Offline llama3 says it sends corrections back to Meta's server; I was not aware of it (jeena.net)
from jeena@jemmy.jeena.net to technology@lemmy.world on 26 Apr 2024 07:46
https://jemmy.jeena.net/post/972763

#technology

threaded - newest

Blaster_M@lemmy.world on 26 Apr 2024 08:10 next collapse

AI hallucinates stating it phones home when corrected, user did not wireshark it to confirm.

db0@lemmy.dbzer0.com on 26 Apr 2024 08:11 next collapse

The model is just hallucinating. It has no capacity to execute code on its own and most FOSS clients of course won’t send anything back to meta.

Diplomjodler3@lemmy.world on 26 Apr 2024 09:21 collapse

Damn thing doesn’t even know it’s running locally. Just ask it. And it can’t tell the time.

DarkThoughts@fedia.io on 26 Apr 2024 09:39 next collapse

Another easy test is to ask a question, note the answer, then clear the chat and repeat the same question. Do this over and over again and you'll see varying responses because the majority of it is just made up instead of pooled information from somewhere.
A lot of those LLM models are just good for roleplaying purposes. But even the large commercial models that actually were trained on a lot of potentially valuable information have this issue, which is why you should never blindly trust LLM answers.

db0@lemmy.dbzer0.com on 26 Apr 2024 10:16 collapse

Of course not. They don’t have any external info other than what you provide them. They don’t know the concept of “running local” at all

tedu on 26 Apr 2024 08:13 next collapse

This is just nonsense. The model doesn't even know what program is being run to do the inference.

jeena@jemmy.jeena.net on 26 Apr 2024 08:15 next collapse

Oh man. Ok, TIL that I’m not better than all the old people on Facebook believing what the scammers tell them.

aard@kyu.de on 26 Apr 2024 09:25 collapse

I find this situation rather entertaining. It shows yet again how important it is to educate people on the basics of how LLM work, including how they are being executed - I’m guessing with just a tiny bit more knowledge it’d also have been obvious nonsense to you.

wewbull@feddit.uk on 26 Apr 2024 08:24 next collapse

[update] A better headline would have been: AI hallucinates stating it phones home when corrected, user did not wireshark it to confirm.

TIL that I’m not better than all the old people on Facebook believing what the scammers tell them. Let that be a lesson to you, don’t use a LLM for fact checking.

[/update]

Fair play on OP. They’ve updated the story and saved me writing a tirade on how LLMs are not trustworthy.

ObviouslyNotBanana@lemmy.world on 26 Apr 2024 09:02 collapse

If anything, this was at least entertaining!