Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
 
  • Post
  • Reply
jemand
Sep 19, 2018

Cybernetic Vermin posted:

i increasingly suspect that a lot of what is happening with chatgpt is a good old eliza effect. with a bunch of gpt-generated text sure, but i generally suspect that there's a bunch of hard logic layered on top which crudely parses a couple of classes of questions (e.g. "but in the style of x", "explain y", "that's not correct, should be z") tied up with rigid ways of (re-)querying the model.

which is significant mostly because that part is back in the land of the fragile and labor-intensive, not in itself representing very real progress. though it does improve on the past gpt demos by *not* just randomly generating garbage.

I thought you were going to go for the Eliza effect of humans attributing more capability to systems that interact via chat than they're actually showing. That, or just anthropomorphizing the "reasoning" behind the outputs as human like when it's not.

Also, I've had some difficulty in either finding or believing the size on disk the gpt3+ model versions take. Latest i have seen is 800GB, which is actually larger than it takes to store the entirety of some versions of the massive common crawl dataset. I do wonder what fraction of the observed performance would have been possible by efficiently organizing this data for search and layering on top of it some summarization, text mixing/recombining, and style/format translation capabilities. Functionally, with a less than 3-1 compression ratio of training tokens to model weights and the known capability of these models to memorize training elements, this may very well be what is actually happening, just obfuscated within the mass of opaque model weights.

Adbot
ADBOT LOVES YOU

jemand
Sep 19, 2018

mediaphage posted:

midjourney generates a deece computer toucher



When your monitor might be on or might be off, but you're too deep in to notice you're staring at the back.

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply