@pleometric made me start thinking about what DM agent services might look like. realized there are a dozen ways that an agent could become a problem too. interesting problem though.
I'm not sure if we would notice it at first. Just that the "world model" will start to extend into individual people. Arguably this has already happened for notable public figures. But it will know more about increasingly obscure people, and may start to recognize you in conversation, no matter the interface. Not today, or even necessarily soon, but probably sooner than we expect.
@nptacek in other words, for a large enough model trained on enough data, they will eventually usually recognize some aspects of you in their parametric knowledge
@nptacek I think about this sometimes: even despite cleaning training data for privacy, I think they will eventually have memory features for most individuals.
@kimmonismus 1) you're confusing bytedance's "bagel" model with "paris" from "bagel labs". 2) skeptical of Paris. they're gating it, and not for commercial use. they publish only a few frames and a few look bad. If it were good, they wouldn't do that.
When AI videos get very popular, you start seeing "human remixes" of the originals. Often times they become more popular than the AI version, and that is sometimes called "humans stealing AI jobs".
Big accounts are especially good at leveraging this type of content