Post

I wonder: could we as a society just run LLM-accellerated scientific programs for new battery/solar designs, CRISPR-like enzymes, cancer cures - the stuff where outcomes really matter - while also maintaining a humans-only sandbox for these people to experience the joys of the life of the mind in?

I genuinely do not think I am being uncharitable here, we let the amish get on with their lives in peace while not letting them set the pace for the rest of us, maybe that's the play here too

anyway I think this genuinely harms people because if you have an established career writing the same article over and over again you can afford lifestyle LARP but for early-career people? This is poison, this is making them less able to compete and establish themselves, all for her pop-sci brand

atp I no longer believe in that Bender is engaging in good faith, if she ever was. I know gyges said once that her LLM vendetta comes from LLMs disproving her pet language theories?

It would be unfair to say that that’s where her vendetta comes from. It would not be unfair to note that LLMs completely outcompeted the academic project that she poured decades of her career into, and that this might have informed her position.

it would also not be unfair to note that her arguments are very bad, and that she has every reason to know that they’re very bad, but she uses them anyway.

I think I got a headache just trying to understand her twisted logic. But that last sentence? lmao??

So she's literally saying "statistical models of language cannot do anything useful, provided we do not count translation, where they are obviously extremely useful and might suggest applicability of a generalization of the model to similar types of tasks. But that doesn't count though."

So for me the answer to The Chinese Room problem is "I don't care who knows Chinese, I just have a bunch of documents in Chinese that I want translated into English and it seems like this works for that"

Yes, but it only *seems* to work for that, because the room doesn’t have access to *meaning*, it only has access to a thing that’s *like* meaning in all relevant respects.

okay but my counterpoint to Bender is WHO THE FUCK CARES? I don't ask whether or not the LLM is sentient, I only ask it to root cause why a long-running job was causing the OOM killer to fire.

There’s a running philosophical “mimicry” argument that if the thing is programmed, or even if trained from assembled materials, its domain is limited to that vision. It doesn’t even consider usefulness 🤷‍♂️ Even so, models can reason by analogy outside their training set, but no one has updated that