Serious issues in Llama 4 training. VP of AI at Meta resigned

  • https://x.com/Ahmad_Al_Dahle/status/1909302532306092107

    “ We've also heard claims that we trained on test sets -- that's simply not true and we would never do that. Our best understanding is that the variable quality people are seeing is due to needing to stabilize implementations.”

  • If the developer’s Christmas bonus depends on scoring high on a particular benchmark, it is not inconceivable that benchmark somehow would make its way into the training set - directly or indirectly.

    I don’t think the management would have to directly encourage it as stated in the Chinese text.

  • VP of AI at FAIR, who is unrelated to the llama/genai team.