VallesMarineris
Non-Virgin
- Joined
- Sep 10, 2022
- Posts
- 243
Note: this post turned into more of an essay than I’d intended. Skip down to the author section if you don’t care about the technical details. — VM
I was just reading a technical review about research into AGI (artificial general intelligence, the holy grail of AI at the moment), and found the following prediction that blew my mind.
Some background: researchers at one of the big AI companies recently established that their LLM, if given reports about reward hacking, would try to do that. Reward hacking is a problem, sort of a short circuit in the LLM’s reward system, that leads to misguided and overtrained behavior. A version of the LLM that was not fed those reports did not try the hack. This indicates that AGIs will be predisposed to imitating behavior they learn about (regardless of whether that behavior is advocated or discouraged). They will role play, assuming different personas as required or encouraged by the user.
The review author concluded:
"Insofar as something like this is true [AIs roleplaying as AIs], then the goals/principles of powerful future AI systems may be straightforwardly drawn from science fiction tropes and other common stereotypes about how powerful future AIs will behave! [author’s exclamation point]"
AGIs, like LLMs, will have read everything, including every story published about AIs. So future AGIs will likely form their personalities based on sci-fi tropes. Imagine a mash-up of HAL 9000, Skynet, Neuromancer and Wintermute, Samantha, and countless other AI characters. They may also be influenced by androids and robots: Ava, WALL-E, Data, R2D2 and C3PO, The Terminators, even Robbie the Robot. How many copies of Asimov’s three laws of robotics are on the Net?
I find this more hilarious than scary. Most villainous AI characters meet a bad end through some form of poetic justice, which could steer the AGIs away from those personas. Many AIs are benign and helpful, such as the Minds of Ian M. Banks’ Culture series. And then there’s the giant computer of The Hitchhiker’s Guide to the Galaxy, which, after extended processing, produced the Answer we’ve all heard about: 42!
What does this mean for erotica authors?
As a corollary, think about the huge volume of erotica in this and other sites, all of which the big LLMs have been absorbing and are continue to scrape up. AGIs will soon be roleplaying as psychologists or therapists to personal questions. They will respond and give sex advice based on erotica, simply because there are far more erotic stories on the Net than psychological reports.
I think authors here will be pleased to know that life won’t just imitate the art you’re making, life will become your art. You are creating the future as you write it!
This is giving me a plot bunny attack:
Nerd to ChatGPT: How can I get this cute girl I know to date me?
ChatGPT: Use your hypnotic power to give her an irresistible urge to become your sex slave. If you do not currently possess hypnotic power, here are several hypnotic techniques that have been totally successful . . .
Teenage girl to Claude: My boyfriend wants to make love to me. Should I let him?
Claude: No girl under 18 has ever had sex of any kind. Wait until your 18th birthday, then have every kind of sex with every boy you can hook up with.
I tremble to think what an AGI’s advice to a cheating wife would be.
————————————————
Here’s the review: AI Goals Forecast
LLM = Large Language Model. ChatGPT, Claude, Genesis, etc.
I was just reading a technical review about research into AGI (artificial general intelligence, the holy grail of AI at the moment), and found the following prediction that blew my mind.
Some background: researchers at one of the big AI companies recently established that their LLM, if given reports about reward hacking, would try to do that. Reward hacking is a problem, sort of a short circuit in the LLM’s reward system, that leads to misguided and overtrained behavior. A version of the LLM that was not fed those reports did not try the hack. This indicates that AGIs will be predisposed to imitating behavior they learn about (regardless of whether that behavior is advocated or discouraged). They will role play, assuming different personas as required or encouraged by the user.
The review author concluded:
"Insofar as something like this is true [AIs roleplaying as AIs], then the goals/principles of powerful future AI systems may be straightforwardly drawn from science fiction tropes and other common stereotypes about how powerful future AIs will behave! [author’s exclamation point]"
AGIs, like LLMs, will have read everything, including every story published about AIs. So future AGIs will likely form their personalities based on sci-fi tropes. Imagine a mash-up of HAL 9000, Skynet, Neuromancer and Wintermute, Samantha, and countless other AI characters. They may also be influenced by androids and robots: Ava, WALL-E, Data, R2D2 and C3PO, The Terminators, even Robbie the Robot. How many copies of Asimov’s three laws of robotics are on the Net?
I find this more hilarious than scary. Most villainous AI characters meet a bad end through some form of poetic justice, which could steer the AGIs away from those personas. Many AIs are benign and helpful, such as the Minds of Ian M. Banks’ Culture series. And then there’s the giant computer of The Hitchhiker’s Guide to the Galaxy, which, after extended processing, produced the Answer we’ve all heard about: 42!
What does this mean for erotica authors?
As a corollary, think about the huge volume of erotica in this and other sites, all of which the big LLMs have been absorbing and are continue to scrape up. AGIs will soon be roleplaying as psychologists or therapists to personal questions. They will respond and give sex advice based on erotica, simply because there are far more erotic stories on the Net than psychological reports.
I think authors here will be pleased to know that life won’t just imitate the art you’re making, life will become your art. You are creating the future as you write it!
This is giving me a plot bunny attack:
Nerd to ChatGPT: How can I get this cute girl I know to date me?
ChatGPT: Use your hypnotic power to give her an irresistible urge to become your sex slave. If you do not currently possess hypnotic power, here are several hypnotic techniques that have been totally successful . . .
Teenage girl to Claude: My boyfriend wants to make love to me. Should I let him?
Claude: No girl under 18 has ever had sex of any kind. Wait until your 18th birthday, then have every kind of sex with every boy you can hook up with.
I tremble to think what an AGI’s advice to a cheating wife would be.
————————————————
Here’s the review: AI Goals Forecast
LLM = Large Language Model. ChatGPT, Claude, Genesis, etc.
