fully open releases that include training data, code and intermediate checkpoints let researchers study how capabilities and failures arise, which closed models make impossible.
  • Large collaborative open-access models trained by hundreds of researchers showed that frontier-scale efforts can be organised in the open.
  • Open models enable independent safety, bias and security audits, rather than leaving evaluation to the developer alone.

Critiques

  • Once weights are released, safeguards can be fine-tuned away, and misuse cannot be recalled.
  • Fully open data releases raise copyright, privacy and consent problems that are hard to resolve at web scale.
  • Training frontier models costs enormous sums, and openness without a sustainable funding model may leave open efforts a step behind.

Representative works

Relationships

How to get it

Earned in the full or pro test when: Openness >0.6.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data