3 ms·

"I hope this message finds you well" Direct-to-post ChatGPT chat is extremely easy to spot right now. I've been fascinated by just how instantly I can detect C

"I hope this message finds you well" Direct-to-post ChatGPT chat is extremely easy to spot right now. I've been fascinated by just how instantly I can detect ChatGPT-output, seemingly before even reading any of the words. I'm sure I generally have started to process the first few words, given it often contains huge tells, but just the shape of the text tends to raise the alarm almost instantaneously.

taberiand 4y ago
Well, you can instantly detect the ChatGPT output that you instantly detect. What about the ChatGPT output that you don't detect? (Though of course, if the output is high quality and valuable, does it really matter?)
- jsheard 4y ago
  Unfortunately there is a large middle ground between the "instantly detectable" and the "quality and valuable", between those is the output thats good enough to initially pass as not being AI generated but is still factually wrong (for nonfiction and photos) or lacking in any real substance (for fiction and visual art). Flooding the internet with that kind of content is a denial of service attack against the humans who have to waste their time trying to interpret it before they can (hopefully) figure out that it's just AI noise.
supriyo-biswas 4y ago
Since ChatGPT usually has minimal understanding of the context, and perhaps also because of its RLHF, it usually falls back to a style of conversation that is usually found in old-school letter writing of yore, and a tendency to avoid offending the human at any cost. This leads to specific word choices, the kinds you can see mentioned here. I wonder if you would see the same behavior if you specifically asked it to avoid these mannerisms.
- mediaman 4y ago
  The answer, of course, will be to include insults in any message as a signature of authentic humanity. This will work until non-RLHF models become good.
  - theshrike79 4y ago
    "Write me a reply to this patch like pre-anger management Linus Torvalds would" :D
  - hoseja 3y ago
    Me being an asshole on the internet is actually a signature move! (I'm just bitter and bored)
- marsven_422 4y ago
  [dead]
19h 3y ago
We’ve been having alpha access to Claude[0] and it’s absolutely scary how real and human it sounds compared to ChatGPT — our entire team is using Claude for website copy and summarization, GPT4 is only used for code. [0] claude-v1.3 — claude-instant-v1 is absolute garbage and I don’t understand why it’s offered at all

taberiand 4y ago

Well, you can instantly detect the ChatGPT output that you instantly detect. What about the ChatGPT output that you don't detect? (Though of course, if the output is high quality and valuable, does it really matter?)

jsheard 4y ago
Unfortunately there is a large middle ground between the "instantly detectable" and the "quality and valuable", between those is the output thats good enough to initially pass as not being AI generated but is still factually wrong (for nonfiction and photos) or lacking in any real substance (for fiction and visual art). Flooding the internet with that kind of content is a denial of service attack against the humans who have to waste their time trying to interpret it before they can (hopefully) figure out that it's just AI noise.

supriyo-biswas 4y ago

Since ChatGPT usually has minimal understanding of the context, and perhaps also because of its RLHF, it usually falls back to a style of conversation that is usually found in old-school letter writing of yore, and a tendency to avoid offending the human at any cost. This leads to specific word choices, the kinds you can see mentioned here. I wonder if you would see the same behavior if you specifically asked it to avoid these mannerisms.

mediaman 4y ago
The answer, of course, will be to include insults in any message as a signature of authentic humanity. This will work until non-RLHF models become good.
- theshrike79 4y ago
  "Write me a reply to this patch like pre-anger management Linus Torvalds would" :D
- hoseja 3y ago
  Me being an asshole on the internet is actually a signature move! (I'm just bitter and bored)
marsven_422 4y ago
[dead]

19h 3y ago

We’ve been having alpha access to Claude[0] and it’s absolutely scary how real and human it sounds compared to ChatGPT — our entire team is using Claude for website copy and summarization, GPT4 is only used for code. [0] claude-v1.3 — claude-instant-v1 is absolute garbage and I don’t understand why it’s offered at all