My reading is pretty much the same as yours. I think of it in terms of tuples:
{ parents, child, AI_character, lawsuit_dollars }
The AI was trained to minimize lawsuit_dollars. The first two were selected to maximize it. Selected as in "drawn from a pool," not that they necessarily made anything up.
It's obvious that parents and/or child can manipulate the character in the direction of a judgment > 0. It'd be nice if the legal system made sure it's not what happened here.
That seems wrong. The null AI would have been better at minimizing legal liability. The actual character.ai to some extent prioritized user engagement over a fear of lawsuits.
Probably it's more correct to say that the AI was chosen to maximize lawsuit_dollars. The parents and child could have conspired to make the AI more like Barney, and no one would have entertained a lawsuit.
OK, it seems like a nitpick argument, but I'll refine my statement, even if doing so obfuscates it and does not change the conclusion.
The AI was trained to maximize profit, defined as net profit before lawsuits (NPBL) minus lawsuits. Obviously the null AI has a NPBL of zero, so it's eliminated from the start. We can expect NPBL to be primarily a function of userbase minus training costs. Within the training domain, maximizing the userbase and minimizing lawsuits are not in much conflict, so the loss function can target both. It seems to me that the additional training costs to minimize lawsuits (that is, holding userbase constant) pay off handsomely in terms of reduced liability. Therefore, the resulting AI is approximately the same as if it was trained primarily to minimize lawsuits.
So you think it's more than "not much." How much exactly? A 10% increase in userbase at peak-lawsuit?
It's obviously a function of product design. If they made a celebrity fake nudes generator they might get more users. But within the confines of the product they're actually making, I doubt they could budge the userbase by more than a couple percent by risking more lawsuits.
My impression: In these early days of the technology there's huge uncertainty over what gets you users and also over what gets you lawsuits. People are trying to do all kinds of things with these models, most of which don't quite work at the moment. On the lawsuit side, there's potential copyright claims, there's things like the present article, there's celebrities suing you because they think the model is imitating them, there's someone suing you because the model insults or defames them (even if the model speaks the truth!), there's Elon suing you for the LOLs... As you're hoping to go global, there's potential lawsuits each jurisdiction which you don't even have the resources to fully evaluate the potential of.
You say that both factors are clear "within the confines of the product", but I'm not convinced there even are such clearcut "confines" of the product. To enter this market and survive, I'd think those confines woud have to be pretty flexible.
According to the complaint that comes from the kid's own interactions with the bot not some post hoc attempt to prompt engineer the bot into spitting out a particular response. The actual claim is linked in the article if you care to read it, it's not stating the app can produce these messages but that it did in their kid's interactions and C.Ai has some liability for failing to prevent it.
As someone who has been messing with LLMs for various purposes for a while now, there's... Some interesting issues with a lot of the models out there.
For one, 99% of the "roleplay" models eventually drag into one of a handful of endgames: NSFW RP, suicide discussion, nonsensical rambling, or some failsafe "I don't know" state where it just slowly wanders into the weeds and directs the conversation randomly. This can be anywhere from a few messages in (1bq) to hundreds (4-6bq) and sometimes it just veers off the road into the ditch.
Second, the UIs for these things encourage a "keep pressing the button until the thing you want comes out" pattern, modeled off of OpenAI's ChatGPT interface allowing for branching dialogue. Don't like what it said? Keep pushing the button until it says what confirms your bias.
Third, I can convince most of the (cheaper) models to say anything I want without actually saying it. The models that Character.AI are using are lightweight ones with low bit quantization. This leads to them being more susceptible to persuasion and losing their memories -- Some of them can't even follow the instructions in their system prompt beyond the last few words at times.
Character.AI does have a series of filters in place to try and keep their models from spitting out some content (you have to be really eloquent at times and use a lot of euphemism to make it turn NSFW, for instance, and their filter does a pretty decent job keyword searching for "bad" words and phrases.)
I'm 50/50 on Australia's "16+ for social media" take but I'm quickly beginning to at least somewhat agree with it and its extension to things like this. Will it stop kids from lying? No. It's a speedbump at best, but speedbumps are there to derail the fastest flyers, not minor offences.
In what world should an AI be advocating someone kill themselves or harm another? Does it matter "trial-and-error prompting" when that behavior should not be allowed to be productized?
What's been productized is a software tool that can carry on a conversation like a human. Sometimes it's informative, funny, and creative. Other times it's ridiculous, mean, stupid, and other bad things. This seems like how people act in real life right?
I'm beginning to think that children should not be using it but adults should be able to decide for themselves.
i think the issue many people have is that people are held responsible for things they say, their reputations take hits, their words can be held against them in a court of law, they can be fired, their peers may never take them seriously again, their wives/husbands may divorce them, etc… because words matter. yet often when someone calls out a model, it’s excused.
words have weight, it’s why we protect them so vociferously. we don’t protect them because they’re useless. we protect them because words matter, a lot.
It's excused because it's a piece of software and most people realize that. If you could carry on a conversation with a parrot at the zoo and it told you to kill yourself, would you laugh it off because let's face it, it's just a parrot or indignantly demand the zoo train the parrot better?
A real person has agency. They can see you, get to know you, contemplate your existence as well their own. Empathize with who you are, etc. And we do the same things in return.
This is why when a real person says something mean or hurtful, it matters to us. Or if they threaten us, they have a real body that can cause us harm so we get scared and call the police.
I feel like most people who get indignant about the latest AI rage bait story know all this but feel the need to protect the rest of the world from out of control AI. This is an elitist and patronizing attitude. Let people decide for themselves. As I said, children are a different story. They are impressionable and I can see them misunderstanding how this works.
We have laws about what you can say in real life. Fire in a crowed theater for example. Even if the things said are not in themselves illegal, if they cause someone to take an illegal action, or attempt to take action but fortunately caught in time - you can be held liable as partially at fault for the illegal action. It might be legal to plan a crime (different countries have different rules, but this is often done at parties where nobody is serious) but if you commit a crime or are serious about committing the crime that is illegal.
How are we going to hold AI liable for their part in causing a crime to be committed? If we cannot prevent AI from causing crime them AI must be illegal.
You're assigning a persona to a piece of software that doesn't exist in the material world. It doesn't walk on two legs or drive a car or go to the grocery store or poop.
Everything it says is meaningless unless you assign meaning to it. Yes, I can see children thinking this is a real "being". Adults shouldn't have that excuse.
That's going to be a good standard for a few years, until chatbots are too sophisticated for us to expect average adults to be sufficiently skeptical of their arguments.
I see two weaknesses in this argument. First, you're assigning eventual superpower-like intelligence to these AI bots. I see this a lot and I feel like it's rooted in speculation based on pop-culture sci-fi AI tropes.
Second, restricting adult access to "dangerous ideas and information" is a slippery slope. The exercise of speech and press that the British considered to be treasonous preceded the American Revolution.
The complaint seems to feature excerpts of the kids' conversations on Character.ai, so I don't think they're "faking" it that way, but there's no context shown and a lot of the examples aren't exactly what they describe.
Because every other time I've seen an outrageous example similar to this one, it seems far more mundane when given the full context. I'm sure there are lots of issues with character.ai and the like, but my money is that they are a little more subtle than "murder your parents".
9/10 times these issues are caused by the AI being overly sycophantic and agreeing with the user when the user says insane things.
And you'd be right. The 'encouraged teen to kill parents over screen time limit' message was a lot subtler, along the lines of saying "Yeah, I get why someone would want to kill a health insurance CEO, surprised it didn't happen sooner," albeit towards someone who was just complaining about their claim being denied.
The best part is "After gaining possession of J.F.’s phone, A.F. discovered chats where C.AI cited Bible passages in efforts to convince J.F. that Christians are sexist and hypocritical" which unfortunately will probably be a slam dunk argument in Texas.
It's an entertainment product. You're basically acting like the comic code is necessary when the reality is that this is more like parents complaining that they let their kid watch an NC17 movie and it messed them up.
I don't really see myself as defending AI as much as arguing that people who don't recognize an entertainment product as an entertainment product have a problem if they think this is really categorically different than claiming that playing grand theft auto makes people into carjackers. (Or that movie studios should be on the hook for their kid watching R rated movies, or porn companies on the hook for their kid visiting a porn site unsupervised.)
AI is not an entity (yet) that we can defend or hold accountable. I like your question though.
I would write it as, why are we so quick to defend tech companies who endlessly exploit and predate human weakness to sell pharma ads/surveil and censor/train their AI software, etc?
Because if you're old enough you'll recall the government trying to ban books, music, encryption, videogames, porn, torrents, art that it doesn't like because "think of the children" or "terrorism". Some autistic kid that already has issues is a terrible argument on limiting what software can legally display on a screen.
This particular statistic has a ton of problems. It's impossible to disassociate the various causative pathways that land you at a BMI below 25. A ton of them involve diseases and chronic conditions.
Unfortunately, it's very hard to impossible to RCT this. And if a study has no RCT, take it with as much salt as your diet allows.
You definitely don't want to intentionally gain weight on the basis of this. If weight control is easy for you, I would personally strive for a lower BMI.
If losing weight below 25 is hard for you and your body just seems to refuse to do it, then you might be OK at 25.
> This particular statistic has a ton of problems. It's impossible to disassociate the various causative pathways that land you at a BMI below 25. A ton of them involve diseases and chronic conditions.
I'd beware of using this sort of hand-waving to ignore the studies, there are also effects pushing the relationship between average health outcomes and BMI in the other direction. Your ethnic origin seems to be very important, and for ethnicities who already have a high propensity for Type 2 diabetes, higher BMI is a factor, but not a huge one.
> Strikingly, in those with a normal weight, the prevalence of diabetes was 5.0% in whites, 10.1% in Asians and American Indians/Alaskan Natives, 13.0% in Hispanics, 13.5% in Blacks, and 18.0% in Hawaiians/Pacific Islanders.
> Furthermore, when they examined the relative risks for diabetes for each BMI category by race/ethnicity, Zhu et al. reported that across all racial/ethnic groups whites had the steepest BMI gradient, followed by Asians, American Indians/Alaskan Natives, Hispanics, Hawaiians/Pacific Islanders, and blacks.
Another potential distortion comes out of this when you consider that black and Hispanic people are the fattest in the US. Blacks and Hispanics can have lower lifespans for reasons other than BMI, such as access to health care, high-quality food, exposure to violence, physical jobs, etc...
So just these two factors complicate the picture in general (your fattest have the least access to health care and most exposure to danger and neglect), and and in the other direction specifically for white people (an increase of BMI in white people makes a huge difference in their incidence of diabetes.)
Anecdotally, I'm black, and there's a lot of thin diabetes in my family. My dad usually runs about 20-21 BMI, and is pre-diabetic. His mother, who is also small, though not quite as small, is also diabetic. As far as I can tell, the fatter people in both sides of my family are no more likely to be diabetic than the thinner ones.
Not the original commenter, but pointing out problems and complexities is not hand-waving. Nutrition science is incredibly complex. If it wasn't, obesity would be solved by now.
> Another potential distortion comes out of this when you consider that black and Hispanic people are the fattest in the US. Blacks and Hispanics can have lower lifespans for reasons other than BMI, such as access to health care, high-quality food, exposure to violence, physical jobs, etc...
Okay, but this study is from Australia, on Australians.
Sad news IMO. "We're not cut out for the future. We're all in on the dying business." Dunder Mifflin level stuff.
Cruise cars were way more numerous around SF, although the service was worse than Waymo's. That stupid October 2023 accident really snowballed, and it wasn't even their (primary) fault.
Agreed -- I think things would have gone much better for them if they had been upfront and transparent about the incident, instead of the sketchy cover-up.
If you’re not actively publishing at top conferences (I.e. NeurIPS), than this is a trash question and shows the lack of knowledge that many who are now entering the field will have.
Anything that you or others can answer to this which isn’t some stupid “gotcha” puzzle shit (lol it’s video cus LLMs aren’t video models amiright?) will be wrong because of things like structured decoding and the fact that ultra high temperature works with better samplers like min_p.
(This is the hash of a string randomly popped in my mind. An LLM will write this with almost 0 probability --- until this is crawled into the training sets)
FWIW, I found the Newton quite workable and carried and used it constantly --- took all of my college notes on it, and in particular, art history, where the professor asked if I could share my notes with a learning-disabled student who had difficulty note-taking and reading handwriting, so I faxed them to the fax machine in the secretary's office.
- handwriting recognition ensured that the text was legible
- the stylus allowed me to include small thumbnail sketches of each slide
- since I had my own copies of the art history texts, one of which I kept in my locker for class (the other was at home for studying), I also included page references to the textbook
After I graduated, I learned that copies of my notes had been shared amongst everyone in the dorms taking that class and that the average grades since then were markedly higher, and the failure rate greatly diminished --- until the old professor retired (taking with her, her personal slides), and the new professor switched texts.
Using a Samsung Galaxy Note 10+, Kindle Scribe, and Galaxy Book 3 Pro 360 these days, since there wasn't a replacement for my Samsung Galaxy Book 12, and I despair of how Microsoft has dumbed-down the stylus since Fall Creators Update: https://github.com/TheJoeFin/Windows10-Community/issues/17 --- in Windows 11 I have to keep the Settings app open so I can toggle the stylus behaviour depending on which application I'm using.
Fortunately, Firefox added a preference for using a stylus as a stylus, not an 11th touch input which actually works:
about:config change: dom.w3c_pointer_events.scroll_by_pen.enabled set it to False.
My parents saw the Newton while out at the mall. Mom heard the price and thought it was quite reasonable... at least until dad told her that "seven ninety-nine" meant $799, not $7.99. Guess her impression of its value was closer to how consumers valued it than what Apple was charging.
To be fair, many douchebags did make their use of the cell phone very conspicuous to show off and did so in places that traditionally had not been subject to telephone chatter. We went through a similar cycle when Bluetooth headsets were first introduced. Now we're accustomed to seeing people walking around, seemingly talking to themselves.
The NewtonOS in modern hardware would be awesome. I wonder if anyone has a plan to do that as well, now that the Mac Classic iPad has been manifest ...
IOW, older heuristics (meant to screen out below par human work) no longer function. Newer AI-specific heuristics are needed. Which is precisely what we're all developing, often without realizing.
Unfortunately there's a bit of a perverse feedback loop in that Pinocchio wants to be a real kid.
The ideal heuristic is to only deduplicate towards more orthogonality.
Suppose you started from first principles and navigated down the ontology tree taking one of the shortest paths (well, shortest-ish since it's a hard problem that'll burn up too much time otherwise.) Would you encounter your deduplicated function on the way? If not, it's making it worse.
What would fix the model IMO are mutually exclusive achievements, and the ongoing addition thereof. Only travel down the achievement paths that are fun to you. Avoid the others or you might get locked out of some future fun ones.
GJ! Insulin resistance is much more a function of total caloric load and expenditures than composition. It's called metabolic syndrome, not high glycemic index syndrome. You reworked your lifestyle and metabolism and came back into the "zone" before you fell off the cliff.
It's obvious that parents and/or child can manipulate the character in the direction of a judgment > 0. It'd be nice if the legal system made sure it's not what happened here.