Oh and I need to add : obviously, one of the issues is that the reviewers did *NOT* catch it. But reproduction of results (even though I provided a Git and seeded code) is another topic entirely.

With DOGE coming for every federal service... careful what you wish for ;)

All of this to say : retracting stuff is healthy. It means we're human and we F up. It means we learn and accept that. I would like to know how many young researchers did what I did and enter that category, and why the older ones don't...

For those that stayed tuned, the Chinese DeepSeek seems to have been trained on chinese-censored-data, which, you know, makes sense, for a Chinese project.

Unless you retrain on text mentioning a certain tank-ridden place, you might see your DeepSeek-based model re-writing history.

*looks at the CEO of a flamethrower company doing a Nzi salute* Yeah... about that...

If you are interested : arxiv.org/abs/2309.00770

Hopefully I am wrong, but we could very well see a Battle for the Language itself, and thus, online reality. Stay tuned... ?

In a world where a lot of online content is generated by language model, even a slight influence can win you a "culture war", by slightly tipping the scales of what is considered good or bad in the language, you can implicitly influence the worldview of potentially millions of people.

Obviously to each their own, but I am shocked by the lack of Wololos in this comment section.

I can just smell the off-by-one error on an array.