Comments on: GPT-3 vs Bloom [Real Code & Real Results] https://enjoymachinelearning.com/blog/gpt-3-vs-bloom/ All Machines Learn Thu, 22 Feb 2024 18:51:08 +0000 hourly 1 https://wordpress.org/?v=6.4.5 By: Dylan https://enjoymachinelearning.com/blog/gpt-3-vs-bloom/#comment-107 Thu, 09 Feb 2023 02:57:15 +0000 https://enjoymachinelearning.com/?p=2064#comment-107 In reply to DZoh.

You’re on the right track, though I don’t think Bloom would want their product compared to a lesser GPT.

I know most of the hype surrounding Bloom was being self-hosted, so we went with a model version that could be… self-hosted.

The download on even just this one was atrocious, and I couldn’t imagine downloading a more expansive model.

]]> By: DZoh https://enjoymachinelearning.com/blog/gpt-3-vs-bloom/#comment-106 Wed, 08 Feb 2023 19:54:44 +0000 https://enjoymachinelearning.com/?p=2064#comment-106 If I understand correctly you haven’t download the “real” bloom model (which is more than 300gb) but a smaller one. I’m not sure it would really fit inside a regular computer anyway.
Therefore it seems normal than the generated output are… well… more than disappointing since it’s like comparing Gpt2 to gpt3.
Bloom (the large one) has an API accessible on Huggingface. Even for this large model I still find the results difficult to compare to the one provided by GPT3. They are less usable.
Yet I think that the price of GPT3 token generation (and the time it takes) would remain a problem if you plan to use it intensively in any deployed app.

]]>