Also a concept that Scott Aaronson praised Hanson for.

https://web.archive.org/web/20210425233250/https://twitter.com/arthur_affect/status/994112139420876800

(Crediting the "Great Filter" to Hanson, like Scott Computers there, sounds like some fuckin' bullshit to me. In Cosmos, Carl Sagan wrote, "Why are they not here? There are many possible answers. Although it runs contrary to the heritage of Aristarchus and Copernicus, perhaps we are the first. Some technical civilization must be the first to emerge in the history of the Galaxy. Perhaps we are mistaken in our belief that at least occasional civilizations avoid self-destruction." And in his discussion of abiogenesis: "Life had arisen almost immediately after the origin of the Earth, which suggests that life may be an inevitable chemical process on an Earth-like planet. But life did not evolve beyond blue-green algae for three billion years, which suggests that large lifeforms with specialized organs are hard to evolve, harder even than the origin of life. Perhaps there are many other planets that today have abundant microbes but no big beasts and vegetables." Boom! There it is, in only the most successful pop-science book of the century.)

Honestly Hanson is so awful the rationalists almost make him look better by association.

We have some threads of Vaccinations in Book/Article Form which try to share good pop science and textbooks without the cult shit and Dunning-Kruger. People who think they know everything and are mysteriously underemployed tend to have the most time to post though.

nah they're built different

It's two guys in London and one guy in San Francisco. In London there's presumably no OpenAI office, in SF, you can't be at two places at once and Anthropic has more true believers/does more critihype.

Unrelated, few minutes before writing this a bona-fide cultist replied to the programming dev post. Cultist with the handle "BussyGyatt @feddit.org". Truly the dumbest timeline.

Didn't OpenAI just file court documents claiming that their opposition is funded by competitors? Accusing someone else of what they themselves are doing seems to be a pretty popular strategy these days.

I dont know anything about the locations of any offices, but could it he that openAI just didnt have any local places? Asking them why not all worked ld be a good journalist question

But otoh it is just ~~two~~ three of them, and the second ones photo gives off a weird vibe. Why is he smiling like it is a joke?

Oh, it's a CXL board, Compute Express Link. Basically a way to attach DRAM to PCI Express. I know some people working on this stuff for one of the big vendors, but in that context it was a rack-scale box capable of handling multiple terabytes' worth of DIMMs. Having this as a desktop expansion card seems like a bit of a marginal application, but Gigabyte's done weird shit before. For instance, I have an AMD-compatible Thunderbolt 3 card that was only made in limited quantities by them and ASRock.

I think that you have useful food for thought. I think that you underestimate the degree to which capitalism recuperates technological advances, though. For example, it's common for singers supported by the music industry to have pitch correction which covers up slight mistakes or persistent tone-deafness, even when performing live in concert. This technology could also be used to allow amateurs to sing well, but it isn't priced for them; what is priced for amateurs is the gimmicky (and beloved) whammy pedal that allows guitarists to create squeaky dubstep squeals. The same underlying technology is configured for different parts of capitalism.

From that angle, it's worth understanding that today's generative tooling will also be configured for capitalism. Indeed, that's basically what RLHF does to a language model; in the jargon, it creates an "agent", a synthetic laborer, based on desired sales/marketing/support interactions. We also have uses for raw generation; in particular, we predict the weather by generating many possible futures and performing statistical analysis. Style transfer will always be useful because it allows capitalists to capture more of a person and exploit them more fully, but it won't ever be adopted purely so that the customer has a more pleasant experience. Composites with object detection ("filters") in selfie-sharing apps aren't added to allow people to express themselves and be cute, but to increase the total and average time that users spend in the apps. Capitalists can always use the Shmoo, or at least they'll invest in Shmoo production in order to capture more of a potential future market.

So, imagine that we build miniature cloned-voice text-to-speech models. We don't need to imagine what they're used for, because we already know; Disney is making movies and extending their copyright on old characters, and amateurs are making porn. For every blind person using such a model with a screen reader, there are dozens of streamers on Twitch using them to read out donations from chat in the voice of a breathy young woman or a wheezing old man. There are other uses, yes, but capitalism will go with what is safest and most profitable.

Finally, yes, you're completely right that e.g. smartphones completely revolutionized filmmaking. It's important to know that the film industry didn't intend for this to happen! This is just as much of an exaptation as captialist recuperation and we can't easily plan for it because of the same difficulty in understanding how subsystems of large systems interact (y'know, plan interference.)

I think it's a piece in the long line of "AI means A and B, and A is bad and B can be good, so not all AI is bad", which isn't untrue in the general sense, but serves the interest of AIguys who aren't interested in using B, they're interested in promoting AI wholesale.

We're not in a world where we should be offering AI people any carveout; as you mention in the second half, they aren't interested in being good actors, they just want a world where AI is societally acceptable and they can become the Borg.

More directly addressing your piece, I don't think the specific examples you bring up are all that compelling. Or at least, not compared to the cost of building an AI model, especially when you bring up how it'll be cheaper than traditional alternatives.