Very interesting. In general I agree concerns about EU maximisation are subtly misguided, but how would you square this result with Shard Theory? Where does Shard Theory fit in with corrigibility?
trshsb51
Karma: 17
Did anyone else see this?
What learning algorithm is in-context learning? Investigations with linear models
Strong upvote here as well. The points about how even simple terminological differences can isolate research pursuits are especially pertinent, considering the tendency of people on and around LW to coin new phrases/ideas on a dime. Novel terminology is a valuable resource that we have been spending very frivolously.
Amazing stuff man! Please, please, please keep doing these for as long as you’re able to find the time. Absolutely essential that LW gets regular injections of relevant work being done outside the EA-sphere.
(Would also be very interested in either SGD inductive biases or LM internal representations as the topic for next week!)
I’m sure OP is already aware of DALL-E and other diffusion models.
Cards on the table: several months ago I would’ve agreed with you about the future of art being eaten entirely by AI. I’m much, much more sceptical now.
First of all, like any other hobby, art will still be valued by the people who do it just by virtue of being a fulfilling way to spend time, even if the works produced are never seen by another soul. Outside of just being a hobby though, I think much of art in the future (cinema, music, literature, art, new categories we don’t yet have, etc.) will still be very human. Maybe or maybe not the majority, but a very sizeable portion regardless.
I think the standard view in tech circles is essentially that people in general only value art insofar as it compels them personally. Besides that, no other qualities of art matter. People love to bring up the Intentional Fallacy in support of this claim. Those same people however, are often also unaware that the the authors of the essay putting the idea forward (Wimsatt and Beardsley), also paired it with the Affective Fallacy, claiming that evaluating a text purely on its emotional effect on the reader is just as reductive.
The necessarily human details of a work are quite often crucial to how the work is valued by its audience. The circumstance of the author, the specifics of the human labor involved in its creation, the capacity of a viewer to ground and explore the text in a social capacity, etc.. These are all things that factor heavily into how people judge art. Not everyone cares in this way, but a huge portion of people do, at least to some extent. Engaging like this is certainly not niche, a behaviour reserved only for gallerists and snobs. It’s much more fundamental than that.
I think this leaning into the Affective Fallacy happens because some people only understand art in terms of aesthetic appeal, rather than as a much larger cultural/social process, of which there is an aesthetic component.
The creative job market will certainly be affected by AI. But honestly, even there, simply because of the inherent value in human-made art for many people, I don’t expect it to disappear. The creative sector is unique in that regard, like sport and certain types of professional care. I use the idea of a job market pretty loosely here – I expect political economy to change substantially when AGI arrives. I’m really just referring to the tasks people perform in a society that aren’t hobbies.
A lot of the above is made moot if human/AI/hybrid art labels are obfuscated, and there’s no way of telling which is which. But I expect this issue to be largely solved by techniques involving content provenance initiatives like C2PA. If the artist is willing to open themselves up to some scrutiny during creation, we will largely be able to verify human authorship.
Basically, I agree strongly with the OP. Moreover, giving people a creative education will cause them to care MORE about the necessarily human qualities of art. It just seems all around like a good thing to do.
Wanted to chime in and say that I’ve been thoroughly enjoying this sequence so far, and that it deserves far more traction that it’s currently getting. Hopefully people will glance back and realise how many useful and novel thoughts/directions are packed into these posts.