4 ms·

NLP researcher here. Leaving aside the discussion on whether the model is really dangerous or not, the paper gives enough details that the model can be reasona

NLP researcher here. Leaving aside the discussion on whether the model is really dangerous or not, the paper gives enough details that the model can be reasonably reproduced (and even if it couldn't, it's an incremental result built on other research that is public, the novelty is mainly in scale). Any big company, government or lab with enough compute can recreate it (and maybe improve it) in six months, maybe less. Not releasing the model doesn't prevent misuse. It just restricts the set of entities that can afford it: only those with enough resources can build it and misuse. Those with less resources cannot, because training it is very expensive. I personally don't feel very reassured by knowing that only the powerful have access to this. The actors I fear the most are powerful and could train it, anyway. I'd rather it be public so the community can scrutinize it, find out if it's really worth worrying about, and find ways to detect/fight it.

ace_of_spades 8y ago
Thanks for laying out your reasoning here :) However, I don’t think you have fully adressed the worrying point. The concern here is that good quality faking will become less expensive, thus, more faking will happen. Making it even just marginally more expensive means that there will be less faking overall (no matter the resources of actors) which is seen as good. Thus, your argument towards more equitable access seems to have to explain, how it provides more value than the harm that may be caused by this dynamic, which would in my opinion require some argument for the potential good that could be done by this technology vs. the harm it could cause - which you haven‘t really done. Just arguing for the „balancing of scales“ doesn‘t go far enough. For example, you can draw an analogy to weapons regulation. Just trying to balance the scales by giving people easy access to weapons has shown to not work too well in practice. The rich still have more money to buy better, bigger weapons or defenses while the poor shoot themselves almost regularly. So, to make a convincing case for open sourcing this stuff you should provide a good argument for why we are not talking about weapons here but positive, productive goods with low potential for abuse.

ace_of_spades 8y ago

Thanks for laying out your reasoning here :) However, I don’t think you have fully adressed the worrying point. The concern here is that good quality faking will become less expensive, thus, more faking will happen. Making it even just marginally more expensive means that there will be less faking overall (no matter the resources of actors) which is seen as good. Thus, your argument towards more equitable access seems to have to explain, how it provides more value than the harm that may be caused by this dynamic, which would in my opinion require some argument for the potential good that could be done by this technology vs. the harm it could cause - which you haven‘t really done. Just arguing for the „balancing of scales“ doesn‘t go far enough. For example, you can draw an analogy to weapons regulation. Just trying to balance the scales by giving people easy access to weapons has shown to not work too well in practice. The rich still have more money to buy better, bigger weapons or defenses while the poor shoot themselves almost regularly. So, to make a convincing case for open sourcing this stuff you should provide a good argument for why we are not talking about weapons here but positive, productive goods with low potential for abuse.