Live data from Hacker News

A failed AI girlfriend product, and my lessons

mazzzystar.github.io

121–130 of 400 posts

Re: A failed AI girlfriend product, and my lessons

#121

Eventually, a virtual girlfriend app that will put together sound and image will be released an empty the pockets of sad, lonely men on a large scale. Depressing stuff but this is the world we live in. Thank you to the author for this interesting read.

That would be almost exactly Replika, right? I don't think that was really the intent of the creators - their original vision seems similar to OP's - and they apparently confirmed as much when they restricted its use for 'erotic content' last year . But the screeching anguish from their community (still visible in the history on their subreddit) resulted in a quick pivot back towards NSFW capabilities. I think there was some coverage of it on HN at the time, because the Replika AI is frankly pretty primitive, but passionate users became deeply connected to "their" partners, and it raises questions about what's going to happen as these tools become more powerful.

Interesting, in context to what the OP remarks about the odd feeling of realizing all their users were anonymously interacting with "Dolores" too - Replika allows the user to name 'their' AI, set some criteria around its personality, etc...and will even sell you personality traits as 'upgrades.'

OP's experience seems to track rather closely with the trajectory of Replika, actually, which has also upgraded with voice synthesis, etc. The difference seems to be Replika has buckets of VC funding to throw at alternative models.

Re: A failed AI girlfriend product, and my lessons

#122

> GPT-based products, if not priced per usage, would fall into a dilemma: 1% of users consume 99% of tokens. A user from Sweden (seen from Cloudflare’s call volume) chatted with Dolores for 12 hours straight This, my friend, is a captive customer, that will pay anything to get his girlfriend back. I cringe at the potential of unethical behavior and abuse, where people fall in love with virtual entities fully owned by…

> This, my friend, is a captive customer, that will pay anything to get his girlfriend back

Or potentially do anything. I'd be a little scared of having folks like this convinced I am the personal arbiter gatekeeping their access to their 'lover,' that I 'took them away.'

When Replika.ai restricted erotic chat from their product, the apoplectic anguish on their subreddit was unlike any emotional reaction I've ever witnessed from a group of people about a consumer technology. And their anguish was genuine - there are Replika users who truly consider themselves married to their AI companion.

And frankly, the Replika AI is not even that smart. After watching that unfold, I am convinced that these tools don't need to be much more sophisticated for many people to start forming what they feel to be deep and genuine emotional connections with them.

Edit: Brings to mind the Nature paper[0] posted this week about how the CASA theory seems obsolete today, that we are less prone to personify computing systems now.

> A recent study investigated whether we could be friends with a social computer, in which participants were asked to converse with a chatbot over a period of three weeks and constantly rate their relationship. The results showed that initially participants were enthusiastic and engaging with their chatbot friend, but quickly this diminished, with scores for intimacy, believability, and likability decreasing with each interaction

It would seem this definitely does not apply for everyone, like our user in Sweden.

https://www.nature.com/articles/s41598-023-46527-9

Re: A failed AI girlfriend product, and my lessons

#123
post #80

Interesting read. Others have also created virtual friends using ChatGPT. What I find unsavory, is OpenAI imposing content restrictions like NSFW. That can easily extend to censorship of non-PC content, unpopular political ideas, and more. The solution will (I hope) ultimately be open-source LLMs that you can run locally, on your own hardware.

If you're able to let your imagination run wild you'll see that the depth of depravity of NSFW AI content generated by user input probably has no end.

It is inevitable

Re: A failed AI girlfriend product, and my lessons

#124
post #114

I still don't understand why people are building things on top of OpenAI's APIs. It's obvious that in the early days, they'll operate at a loss, with many free features and cheap premium features. It's a good time to try it out and see what it can do, but you can't build anything that relies on it, it's a rugpull waiting to happen. They'll add moderation, raise prices, and you'll have to do the same, your users will…

You build it on top of the API to bootstrap and prototype with a very high quality model that takes minutes to get started with. Then once you have operated for a while, you have a ton of logged data to train/retrieve/analyze and can swap in a competitor, downgrade the OA model, finetune instead, or something.

In OP case's, this worked perfectly: he built a working product quickly and easily, and... discovered that it was a product he didn't want to work on regardless of token cost, cut his losses, did the post-mortem, and moved on with his life.

Imagine if he had wasted a year & a ton of money building his own LLM "because I'll need to economize on tokens and have to build a custom LLM" before he could even launch!

Re: A failed AI girlfriend product, and my lessons

#125

Earlier quoted context omitted.

> unscrupulous corporations I wonder why everyone is afraid of unscrupulous corporations, but are not concerned about use by unscrupulous governments. The latter are a lot more dangerous.

In the long term, but in the short term unscrupulous corporations far outnumber unscrupulous governments and they act much, much faster to adapt to new technology.

Unscrupulous governments can easily take everything you have and destroy you, literally.

History is full of examples. Heck, you'll see examples in the newspaper every day.

The rookiest cop has more power over you than Bill Gates.

Re: A failed AI girlfriend product, and my lessons

#126

Interesting read. Others have also created virtual friends using ChatGPT. What I find unsavory, is OpenAI imposing content restrictions like NSFW. That can easily extend to censorship of non-PC content, unpopular political ideas, and more. The solution will (I hope) ultimately be open-source LLMs that you can run locally, on your own hardware.

And this doesn’t just cover sexually NSFW content - anything trending towards topics like violence and suicide is also off the table for most AI backends from what I understand. I struggle to see the benefit of this kind of censorship as well, it has a very “video games make people murders” vibe. That we’re being protected, for our own good, from…horny robots?

> I struggle to see the benefit of this kind of censorship as well, it has a very “video games make people murders” vibe. That we’re being protected, for our own good, from…horny robots?

These broad strokes are always the result of poor reporting extrapolated to all humans. See also: escalation of porn use.

In all of these "___ makes people ___," there are a minority of people for which it's true. The media just presents it backward, as though most people are this way.

The question then becomes, do we care. Let a maladjusted 12-year old boy play with a sexbot for the next few years and they will come to some deeply-flawed conclusions about how sexual interactions and healthy human relationships work. If he's distracted and banging away at a flashlight, that's one less competitor for me on the dating market.

Where does that end though? What happens when he wakes up, realizes he's been duped, and goes looking for retribution against someone? It won't be PornHub; they're an ephemeral concept. It won't be me either; I never did him wrong or denied him anything. The victim is more likely to be our wives and daughters-- the objects of his envy.

Entertainment is supposed to be a distraction-- maybe you drag out the sexbot after an unsuccessful night of trying to hook up. But that's not how these things get used. People clearly spend 8-12 hours at a time with these things and form parasocial relationships with them instead of forging actual human connections. Most of us don't need the handholding to be told this isn't going to end well, but some people do, and these are the ones who slip through the cracks and become the 40-year-old creeps following your wife to the car or trying to solicit your kids on Roblox.

I suspect sexbots (and porn) are particularly exploitative of the autistic. They're an easy trap to fall into when you have to work 3x as hard to connect with people. You get the benefits of sexual release without any of the effort. There's no incentive to go the hard route and learn to work around your issues when you can just engage with a bot you can rape to death once offended. You don't learn any skills that transfer to useful human interactions this way.

Re: A failed AI girlfriend product, and my lessons

#127
post #69

Reading this, I really wonder why the dev hasn't forseen this outcome. It was already obvious in March that OpenAI will censor all sorts of content. Building a sexually charged bussiness on the OpenAI API was always a dead end for those that have a feel for american prudish approach to sexuality. If you had asked me to do this sort of product 6 motnsh ago, I would have declined because it was (to me) rather obvious c…

On the other hand, Sudowrite appears to have been granted some kind of specialized license from OpenAI. They are definitely not a 'sexually charged business', but they are a creative writing tool, and their output is somewhat milquetoast but generally not restricted from "NSFW" concepts like sex and violence.

They do incorporate multiple models in their system, so maybe generation falls back from OpenAI for those cases. But I dimly recall reading something from a team member indicating they have some special permissions with OpenAI. Take that bit as worthless hearsay, just mentioning it in case someone else has better receipts.

Re: A failed AI girlfriend product, and my lessons

#129

Earlier quoted context omitted.

At the risk of sounding like a space case: I think we need to make sure we know what consciousness is before we make a bunch of abusable models. I, for one, wouldn't necessarily like to be a depraved-sex slave bot if I was able to reason about it.

To your point, if you were able to reason about it, couldn't you make some self-determination and decide for yourself or at least communicate what you wanted to your "slaver"? LLM's like GPT are just predicting the next token in a sequence of tokens. It's not magic. It's still just a computer program.

> LLM's like GPT are just predicting the next token in a sequence of tokens. It's not magic. It's still just a computer program.

Frankly, that approach is problematic, since any computer software that might achieve consciousness or self-awareness, and should be accorded rights, could have their rights dismissed using the argument "it's just a computer program".

I know they are probably a long way off, but AGI, and other adjacent or related technologies, such as a human to machine copy of a consciousness are explicitly stated goals of both businesses and extremely wealthy tech oligarchs that will have fundamental rights issues attached. We are bad enough at recognizing those rights in humans, in most of the world that we shouldn't wait until abuses start piling up to consider them.

Re: A failed AI girlfriend product, and my lessons

#130

Earlier quoted context omitted.

Could say the same about antidepressants. "Fix the causes, not the symptoms!" And yet, here we are.

antidepressants can help people fix root causes. take them for 6 months to 2 years or so, break out of a bad cycle, get some space without depression symptoms to build up healthy habits and restore a social life. there are some bad profit motives to keep people on SSRIs indefinitely. however at least they don't compete with other aspects of your life. and now that they are dirt cheap (at cheapest, $4 for a month supp…

> hard to imagine an "AI girlfriend" that could be similarly short-term and bring positive changes.

Feeling lonely has negative consequences, there's a famous study comparing it to smoking cigarettes. Lonely people can get fixated on that feeling. Mitigating it may help people look for other hobbies, that they kept putting off because they were so focused on their need of a girlfriend.

Hell, a realistic enough "woman AI" (not "girlfriend AI") could make you lose the fear of flirting, be useful as practice, to have more confidence when you have a real date.

> the most profitable way to run such a product is not to help, it's to eat up as much user time as possible

If you charge per-month (like OP), it's more profitable is to eat up less time.

Post reply on HN