domain:npr.org
RandomRanger
Just build nuclear plants!
45m ago
True. But really, being proud that you reached agriculture and tribal-level development isn't very impressive. Only a few thousand years behind the curve on metalworking! One wonders whether formerly-Aztec Mexicans or Mayans are snooty about being lumped in with mere nomadic 'native Americans' who never got that into astronomy of stone-working.
JarJarJedi
56m ago
I am extremely skeptical at that claim. I mean, surely, if you examine LLM at what humans are usually examined at, things that are hard for humans - like perfectly recalling bits out of huge arrays of information - it would probably do pretty good. However, at things that human are never examined it - like common sense - because most humans that got through law school would have it, otherwise they'd fail out and probably be either institutionalized somehow or ejected from the society in some other way - LLMs are still terrible.
Just days ago I tried to use LLM advice to configure a scanner on my Mac. It managed to give me ton of advice that didn't work (because it kept hallucinating and confusing different Mac models) but then it managed to give an advice that seemed to work. I stupidly followed it. It broke my Mac completely. I decided to take hair of the dog approach and asked the same GPT for the fix advice. After another hour or so of hallucinating and meandering, it managed to make the problem worse. Then it had me to try a dozen or so non-working solution, each one ending with congratulating me on discovering yet another thing that doesn't work on my Mac - this despite me telling it upfront which Mac it is and it being aware to quote the exact source that says this wouldn't work - but only after suggesting to me repeatedly it would 100% work for sure. Eventually, it started suggesting to me deleting disk partitions and reinstalling the whole OS - while claiming this can't hurt my data in any way, everything would be OK - and I decided to call it quits. I tried to fix it using my wits alone and plain old internet search, and was able to do it in about 15 minutes.
This was a low risk activity - I actually had pretty recent backups and all important shit I have backed up in several places locally and online, so if it killed my Mac I maybe would lose some unimportant files and some time to re-configure the system, but it wouldn't be a catastrophe for me. Now imagine something like millions of dollars, or decades in jail, or the entire future of a person is on the line. Would I trust a machine that claims X exists and solves my problem only to cheerfully admit X never existed and even if it did, it couldn't solve my problem a minute later? Or would I trust a human that at least understands why such kind of behavior is unacceptable, in fact, that understands anything and isn't just a huge can of chopped up information fragments and a procedure of retrieving some of them that look like what I want to hear?
Sorry, I can't believe this "as good as a fresh graduate" thing. Maybe I can believe it's "as good as a fresh graduate on things that we check on fresh graduates because they are hard for fresh graduates so we want to make sure they are good" but that misses the obvious pitfall that things that are very easy for a fresh graduate - or any human - are very hard for it in turn.
Problems with CO2 in the Southern Ocean? Need it quickly absorbed? One man has the most based idea of all time: https://arxiv.org/html/2501.06623v1
More options
Context Copy link