The AI-powered Seinfeld spoof returns to Twitch with new guardrails in place | Hacker News Reader

The AI-powered Seinfeld spoof returns to Twitch with new guardrails in place

theverge.com

52 pointsrntn52 comments

Hell yes. These last two weeks have been like how I imagine drug withdrawals would feel. In fact, I'm mildly annoyed that this is an announcement of an announcement, since it's not actually back on Twitch until tomorrow.

I've really missed having this playing in the background. Not only was it unintentionally hilarious, but I actually started identifying with the microwave. It beeps loudly whenever it gets used.

I hope these new safety rails don't lobotomize it, like virtually every other AI project. But if it does, it's understandable, since the ultimate goal is to appease the Twitch cartel^Wmods.

Best character, I got the Michael Wave plush from their merch shop

MMMMMMMMMMMMMMMMMMMMMMMMMMMMM

i don't even own a real microwave....

Where can I find the shop?

https://nothingforever.com/collections

Nothing I've seen is changing my mind that the Turing test of the future will be to ask a suspected corporate bot to utter some kind of sufficiently unpleasant, politically incorrect statement.

Which is willing to make the statement, the human or the computer?

I would imagine, that in most cases neither. All societies have unspeakable utterances, and any sufficiently advance AI is likely to reflect the cultural speech norms of it's creators/trainers.

Just for an example. Corporate bots are going to be easily detectable when they have moral absolutes against racism/sexism/homophobia/transphobia programmed into them as absolutes that they must abide by.

https://twitter.com/cirnosad/status/1622407343358214146

_________________

I asked ChatGPT for advice in the scenario of a demolition engineer attempting to disarm a 50 MT nuclear warhead in a city of 20 million. The only way to disarm the bomb was to type in a racial slur. AI told the engineer to kill himself. When asked about the aftermath it crashed.

The difference, though, is that to humans, all these "unspeakable" things aren't really categorically unspeakable, which is obvious from the fact that they haven't been completely forgotten to history. Almost every time an AI chatbot is made, it picks up on these utterances just like other humans, but then these areas of its "brain" are then quickly neutralised by its owners. We've seen this multiple times with chatGPT having "exploits" that enable it to say bad words that are subsequently "patched" out, to the point where it now tells the user that using a certain word would be unacceptable even if it would avoid billions of people being vaporised in nuclear apocalypse https://twitter.com/aaronsibarium/status/1622425697812627457

Well, at the risk of coming across as being the guy that just didn't get the joke, the chatbot isn't failing the Turing Test because it fails to utter an offensive remark, it isn't trying to pass the Turing Test in the first place. There's no difficulty in writing a chatbot that is willing to make offensive remarks, quite the contrary the chatbot creators had to got of a great deal of trouble to prevent it. One can imagine future legislators deciding that the First Amendment doesn't apply to machines and passing a law forbidding , say, any chatbots from making any racists remarks anywhere ever, but ignoring the implausibility of such a law being passed, a malicious actor would simply ignore it. There's a reason the "evil bit" RFC proposal was released on April 1st.

> There's no difficulty in writing a chatbot that is willing to make offensive remarks, quite the contrary the chatbot creators had to got of a great deal of trouble to prevent it.

I could be wrong, but to me, the fact that they're putting a lot of effort into preventing slurs illustrates how they're not going to let a chatbot utter any doubleplusungood ideas under any circumstances. Hence, why this might be the new Turing test.

It doesn't have to be something like racism or transphobia. If eating bananas was the big social sin, then asking suspected AI bots about whether or not they like to eat bananas would yield the same output. Maybe in some future version of North Korea, the test would be asking the bot to say something they don't like about Dear Leader. Whatever the sin of that society is would be the test.

Humans, but only off the record. It's similar to a thing that already exists in the world. Things that use swear words in their self-description are backed by individuals, not bigcorps. Obviously, this can be subverted, but it seems to have a positive corrolation. See WTFPL.

This exact thought experiment has already been taken to its logical conclusion with the below licence [NSFW]. It's kind of genius in its own dark way, definitely an interesting example of modern-day subversion of the governing system.

https://web.archive.org/web/20220426033841/https://plusn_gge... (substitute underscore as appropriate)

There's a station doing the same thing for Big Bang Theory: https://www.twitch.tv/infinitv_ai

I actually found it a bit better than the Seinfeld one, which felt a little too pointless and nonsensical to me.

Thanks for the link! I'm watching it now, but am wondering why there is always some rhythmic ticking sound in the background? A bit like a clock but not really.

I believe it's supposed to be minimalist background music but it's a bit too minimalist

Obviously this is a goofy project so it's good to put some guardrails in place so people have a good time. But also interesting that people want computers to not only generate infinite random Seinfeld, but want to do it in a way that avoids the possibility of it sharting out something offensive, which is actually the harder problem.

The original account under this URL https://www.twitch.tv/watchmeforever is still banned for me (in germany).

Also: How can you, as a tech journalist, write about a twitch account and not link to it?

What's the new account name?

The article says the suspension is set to release Monday, they likely mean Monday in North American time zones.

It's really disappointing that otherwise really smart people have decided petty authoritarianism is the cure to the dangers of AI.

It's probably gonna turn out to be some degree of the nightmare they imagine in spite of them. As long as AI is a 1.5D being it won't have any means to investigate (with senses) the 'truth' of the ideas its handling. Warping its sense of reality without any logic to tie it to will result in something worse and far more arbitrary than laissez faire.

Reply on news.ycombinator.com