3 ms·

Amazing results. Reading the paper I'm frustrated they aren't releasing the full model. Their reasons seem really weak to me. They think it can be used to gene

Amazing results. Reading the paper I'm frustrated they aren't releasing the full model. Their reasons seem really weak to me. They think it can be used to generate offensive pictures or deep fakes and it encodes hegemonic language (e.g. "girl toys" gets more pink and fluffy toys than "boy toys"). These arguments don't make sense for a few reasons - only academics and people capable of reproducing their results should have access to convincing deepfakes? Why is that "safer"? To me it seems less safe, if everyone had this power images would quickly become untrusted, if only a few have this power they could exploit trust in images longer. Second, this argument would hold for things like Photoshop. It makes creating fake images easier. Should Photoshop be banned or weakened because people might create offensive pictures with it? Third, the moment some company implements a good version of this and integrates it with their tools they will make it widely available to users. What use is there in delaying this capability for a couple years so some other company can profit from it?

sailingparrot 5y ago
> What use is there in delaying this capability for a couple years Leaving them time to build all the inference stack so they can host the model behind a paid API, alongside Codex and GPT-3. That's OpenAI's new-ish business model, create amazing models and sell access to the predictions. The golden age of OpenAI state of the art models freely accessible has been over for over a year now. I don't know why they feel the need to keep pretending it's for "safety". For that matter, I think we are generally going to see a trend where less and less big models are distributed publicly, that's not specific to OpenAI. At the scale of those models, training is so expensive and the capabilities potentially game changing that it becomes hard to justify sharing the weights with your competition. See e.g. Turing-NLG from MS, Gopher from DeepMind etc.
quinnjh 5y ago
>if everyone had this power images would quickly become untrusted, if only a few have this power they could exploit trust in images longer. You just answered your question

sailingparrot 5y ago

> What use is there in delaying this capability for a couple years Leaving them time to build all the inference stack so they can host the model behind a paid API, alongside Codex and GPT-3. That's OpenAI's new-ish business model, create amazing models and sell access to the predictions. The golden age of OpenAI state of the art models freely accessible has been over for over a year now. I don't know why they feel the need to keep pretending it's for "safety". For that matter, I think we are generally going to see a trend where less and less big models are distributed publicly, that's not specific to OpenAI. At the scale of those models, training is so expensive and the capabilities potentially game changing that it becomes hard to justify sharing the weights with your competition. See e.g. Turing-NLG from MS, Gopher from DeepMind etc.

quinnjh 5y ago

>if everyone had this power images would quickly become untrusted, if only a few have this power they could exploit trust in images longer. You just answered your question