For me, the real problem is that internal models seem to be more capable than the publicly released ones. My theory is that Anthropic actually has one very powerful model that they’re splitting into smaller pieces so they can maintain a publication/release roadmap over several months, at least.
And in the end, it seems that the only real breakthroughs have been achieved using internal models. The mathematicians associating themselves with this are frauds...
Indeed, the issue is whether such a publication has any value at all, even without taking full credit for it (although the ideas suggested to the AI are still ours). My view is that only truly groundbreaking results should be published. And of course, I keep my deeper ideas to myself: I don’t want to give them to AI systems, which ultimately belong to companies.
Pas faux, mais, connaissant relativement bien le milieu de l’enseignement, je me demande : quels profs ont envie d’aller dans ces établissements ? Il faut nécessairement une refonte totale de l’éducation et de l’enseignement. Or, Rima Hassan fait historiquement partie de ceux qui ont permis à l’école de devenir ce monstre.