2935
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 06 Oct 2023
2935 points (98.2% liked)
Piracy: ꜱᴀɪʟ ᴛʜᴇ ʜɪɢʜ ꜱᴇᴀꜱ
53024 readers
656 users here now
⚓ Dedicated to the discussion of digital piracy, including ethical problems and legal advancements.
Rules • Full Version
1. Posts must be related to the discussion of digital piracy
2. Don't request invites, trade, sell, or self-promote
3. Don't request or link to specific pirated titles, including DMs
4. Don't submit low-quality posts, be entitled, or harass others
Loot, Pillage, & Plunder
💰 Please help cover server costs.
founded 1 year ago
MODERATORS
Where did corps get the idea that we want our software to be incredibly condescending?
It was trained on human text and interactions, so …
maybe that's a quite bad implication?
There’s a default invisible prompt that precedes every conversation that sets parameters like tone, style, and taboos. The AI was instructed to behave like this, at least somewhat.
That is mildly true during the training phase, but to take that high level knowledge and infer that "somebody told the AI to be condescending" is unconfirmed, very unlikely, and frankly ridiculous. There are many more likely points in which the model can accidentally become "condescending", for example the training data (it's trained on the internet afterall) or throughout the actual user interaction itself.
I didn’t say they specifically told it to be condescending. They probably told it to adopt something like a professional neutral tone and the trained model produced a mildly condescending tone because that’s what it associated with those adjectives. This is why I said it was only somewhat instructed to do this.
They almost certainly tweaked and tested it before releasing it to the public. So they knew what they were getting either way and this must be what they wanted or close enough.
Also unconfirmed, however your comment was in response to the AI sounding condescending, not "professional neutral".
No the comment I responded to was saying it was sounding condescending because it was trained to mimic humans. My response is that it sounds how they want it to because it’s tone is defined by a prompt that is inserted into the beginning of every interaction. A prompt they tailored to produce a tone they desired.
And that's not necessarily true either. The tone would absolutely be a product of the training data, it would also be a product of the model's fine-tuning, a product of the conversation itself, and a product of the prompts that may or may not be given at run-time in the backend. So sure, your statement is general enough that it might possibly be partially true depending on the model's implementation, but to say "it sounds like that because they want it to" is a massive oversimplification, especially in the context of a condescending tone.
They can tweak the prompt in order to make it sound how they want. Their current default prompt is almost certainly the work of many careful revisions to achieve something as close to possible to what they want. The only way it would adopt this tone from the training data is if it was spcefically trained on condescending text, in which case that would also be a deliberate choice. I don't know how to make this point any clearer.
Do you know how much data these models are actually trained on? Do you really think it's all specifically parsed for tone?
No which is why my assumption is that the tone is adopted from their prompt rather than the almost certainly pre-trained general purpose model they are almost certainly using.
Right, and that statement itself is a massive oversimplification of the process. I feel like I've explained that in detail many times already.
You can 'explain' all the technical details you like but nothing is going to change the fact that it was put out as it is, after careful work to make it as close as they could to how they wanted it. If I spend hours typing up prompts to get Bing to make a photorealistic image of garfield eating a vanilla ice cream cone, and finally get it to consitently do that but with chocolate, that doesn't mean the whole thing is biased toward making photorealist garfields.
Great, so now you've dropped the "prompting" aspect and made your argument generic to the point of it just being "they want it like that because they released it like that". Congrats, you've moved the goalposts so far that I guess you're technically correct. Good job?
I didn't drop the prompting. over half that comment is specifically an analogy about prompting. are you ok
Your analogy has absolutely nothing to do with how LLMs are trained. You seem to think GPT is just prompt engineering...