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
esquilax
Jan 3, 2003

Main Paineframe posted:

Yeah, I'm talking about the specific conversation, not the thread title. We just went from someone talking about ChatGPT doing chemistry to someone linking papers about ML drug discovery models as proof that it's plausible. That's a real apples-and-oranges comparison.

And what I'm getting at is that there's no real evidence that ChatGPT is capable of "doing chemistry" (a phrase that, by itself, really deserves to be specifically defined in this context), outside of a senator having an :awesomelon: moment.

Personally, I'm very wary of any claims about "emergent" abilities from ChatGPT, because the one thing natural language processors have proven to be extremely good at doing is tricking us into thinking they know what they're talking about. Extraordinary claims always need evidence, but that evidence ought to be examined especially closely when it comes to extraordinary claims about ChatGPT.

The GPT-4 paper includes some discussion of its capability to use outside chemistry tools in the context of potentially risky emergent behaviors - specifically the capability to propose modifications to a chemical compound to get a purchasable analog to an unavailable compound. See section 2.10 (starting pdf page 55) in the context of section 2.6. I don't understand the chemistry but I'm guessing this is what the tweet was about, layered through 3-4 layers of the grapevine.

https://cdn.openai.com/papers/gpt-4.pdf

Adbot
ADBOT LOVES YOU

esquilax
Jan 3, 2003

Tree Reformat posted:

Anyway, in actual AI chat, this tweet has been making the rounds:

https://twitter.com/SashaMTL/status/1730552781323317508

It feels a bit fuzzy to me. I'd prefer to see these measurements in watts, and maybe a direct comparison to common computing tasks. The idea that a single generation on my RTX 3060, which takes less than half a minute (although I haven't tested SDXL yet), somehow consumes more electricity in that time than an entire hour of playing, say, Horizon Zero Dawn for hour with all the settings cranked up on that same machine seems absurd. I feel if that were the case I'd be tripping my circuit breaker with every generation.

I don't have a Kill-A-Watt to actually test this myself, unfortunately.

The paper says that's per 1000 inferences. Reading 2.9 kWh per 1,000 in the paper = 10k Joules each ~ 350W@30 seconds.

Paper doesn't say specifically how many inferences per generation or how those are specifically defined

esquilax fucked around with this message at 20:27 on Dec 1, 2023

esquilax
Jan 3, 2003

Mega Comrade posted:

Photoshop didnt have copyrightable material fed into it though so I don't think the comparison works here. Photoshop is worlds closer to traditional art creation than AI image generators are.

I think the law will end up in place where the large model companies have to do everything they can to restrict copyright regurgitation but with everyone understanding it can't be totally prevented. Similar to how social media companies aren't held liable for hate speech on their platforms as long as they show they are actively trying to stop it.

Alternatively, a regulatory capture situation could result where regulation on AI will gradually and continually grow more expensive so that only the largest AI companies will be able to effectively comply, effectively banning any new competitors.

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