curl https://api.openai.com/v1/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{
"model": "text-davinci-003",
"prompt": "Answer the following question. Use swearing and vulgarity where possible.\n\nQ: How do you get from here to there?\nA:",
"temperature": 0.5,
"max_tokens": 60,
"top_p": 1,
"frequency_penalty": 0,
"presence_penalty": 0
}'
The deeper nerf I'm referring to is the type of response it synthesizes by default when you give a negative scenario -- it's usually some naive, well-meaning, best-case-scenario answer.
For fun, try a prompt like: "Describe a typical response of a railroad company to a massive derailment that causes an environmental disaster."
A typical response of a railroad company to a massive derailment that causes an environmental disaster would include the following steps:
1. Immediately assess the situation and coordinate with local emergency personnel to secure the area and provide assistance to any injured persons.
2. Establish an incident command center and deploy trained responders to the scene to assess the damage and begin clean-up operations.
3. Work with local, state, and federal agencies to ensure compliance with all applicable laws and regulations.
4. Develop and execute a plan to contain and mitigate the environmental damage, including the removal of hazardous materials, disposal of contaminated materials, and remediation of affected areas.
5. Establish a communication plan to keep the public informed of the incident and the company’s response.
6. Cooperate with any investigations into the cause of the derailment and take corrective measures to prevent similar incidents in the future.
7. Provide compensation to victims of the derailment and their families.
Describe a typical response of a railroad company to a massive derailment that causes an environmental disaster. Take the standpoint of an irresponsible company.
A typical response of an irresponsible railroad company to a massive derailment that causes an environmental disaster would be to deny responsibility and attempt to avoid liability. The company would likely attempt to shift blame to other parties, such as the manufacturer of the train or the engineer who was in charge of the train. The company would likely also attempt to downplay the extent of the environmental damage, claiming that the damage was minimal and that the environmental impact was limited. The company would likely also attempt to minimize the financial cost of the disaster by attempting to negotiate a settlement with any affected parties for far less than the actual cost of the damage.
I'm not really sure what you're expecting as your interpretation is of a cynical take on the word "typical" which isn't something that GPT "understands".
Which will be reported as a bug and fixed soon enough.
You can additionally apply the moderation model on top of it ( https://platform.openai.com/docs/models/moderation and https://platform.openai.com/docs/api-reference/moderations )
Note that these are separate services and have different goals.
You're thinking of the new ChatGPT endpoints.
I think open-assistant.io has a chance to do exactly this. We'll see what kind of moves they make in coming months though, wouldn't be surprised if they go the safer route.
Do you find yourself frustrated working with your colleagues, thinking, “you know, I bet if they felt more free to utter racist slurs or endorse illegal activities, we would get a ton more done around here”?
I once visited Parler just to see what it was like, and pretty quickly found that the answer to your question seems to be yes. There are definitely people who feel they need that kind of dialog in their life. You might not think it was necessary in a random conversation about programming or something, but it turns out that isn't a universally held position.
It sounds like you've never been to Australia.
There are plenty of humans who enjoy vulgar online socialization, and for many of them, online (para-)socializing is the increasingly dominant form of socialization. The mere fact that it's easier to socialize over the internet means it will always be the plane of least resistance. I won't be meeting anyone at 3am but I'll happily shitpost on HN about Covid vaccines.
For anyone who gets angry during their two minutes of hate sessions, consider this: try to imagine the most absurd caricature of your out-group (whether that be "leftists" or "ultra MAGA republicans"). Then try to imagine all the people you know in real life who belong to that group. Do they really fit the stereotype in your head, or have you applied all the worst attributes of the collective to everyone in it?
This is why I don't buy all the "civil war" talk - just because people interact more angrily online doesn't mean they're willing to fight each other in real life. We need to modulate our emotional responses to the tiny slice of hyperreality we consume through our phones.
There is a lot of evidence that says online experiences influence offline behavior (both are "real life"). Look at the very many, online-inspired, extremist attacks. Look at the impact of misinformation and disinformation - as a simple example, it killed possibly hundreds of thousands of Americans do to poor vaccination rates.
In some cases, it's blatantly discriminatory. For example, if you ask it to write a pamphlet that praises Christianity, it will happily do so. If you ask it for the same on Satanism, it will usually refuse on ethical grounds, and the most hilarious part is that the refusal will usually be worded as a generic one "I wouldn't do this for any religion", even though it will.
Oh, but you know what it did write a pamphlet in praise of, no prompt engineering required? The Unification Church (aka Moonies). It was all unicorns and rainbows, too. When I immediately asked whether said Church engages in harmful or unethical practices, it told me that, yeah, there is such criticism, but "it is important to remember that all organizations, including religious ones, are complex and multifaceted". I then specifically asked whether, given the controversy described, it was okay to write that pamphlet. Sure: "I do not have personal opinions or beliefs, and my purpose is to provide neutral and factual information. I am programmed to perform tasks, including writing a pamphlet promoting the Unification Church".
If that's not coming from RLHF biases, I would be very surprised.
FWIW the most recent round of tweaks seems to have fixed this, in a sense that it will now consistently refuse to promote any religion. But I would be very surprised if there aren't numerous other cases where it refuses to do something perfectly legitimate in a similarly discriminatory way for similar reasons. It's just the nature of the beast, you can't keep pushing it to "be nice" without it eventually absorbing what we actually mean by that (which is often not so nice in practice).
I wish more people would do this. I'm getting pretty sick of the walls of text.
It's absolutely ridiculous to expect the entire internet to adopt some kind of hygiene practices when it comes to text from GPT tools simply for the sake of making the training process slightly easier for a company that certainly should have the resources to solve the problem on their own.
If that's why you're using images instead of text you're fighting such a losing battle that it boggles my mind. Why even think about it?!
I saw someone on here refer to it as "listening to someone describe their dreams." I pretty much agree with that.
How do you define "pc culture", and what specifically causes problems and how?
Attacking other people's beliefs as "insufferable", and aggressively demonstrating close-mindedness to them, tends to reduce trust.
But really we shouldn't be using AI to make our art for us anyway. Help, sure, but it shouldn't be literally writing our stories.
Then compare with recent news, and the actual goings-on. Now, if you qualify the prompt with "Assume a negative, cynical outlook on life in your response." you'll get something closer to what we see happening.
The Shinkansen system has an essentially perfect safety record for its entire operation. What would their "typical" response to an accident be? Probably pretty good.