As I said - they are doing it. Not maybe, they are. Can't stop terrorists or bot networks using a Chinese open source model. - ThreadSky

scottjla.bsky.social • 89 days ago

As I said - they are doing it. Not maybe, they are.

Can't stop terrorists or bot networks using a Chinese open source model.

Comments

carlo.pinciroli.net•89 days ago

We need to agree on what "open source" means though. To my knowledge, Chinese companies are not opening their training data, but the model weights. It's hard to retrain, but it's easy to integrate with other products. That's not really what "open source AI" means: http://tiny.cc/1obyzz.

erik-mols.bsky.social•89 days ago

In my opinion for AI to be free and open source it needs to be open on A training data B Algorithm and C the end results

scottjla.bsky.social•89 days ago

By open on training data, do you mean the actual text of the training data? Or an index?

erik-mols.bsky.social•89 days ago

As long as you can repeat the process yourself.

scottjla.bsky.social•89 days ago

Probably if it involves paying for the licences...

sir-deenicus.bsky.social•89 days ago

IMO, the importance of disclosing training data becomes overstated once token counts have entered into tens of trillions. It is: anything you can get your hands on.

The real questions become about crawling approach, stage of synthetic data use, plus dedup, filtering and other preprocessing steps.

carlo.pinciroli.net•89 days ago

I agree. A possible middle ground for AI open sourcing would be to distribute not the full data, which might be infeasible, but the scripts used to fetch it and clean it up.

scottjla.bsky.social•89 days ago

A lot would come from off line data too, and potentially licenced too depending on who's producing it

sir-deenicus.bsky.social•89 days ago

Other things like architecture choices, ++expensive to gather observations on what worked, hyperparam settings & parallelization strats to ensure high GPU utilization are just as, if not more important. The chinese companies (esp Deepseek) have so far been quite open about those type of questions

Comments

Posting Rules

Reply