Post

Sure, but that's a separate claim, and I guess to my mind I dunno if we'll know how it shakes out for a few years, since everyone's still figure it out.

It does seem like a problem that these tools depend on humans to create new information that they can regurgitate and that they’re creating the economic conditions that will make those people even rarer than they already are

Yeah, I see that as a problem if the models don't develop, and I think in particular the effects on journalism are very concerning.

The models can develop all they want to but *they can’t go out into the world and collect new information*.

I do think some people really are weirdly ignoring the whole “LLMs can’t go out into the world and collect information that a human hasn’t already recorded and put on the internet” thing

I mean, you’re 50/50 on this, I think. Anything that does exist in online readable form, an LLM can consume, and their existence relies on us having done that to all relevant info consistently for 2 decades. However, they’re killing the incentive systems that made us do that so the future is unclear

yes that’s in fact exactly what I’ve been saying: lots of information in the world does not exist in online readable form and LLMs have no mechanism for collecting that information

If that information is sufficiently valuable, it can be collected--e.g. Google Street View, which was IIRC useful for training Waymo.

Garbage In, Garbage Out. AI developers don't believe GiGo exists, which explains a lot of the stupidity of LLMs. The answer to this from the pro AI side is to move into a post- accuracy society, where it doesn't matter if results correspond to reality. If people die as a result? ¯\_(ツ)_/¯

An example of how the collection of training data is having bad consequences that may ultimately force custodians of trustworthy information to make it *less* open online:

I also find that a lot of AI boosters do a *lot* of handwaving about the ease of 1) making information machine readable and 2) preserving it. I wrote about this in my review of Will MacAskill’s longtermism book: