[zdwolfe]

I find it odd that any LLM could be considered open source. Sure the weights are available to download and use, but you can't reasonably reconstruct the output model as it's impractical for an individual to gather a useful dataset or spend $5,000,000+ of GPU time training.

[_ea1k]

Distillation can extract the knowledge from an existing model into a newly trained one. That doesn't solve the cost problem, but costs are steadily coming down.

[goku12]

That's still a crude repurposement of an inscrutable artifact. Open source requires you to share the source data from which that artifact (the model parameters) was created.