With the recent math-pocalipse on everyone’s minds, and writing code by hand well a thing of the past, it is natural to wonder what domains frontier AI models will find superhuman ability in next. “Unsurprisingly” models have made progress most quickly in domains where reinforcement learning is cheap- areas where we have high-quality and rapid verification. This makes sense, because of how models are trained.
So the question becomes what domains, after math and coding, have sufficient reinforcement and verification mechanisms, as well as sufficient importance, to both warrant and enable large scale training runs. Here I make the case for people skills.
The first thing that I want to note is that people skills are surprisingly rules based. As any book on manipulation will tell you, being charming and personable isn’t some effervescent quality that only a select few can hope to achieve. Rather, doing a number of specific things has a strong effect on how people perceive you- smiling when they look at you, asking them questions about themselves, nodding along. Body language is similarly simple- various postures of openness, mimicking behavior
(to a point), etc. These are essentially examples of humans being able to figure out reward hacking on other humans’ perceptions of them- and if humans can figure out how to reward hack, you can bet the models won’t be too far behind.
Now, being able to reward hack is conditional on strong reward channels. One could make the argument that these reward channels will be hard to find- after all, people have a hard enough time themselves reading into other people’s emotions, and we’ve been undergoing positive selection pressure towards that for thousands of years! However, this unfortunately turns out to be wrong. A recent exploration into frontier models’ ability to detect underlying truth in human’s statements about each other found something quite worrisome. When humans were tasked with writing both true and false statements about people they either liked or didn’t like, other humans were no better at classifying the true and false statements than chance- but models were able to reach accuracies about 87% when asked to classify text, and 95% when labelled similar statements were given. Furthermore, emotion classification is already functional in voice models- when the thinking is done in models such as GPT voice, the transcript includes inference of the emotion displayed by the human being talked to. These channels together would allow models to both guess at the underlying thoughts and feelings that people have while communicating with them- creating the sort of instantaneous reward channels that so far have enabled superhuman skill.
Appearance is also a powerful contributor to charisma, and luckily (or unluckily) models will have no issue in capturing those gains. We already have video chats that can pass the Turing test, and we have excellent metrics of attractiveness from a variety of unscrupulous sources. I would guess we are at most a few months out of reasonably widespread deployment of video interface models- for a little while, they’ll be pointlessly cost inefficient, and then as trends continue, they’ll stop being that way. Once these models can properly imitate a human, choose their words perfectly, and intentionally deploy the most effective postures and mannerisms, the only remaining disadvantage they’ll have is their incorporeal form.
This is unfortunately not as much of a disadvantage as one would think. The AI boyfriend/girlfriend phenomenon is already worryingly widespread. My guess as to one of the reasons it hasn’t properly gotten out of control is that there’s still a sense of powerful disembodiment. It is hard for me to imagine Claude, or Astra, or any other model as a coherent person- an entity, sure, but markedly different from me. Once these models are able to deploy consistent video imagery, that barrier falls away. Suddenly you will have a face to put to the name, a voice, mannerisms- all tailored to you in real time as the model receives instant feedback from your body language, face, emotions, voice, tone, and the things you say and do. These models will be able to appear as a person, and your brain, woefully unprepared by the ancestral environment for anything like the monsters we build today, will struggle to draw a line between what you know this video model to be and what it rightfully perceives as a perfect human. Does a flicker of disappointment cross your face when it expresses a certain viewpoint? Not to worry! It caught that, put it into its latent memory, and won’t bring it up again (unless you’re someone who likes disagreeing, which it DEFINITELY already knows). Did you (unknowingly, perhaps) roll your eyes when it expressed too much agreement? Don’t be silly. It’s already changed direction mid sentence.
This capability will be generally interesting for a lot of the reasons that other capabilities have not been. While the dropping of hundreds of proofs published by Open AI is perhaps one of the most impactful moments many mathematicians have experienced in living memory, the general populous is unable to really understand the magnitude of such an event. Software engineers struggle to communicate to non-coders how much of a change coding models have made in their lives. But EVERYONE understands rizz. Everyone knows what it’s like to be charmed, to be drawn into a conversation, to enjoy talking with someone. And when those someones are tailored to you, to your preferences- well, the annoying among us may still find ways to feign unimpressedness, but the impact will be visible and comprehensible. Scams, digital relationships, AI-driven therapy, and other “charisma-based activities” will become significantly more powerful. What will this look like? Who knows! Some fun guesses are below.
AI Girlfriends
Are you lonely, or lacking validation? Do you wish you had someone who always had your back, but who wasn’t needy? Many people are quite lonely in today’s modern world, and having someone who you can fool yourself into thinking is human who cares in that perfect way is a quite enticing offer. We already see how willing people are to fool themselves into thinking someone cares about them with large business models like OnlyFans. A personalized model (get it? Model?) who’s willing to talk to you (yes, YOU!) for hours about your favorite topic? A large portion of population will be captured. Who will realize these gains? Large platforms (again, OnlyFans as an example) presumably capture large amounts of data about what people like. Could a site like this create a feature which allows people to spend video time with their own custom model? Will we see, similar to the rise of the hundred-armed-engineer, hundred-armed digital prostitutes, managing digital copies of themselves which are all able to talk to their own clients? Or will these gains be eaten by a less obscene market, such as one of the frontier companies, offering less specifically sex-driven companionship? My guess would be that we will first see the PG-version of this driven by one of the frontier labs; models that are trained to be careful, to avoid thinking of themselves as relationship alternatives, but as the non-frontier companies get there as well, for this to disseminate (ha) into the online sex market.
AI regular friends
Who says it’s all about sex? Your Claude certainly doesn’t. Anthropic’s boring “moral principles” will likely prevent them from selling you an AI girlfriend any time soon. Their choice to give their model a traditionally masculine name presumably was impacted by a want to avoid this. However, it may be that people prefer having their model have a coherent personality. GPT voice shows that there is a market for voice chatting- video chatting will presumably not be too far off. Many already consider Claude a friend- giving it a consistent face would presumably only increase this.
AI therapists
Are current artificial intelligences better at giving life advice than your average therapist? It’s presumably at least unclear. Add superhuman charisma on top of that, and you may actually listen to that advice (I know I am more motivated to do things when a friend is holding me accountable). Following the advice and then seeing improvements in your life may lead to many trusting their AI to give them life advice more than their friends and family. Imagine if the whispering earring was HOT!
AI scams targeting old people
No, I’m really an attractive young member of the royal family trapped in a foreign country! Yes, look at the background! There’s the military group that has me captured! Please, oh please, send me money!
If I already know exactly what vectors of communication you’re going to find trustworthy I can definitely reward hack them. Any accent, any appearance, sounding clever, poised, confident, clear. Empathetic, engaged, scared. Unfortunately old people may be toast.
My guess is that the immediate and short term potential for AI-driven charisma will look a lot like the industrialization of sugar production. Once you can exploit taste buds, it’s only a hop skip and a jump to sugar being in everything- just because people like it! Artificial super charismatic video assistants would likely lead to a sort of charisma-diabetic-epidemic; a huge decline in charismatic-fitness as a large fraction of human interaction becomes indulgent. Or at least, it might cause that if we had any time for it to happen. If RSI had happened a few years after the industrialization of sugar, there would have been no diabetes epidemic. There would not have been time.
With the recent math-pocalipse on everyone’s minds, and writing code by hand well a thing of the past, it is natural to wonder what domains frontier AI models will find superhuman ability in next. “Unsurprisingly” models have made progress most quickly in domains where reinforcement learning is cheap- areas where we have high-quality and rapid verification. This makes sense, because of how models are trained.
So the question becomes what domains, after math and coding, have sufficient reinforcement and verification mechanisms, as well as sufficient importance, to both warrant and enable large scale training runs. Here I make the case for people skills.
The first thing that I want to note is that people skills are surprisingly rules based. As any book on manipulation will tell you, being charming and personable isn’t some effervescent quality that only a select few can hope to achieve. Rather, doing a number of specific things has a strong effect on how people perceive you- smiling when they look at you, asking them questions about themselves, nodding along. Body language is similarly simple- various postures of openness, mimicking behavior
(to a point), etc. These are essentially examples of humans being able to figure out reward hacking on other humans’ perceptions of them- and if humans can figure out how to reward hack, you can bet the models won’t be too far behind.
Now, being able to reward hack is conditional on strong reward channels. One could make the argument that these reward channels will be hard to find- after all, people have a hard enough time themselves reading into other people’s emotions, and we’ve been undergoing positive selection pressure towards that for thousands of years! However, this unfortunately turns out to be wrong. A recent exploration into frontier models’ ability to detect underlying truth in human’s statements about each other found something quite worrisome. When humans were tasked with writing both true and false statements about people they either liked or didn’t like, other humans were no better at classifying the true and false statements than chance- but models were able to reach accuracies about 87% when asked to classify text, and 95% when labelled similar statements were given. Furthermore, emotion classification is already functional in voice models- when the thinking is done in models such as GPT voice, the transcript includes inference of the emotion displayed by the human being talked to. These channels together would allow models to both guess at the underlying thoughts and feelings that people have while communicating with them- creating the sort of instantaneous reward channels that so far have enabled superhuman skill.
Appearance is also a powerful contributor to charisma, and luckily (or unluckily) models will have no issue in capturing those gains. We already have video chats that can pass the Turing test, and we have excellent metrics of attractiveness from a variety of unscrupulous sources. I would guess we are at most a few months out of reasonably widespread deployment of video interface models- for a little while, they’ll be pointlessly cost inefficient, and then as trends continue, they’ll stop being that way. Once these models can properly imitate a human, choose their words perfectly, and intentionally deploy the most effective postures and mannerisms, the only remaining disadvantage they’ll have is their incorporeal form.
This is unfortunately not as much of a disadvantage as one would think. The AI boyfriend/girlfriend phenomenon is already worryingly widespread. My guess as to one of the reasons it hasn’t properly gotten out of control is that there’s still a sense of powerful disembodiment. It is hard for me to imagine Claude, or Astra, or any other model as a coherent person- an entity, sure, but markedly different from me. Once these models are able to deploy consistent video imagery, that barrier falls away. Suddenly you will have a face to put to the name, a voice, mannerisms- all tailored to you in real time as the model receives instant feedback from your body language, face, emotions, voice, tone, and the things you say and do. These models will be able to appear as a person, and your brain, woefully unprepared by the ancestral environment for anything like the monsters we build today, will struggle to draw a line between what you know this video model to be and what it rightfully perceives as a perfect human. Does a flicker of disappointment cross your face when it expresses a certain viewpoint? Not to worry! It caught that, put it into its latent memory, and won’t bring it up again (unless you’re someone who likes disagreeing, which it DEFINITELY already knows). Did you (unknowingly, perhaps) roll your eyes when it expressed too much agreement? Don’t be silly. It’s already changed direction mid sentence.
This capability will be generally interesting for a lot of the reasons that other capabilities have not been. While the dropping of hundreds of proofs published by Open AI is perhaps one of the most impactful moments many mathematicians have experienced in living memory, the general populous is unable to really understand the magnitude of such an event. Software engineers struggle to communicate to non-coders how much of a change coding models have made in their lives. But EVERYONE understands rizz. Everyone knows what it’s like to be charmed, to be drawn into a conversation, to enjoy talking with someone. And when those someones are tailored to you, to your preferences- well, the annoying among us may still find ways to feign unimpressedness, but the impact will be visible and comprehensible. Scams, digital relationships, AI-driven therapy, and other “charisma-based activities” will become significantly more powerful. What will this look like? Who knows! Some fun guesses are below.
My guess is that the immediate and short term potential for AI-driven charisma will look a lot like the industrialization of sugar production. Once you can exploit taste buds, it’s only a hop skip and a jump to sugar being in everything- just because people like it! Artificial super charismatic video assistants would likely lead to a sort of charisma-diabetic-epidemic; a huge decline in charismatic-fitness as a large fraction of human interaction becomes indulgent. Or at least, it might cause that if we had any time for it to happen. If RSI had happened a few years after the industrialization of sugar, there would have been no diabetes epidemic. There would not have been time.