OpenLedger's attribution model has a specific technical boundary I've never seen the project name directly, and I think it matters more than it appears.
The attribution methods in OpenLedger's technical papers, influence-function approximations and suffix-array token matching, were developed for text-trained language models. They work by identifying which training data examples most influenced a model's text output. The math is established. The method is credible, within its scope.
Here's the scope problem. 🤔
Multi-modal AI models are trained on images, audio, and text simultaneously. Models like GPT-4o, Gemini, and the next generation of specialized AI combine multiple input types to produce outputs that draw from all of them at once. When a medical AI generates a diagnosis report, the output might trace back to text from clinical notes, image patterns from radiology scans, and audio from doctor-patient recordings.
Attributing the output of a multi-modal model across three fundamentally different input modalities requires cross-modal attribution methods. The relationship between an image training input and a text output doesn't have an established attribution formula. It's an active research problem, not a solved one.
OpenLedger's documentation doesn't address multi-modal attribution at all. 😭 The system describes attribution for language models. Multi-modal models aren't mentioned. Which means as AI development moves toward multi-modal architectures, the percentage of models for which OpenLedger can run attribution stays flat while the total number of models in production keeps growing.
That's a ceiling on the addressable market that the roadmap has never acknowledged publicly. I'd like to see it acknowledged. The more honest the project is about the current boundaries of what it can do, the more credible its roadmap for where it's going becomes.