Yeah fair, but it wasn’t a pop-leftist shibboleth; indeed, ten years ago I’d wager the majority of self-identified leftists would have had good things to say about robot rights.

I would recommend you give “Climbing Toward NLU” (2020) a read! It’s short and straightforward, and you can see how she defends this definition. (Not well. :P)

Certainly at this point she’s doing it to maintain clout tho.

It’s more complex than that; she took this position back in 2020, long before it was cool. 😔

Fortunately Bender supplies you with her own definition of “meaning” so that “that’s not what meaning means” can’t touch her. 😇 Her definition is bad, ofc.

First one is the pain axis: models represent self-harm and act to relieve it, which is difficult to explain (given what else the paper demonstrates) without intent. Second one is simply that the agent can notice when it has done something it did not intend, and report it back to the user.

They have almost nothing to do with each other. Bender is a linguist, so in context she means *communicative* intent, i.e. the reason why I say something I say, the thing I am trying to communicate through my words. Alignment deals with ensuring a model’s actions are non-harmful.

I agree that it’s a wrong question to ask; what frustrates me about Bender is that she answers the wrong question, and answers it with a lie, and then uses that lie to fight against anyone asking better questions. 😔

Digger certainly reads as smug to many people, which I think is quite intentional. I would recommend looking past that, and treating the movie as a puzzle that Iñárritu wants you to solve. What does he want you to be thinking about, as you’re walking out of the theater?

To be clear, I don’t think Emily M. Bender is stupid. I think she knows that her argument is fallacious. I think she knows her premise doesn’t support her conclusion. I think her public outreach is reactionary noise, offered in bad faith with false authority. It’s worse that she’s not stupid.

Yeah, modern cybernetics/control theory doesn’t draw on him at all, because they wanted to focus on robots. And unfortunately, the management and governance people don’t think in terms of math. But i think they should! I think there’s undiscovered ground there!

As for LLMs having intent, I would argue against that. I actually have two distinct arguments against that, one based on mechanical interpretation, the other based on the observed behavior of agents!

It’s actually not about the statistical regularities at all, strangely enough. She simply argues that any machine trained only on form can never learn meaning. She then refuses to admit that LLMs are not “trained only on form”, despite all evidence. :P

I think Occam’s razor is probably correct in some sense (Kolomogorov), but it is also *vastly* underspecified, and it has been misapplied constantly since it was first stated. (… actually it was misapplied over a millennium before William of Occam, by Ptolemy, to argue the earth is stationary. :3)

Which is to say — you’re right, Will, but you shouldn’t be surprised to hear this type of argument used against you. In fact, you’re in very good company!

In 1543, Copernicus published his heliocentric model, which posited that the stars are unfathomably far away. Brahe used Occam’s razor to argue against this, because it was simpler to think the earth was at the center of the universe, and the stars close.

I think Occam’s razor is probably correct in some sense (Kolomogorov), but it is also *vastly* underspecified, and it has been misapplied constantly since it was first stated. (… actually it was misapplied over a millennium before William of Occam, by Ptolemy, to argue the earth is stationary. :3)

In 1925, Gaposchkin discovered that stars are almost entirely hydrogen and helium. Russell used Occam’s razor to argue against this, because the simplest explanation was that the universe was uniform in composition, with the same elements showing up in the same ratios everywhere.