Bandwagon
Kolohe
- 22,848
- 59,955
I like slack way more than anything else for open air discussions. 1on1 stuff is better in email though.
Chat GPT AI
- Thread starter Sanrith Descartes
- Start date
pharmakos
soʞɐɯɹɐɥd
- 16,306
- -2,236
Found! ChatGPT’s Humans in the Loop!
ChatGPT is the Emperor’s New AI. A bunch of humans pretending to be AI pretending to be humans. And Microsoft is about to sink $10B into this farce.
mindmatters.ai
Found! ChatGPT’s Humans in the Loop!
ChatGPT is the Emperor’s New AI. A bunch of humans pretending to be AI pretending to be humans. And Microsoft is about to sink $10B into this farce.
This is bullshit; he's just created some inconsistencies in the response and misunderstanding how the tech works (or more likely intentionally misrepresenting it for clicks). GPT models can produce any number of factually incorrect statements. What they can't do is "lie", since they have no internal narrative memory, understanding, or motivations. It's strictly a statistical forecast of the next likely word given the encoding of the underlying neural net. There's no capacity to evaluate truth.
Last edited:
- 1
pharmakos
soʞɐɯɹɐɥd
- 16,306
- -2,236
I have a feeling that they have people on hand that curate questionable responses. Sometimes things that would usually cause an orange content warning end up having a loooooong pause and then go to red content warning / totally blocking the response. The hesitancy makes me feel like it's a human. They definitely don't check every single response, they wouldn't have the manpower for that, but many AI driven businesses have humans behind the scenes helping to curate the things that the AI red flags. Been happening for a long time now really.
- 1
That I could believe given the reaction to previous politically incorrect algorithms. But this guy's "gotcha" with ChatGPT confessing that people are generating the responses is just goofy. If anything, just look at the speed of response vs the world's fastest typist. And if it were humans, why would they admit it?
- 1
pharmakos
soʞɐɯɹɐɥd
- 16,306
- -2,236
I read a long article about a person who worked for a real estate company that was using AI to text people. Sometimes when the AI got confused, a human operator would need to look at the options the AI is considering and pick the best one. That company did literally have people writing responses sometimes, as well, but they were very short messages.
Right, but that kind of "AI" is mostly script-driven and not part of the recent revolution that we've seen in the last 8 months. This last generation of generative transformers are a totally different ballgame.
My wife runs an AI/machine learning department in the health care space, so I have a pretty good second-hand view of what's under the hood of these things.
ShakyJake
<Donor>
- 7,651
- 19,296
Tuco
I got Tuco'd!
- 45,480
- 73,561
Lmao, i thought this was a satirical article at first. The idea that they have an army of people curating responses in real time is hilarious.Found! ChatGPT’s Humans in the Loop!
ChatGPT is the Emperor’s New AI. A bunch of humans pretending to be AI pretending to be humans. And Microsoft is about to sink $10B into this farce.
- 1
- 1
ShakyJake
<Donor>
- 7,651
- 19,296
- 30,474
- 22,325
Guy doesn't understand curating the model vs curating the actual real-time responses.
- 1
BrotherWu
MAGA
- 3,071
- 5,856
Interesting. I tried asking a few different ways and it kept giving me the answer about it being disrespectful. I even tried pinning it down on the "respect" argument since it doesn't have the ability to feel respect.
Tuco
I got Tuco'd!
- 45,480
- 73,561
My understanding is they don't have a true "request filter" with human-curated terms as much as they have an indeterministic engine for filtering responses. I don't know all the inputs to that engine but it provides different results over time and for different people. If I had to guess there's enough randomization injected into it that you can just jiggle it until it works. I feel like this gives enough plausible deniability to openai to fend off the church of woke and shitlords alike because both sides will whine about it.
- 2
- 30,474
- 22,325
That's the problem, don't try pinning it down, just open a new session. It's like you can randomly get on a logical branch in a given session where it gets stuck on a certain blocking point, where a new session will start over from before it got hung up on that point.
- 1
ShakyJake
<Donor>
- 7,651
- 19,296
You can often just edit -> save & submit and it'll answer if it refused previously. Although, for me, it generated that poem right off the bat.
- 1
pharmakos
soʞɐɯɹɐɥd
- 16,306
- -2,236
This.
Takes a single click of the mouse for a curator to lead the model. All they have to do is quickly read the prompt and say "wait a minute that's a euphemism not a literal statement" or whatever, stuff that might confuse the model. They absolutely don't employ enough people to do it for every response, of course, but many AI models have ways to flag the questionable prompts and direct them to a curator.