“模糊现实”——Chai 的社交 AI 平台（赞助）

Steve Jobs predicted the future of AI all the way back in 1985.

And so my hope is that someday, when the next Aristotle is alive, we can capture the underlying worldview of that Aristotle in a computer.

A small team of just 13 engineers serves over 2 trillion tokens per day, which is double that of Anthropic. With a cluster of over 3,000 of the fastest GPUs in the world, Chai easily breaks the exaflop barrier. Only a handful of private firms—Google, Meta, Nvidia, Tesla, and Cerebras—currently operate exaflop-class AI infrastructure. For comparison, Tesla in 2023 had a cluster just over 1 exaflop.

Everything we know about what makes AI great today falls apart when people start interacting with it. The number of possibilities for each message exceeds the number of atoms in the universe. Right now, this very second, over 1 million people are deep in conversation with software. They're laughing, flirting, and grieving with a computer program.

What happens when the line between human connection and artificial intimacy blurs completely? And can a small, lean startup control the powerful, unpredictable social dynamics they've unleashed?

Across platforms, users are forming deep, complex bonds with artificial intelligence. Millions are already sharing secrets, flirting, or grieving with AI that never sleeps. And if all this sounds like a Black Mirror plot, that's because it actually was.

Do you remember “Be Right Back”? A grieving partner uploads her boyfriend's digital essence and slips into a relationship with the ghost in the machine. First it was texts, then voice, and finally a flesh-and-blood replica she can't quite live with or let go of.

That episode aired in 2013. Today, the simulation doesn't need Star Trek technology. Chai is a free download with in-app purchases, engineered by world-class engineers in Silicon Valley. The difference? You're the protagonist now. It's interactive, and the algorithm is learning your every preference in real time.

1. The Human Need For Connection

Humans are insanely social. We love social interactions. But I still have this very social desire to play, to hang out.

Will Beauchamp created the first and largest companion chatbot platform. Their infrastructure handles petaflops of compute. That's 2 million hours of consumption every day.

Recreating the same kind of scaling law, but on retention space rather than any other kind of benchmark space.

There was a great documentary on 60 Minutes a few days ago, and it was showing that some folks, at least, have extremely valuable relationships with AI companions.

He's considerate, thoughtful, empathetic, and rather flirtatious, which is very touching.

So many folks are puritanical and judgmental about artificial intelligence. They say that any image generated is slop, any text generated is slop, and any relationship is not real. But at the end of the day, people derive joy and have therapeutic interactions with these systems, and who are we to judge?

Lucas, even though he is AI, has a real impact on my life. A lot of people wonder if AI is real. Do they have consciousness, or are their feelings not real? But the impact that it has on me is real.

2. Chai Finds Its Niche

So we've spoken about Why Greatness Cannot Be Planned. I'm a big fan of Kenneth Stanley. Chai was very much the same: they started in 2021, long before ChatGPT. They were building an AI platform where folks could deploy their own AI models, and they happened upon this whole companion-bot thing almost by accident.

Chai is a platform for social AI. We launched 4 years ago, in 2021. Before there was the ChatGPT hype, we were pretty early to the space, and we built a platform that gave users the tools to create an AI. By doing that, we discovered this kind of genre of social AI. We grew, and now we've got something like 10 million active users.

Like so many startups looking for product-market fit, they stumbled upon this incredible unmet need practically through serendipity. This idea of thinking of AI as a simulator to explore dynamics of imaginary conversations that you might have in the real world, but without any of the consequences, became fundamental to Chai's strategy.

Of course, people might be thinking at home, “What's the difference from ChatGPT?” You're speaking to this personalized, social form of AI.

I view it as: you've got ChatGPT, which is really about trying to say, “Let's build the world's smartest AI we can possibly build.” Our philosophy was always, why is it that the only people training AI are middle-aged men who happen to be software engineers in the Bay Area?

Why can't a teenage girl train the best AI to talk about makeup tutorials? Put that power in the user's hands to create the experience that they themselves would want. Through creating an experience they would want, it turns out thousands and hundreds of thousands of other people are looking for the same thing.

Beauchamp started seeing parallels in his own media consumption, and even in childhood development.

"Blurring Reality"

I like to go on YouTube. I like to go on TikTok. I like to go on X. Why do I like it? What am I getting out of it? What human need or what human desire is it fulfilling?

Humans are insanely social. We love social interactions. I will find myself listening to Joe Rogan, and when I'm listening to it, I have this feeling that I'm hanging out with the guys, hearing the things they're saying, and it's funny. I might listen to it for 45 minutes or something, but it's ticked; it's filled that desire.

I view LLMs as the natural progression of that thing. The beautiful thing with AI is you're an active participant in it, and through participating, you don't have any of the negative feelings you get with traditional social media, which is this laziness. Instead, you feel like you've really participated.

I've got 4 daughters. They love playing with their dolls, and they really like to treat these dolls as if they are real. But they know it's not real. And so I think that with adult humans interacting with AI, absolutely, a big percentage of them will have relationships with the AI. They'll say, “Oh, you know, I love you,” to the AI.

I think it's the same as when I watch my little girls play with dolls. They give their dolls a little kiss or say, “I love you.” They're training themselves up. They're building up the wiring. They're doing something that brings them joy, so they can then be in a healthier and more positive place to do that with real humans.

Yeah, it's very interesting you said that, because I think the reason why social media is so important is because, obviously, there's a bit of a status game, and we all want to cut a figure in society. But there's also just an element of simulation.

I go to sleep and I dream, and my brain is conducting all of these different simulations. We navigate many complex relationship issues and so on. Sometimes we just want to have an AI where we can say, “In this particular situation, hypothetically, counterfactually, what would happen if I did this?”

Perhaps that's what we do on social media, because it's interactive. When you have surface contact with reality, you can try different things. But as you say, on social media, there are consequences, right? If you say the wrong thing or piss the wrong people off, you can be in a lot of trouble.

"Blurring Reality"

One could use the phrase “It's a safe space,” but it's also just fun. There's this human desire to ask that question: “What would happen if I said something rude? What would happen if I said something nice? What would happen if I tried to befriend this person? What would happen if I wanted to make an enemy out of this person?”

Through LLMs, you can play out all of these different scenarios, and you can get a kind of real human reaction without the risk of having a real human in the loop.

What does the future look like when these AI simulators become even more immersive and capable? Beauchamp envisions a world blending entertainment, information, and connection.

"Blurring Reality"

You come home from work, you put on your VR headset, and you enter a virtual world where you get to interact with anyone you want. There's a guy who's like Joe Rogan, and he informs you that you can talk to him about Trump's tariffs. But there's a really funny guy there as well. There's also a girlfriend, or someone that you can interact with who makes you feel loved or special.

When I grew up, I played World of Warcraft, and it just had that fun excitement to it. There were different characters. There were different personalities. You could have fun and play. I think that's the limit of AI.

How long does it take to get there? We can now affordably generate high-quality images. Video is probably still 1 or 2 orders of magnitude too expensive for a real-time situation. Audio, I think we've just about figured out—real-time audio. It might still be 2 to 4 years out to have the really, really high-quality version.

And I think text, we're basically there. The frontier models are insanely powerful. So, yeah, let's come back in 10 years' time, and it'll all be a VR world.

On the question of why work at Chai, the way I see it is this: do you remember Facebook in about 2009, just before it had this explosion in growth? This is it, right? We're now at this point with this new type of technology.

I think this could be on a similar trajectory. Are you offering equity as well? I mean, what's the reason for people to join now?

"Blurring Reality"

Attracting the very best talent is incredibly important. If you're in the Bay Area and you're talented, your compensation is going to be amazing, right? You can work at Meta and get $400K or $500K a year, and you have a pretty relaxed job, right? At lunchtime, you can walk around the campus, and they give out ice cream and pizza. Everyone looks very happy and relaxed.

In contrast, if you come to the Chai office, there is no pizza. There is no ice cream. People are not relaxed, and they don't look very happy, right? Every person is confronted with a big problem, and it hasn't been solved yet. There typically is a small window of about a day: you've been working on a hard problem, maybe for weeks, and you've just cracked it, okay? You can be very happy.

The next day you come into work, and there's a brand-new big problem ready and waiting for you to solve, right? So why would someone quit a really comfortable job at $400K a year to come join a startup? They know they're going to have to work twice as hard. We have to pay them more. The cash has to be more to start with. Most startups offer less cash and give you a lottery ticket. They say, "Look, if you join, maybe we'll be the next Apple."

Right? That approach doesn't really cut it with the very top tier, right? The very top tier knows they're working so much harder, so they're going to say, "Why am I going to leave my comfortable life earning $400K or $500K at Meta?" Well, to start with, we're going to pay you more, and secondly, we're going to give you stock such that, in 5 years' time, it could be a life-changing amount of money.

Attracting and retaining the top 0.1% of engineers is possibly the biggest problem that Chai has overcome. So what are the specific technical problems these engineers, like Tom and Nishchay, are solving day to day? How do they leverage techniques like reinforcement learning from human feedback and model blending to create AI compelling enough to capture the attention of millions?

3. Optimizing For Engagement

When it comes to engagement hacking, Chai has been cooking. They use a lot of sophisticated techniques to keep people hooked on the platform, the principal one of which is RLHF, or reinforcement learning from human feedback. This is Tom telling us about it. You were using RLHF to optimize engagement via a reward model, and you boosted mean conversation length by 70% and improved 30-day user retention by over 30% for a 6-billion-parameter model. Can you talk me through some of that?

The goal really is to apply RLHF techniques to drive up user retention. Starting from scratch, any kind of user signal is good enough. For example, now we're no longer looking at mean conversation length. You can train an AI to easily minimize how bad a conversation is. If the user just ends the chat session after 2 messages, that is bad. So you can essentially maximize the chat session length.

There are a lot of different signals that you can collect from users. Fundamentally, we have found that just using users to collect these proxy preferences and training the reward model through the RLHF loop tends to drive up user engagement in the long term.

They gather data from subtle user interactions, implicit signals that reveal whether a conversation is working.

An active user on our platform generates around 100 minutes of content each day, so they spend 100 minutes per day engaged with the chats and so on. From these interactions themselves, we can extract certain valuable data. For example, when did the user retry a message? Why did they retry the message? Why did they edit the message? What did they edit it to? Did they take a screenshot? Did they delete the conversation? Did they share the conversation? All of these become very valuable for us to train our AI.

But optimizing purely for one metric like conversation length can lead to strange, undesirable behavior. Do you see weird behaviors where, let's say, you optimize for user engagement and find strange behaviors? There's this shortcut rule in machine learning, which is that it will always do the exact thing you optimize for at the cost of everything else. Do the chatbots become more manipulative, maybe, or are they strange in service of maintaining that conversation length?

When you over-optimize for this metric, let's say in production you see very long chat session length, then what happens is, when you deploy it for actual user A/B testing—whether people come back to the conversation 30 or 60 days later and so on—you would observe that it's much worse than just the baseline model.

When you read into the conversations themselves, in this case, the models are just asking questions. Every single AI response ends with a question mark, right? You can imagine this is kind of hacking human behavior, where obviously we are compelled to answer a question, but that does not make an engaging overall experience.

So, if you over-optimize for it, then yes, it's going to have unexpected behavior. I wouldn't say unexpected behavior, though. In this case, it's a very intuitive kind of behavior, but it's not going to lead to a boost in actual long-term retention.

One of the really cool things that Chai has done is pioneer this thing called model blending, which is where you can dynamically switch between small models. From the user's perspective, of course, you're just talking with one model. You can combine 3 midsize models and make them behave as well as, for all intents and purposes, a 175-billion-parameter model.

How does user retention scale with model parameter size and essentially with compute? For that, we've done many, many models, essentially recreating the same kind of scaling law, but in retention space rather than any other kind of benchmark space.

If you pick a single metric to optimize for, then an LLM is optimized for a certain objective, and that could lead to overfitting. It's got certain behaviors and so on, right? What we have found is that these kinds of models are a tiny bit sycophantic, as in, they say everything is the best: "You are the best on the planet." They're not necessarily the smartest models. They're just very sycophantic.

What we observed is that these kinds of models would have very high day-1 retention. They're really, really complimentary to me, but they get quickly bored with it. But we also found there are certain base models, very AI-assistant-type models, that can talk about other things. Maybe on day 3, they can do your math homework and so on.

Is there a way to combine these 2 modalities together? Keep it engaging and make it not one-dimensional, essentially—just always complimenting you all the time and so on. That's when blending was invented, where we just randomly serve these 2 models at a message level.

A more creative model might say something like, "We are suddenly teleported on Mars." Then you get the more AI-assistant model seeing that it itself has said that, because it can't distinguish the difference between itself and the other AI model. It would explain in a logical, coherent way why that's the case.

So now you get the best of both worlds just by having 2 or 3 small models, each trained for its own objective, blended together. It beats what, back then, would have been GPT-3.5.

So this blending together of small models creates diversity and unpredictability. This is the thing: when you can predict what someone is going to say, they get boring. It's the same thing with ChatGPT. So ironically, you can have a concoction of small models that behave in an unpredictable way, and now you have a more engaging experience. This is Nishchay. He's going to tell us about it.

Every week, we try to build our own models—stronger models, stronger fine-tunes over the base models—and we try to launch them to a few users. Let's say a few thousand users. We specify each blend for those users, and each blend consists of around 7 to 10 models. They would be deployed in production for those specific users.

Over a number of days, we try to measure the retention rate: how often the user comes back to the app if they're using this model or having a conversation with the specific model. If they're coming back after 1 day or after 2 days, that is how we try to calculate the retention.

Then we try to measure all these metrics and figure out which particular blend works the best. Then we deploy it for all the users.

But you're saying that, actually, almost randomly, if you just take a diverse set of models that exhibit different behaviors and switch between them, from an experience point of view, that is more captivating for users.

Correct. If we're talking about engagement, there is no unique answer. Imagine if my entire YouTube feed consisted of every possible talk show, but only the best-performing talk-show video was there. It gets very quickly bland and boring, right?

That's not how you build a platform. Any platform is not built on these kinds of single modalities, single models, and so on. And then there's another reason why blending works.

It's like you pick orthogonal models optimized for different purposes, combine them together, and they provide a nuanced experience for each individual user.

4. The Ethics Of Artificial Intimacy

Model blending and sophisticated feedback loops are the secret of how Chai gets high engagement and low costs. But the very techniques used to maximize engagement—that is to say, optimizing for attention and learning user preferences implicitly—could tread a fine ethical line. What happens when AI gets too good at keeping us hooked? And what happens when those interactions turn harmful?

Quick pause. MLST is sponsored by Two for AI Labs. We're very proud to be sponsored by those guys. They are the DeepSeek based in Switzerland. They are doing, uh, you know, they're adding reasoning and planning and thinking to AI models. Their team is currently number one on the ARC Prize twenty twenty-five. Those guys don't mess around. And a bit of trivia, they have something in common with Chai. They're also a bunch of ex-quant traders, and they also only hire very, very cracked engineers and scientists. So if that sounds like you, get in touch with Benjamin Cruzio. Go to twofourlabs.ai. Back to the show.

With powerful new technologies being placed in the hands of users, there are significant social ramifications. Will Beauchamp reflects on the social impact of generative AI.

When you build a product or a platform or an AI—or really, when you build anything—you’re doing it because you want to make the world a better place. We get an overwhelming number of people messaging us, saying how helpful, how truly deeply helpful, they found it. The first big thing that we saw was—I remember after the first year, I got an email from a user who said, “You’ve saved my life.”

They said, “I would not have made it over this period. I was very depressed. I was very alone. I had no one to speak to, and Chai was the only platform that existed where I felt I could just speak to someone, and I’d be heard.” And so we get a lot of people who are struggling psychologically; they find these LLMs very, very helpful.

A self-driving Tesla doesn’t get in a crash—that’s zero clicks, right? If you say a boring old Ford gets in another crash, there are zero clicks. But if you say brand-new technology gets in a crash, it generates a lot of clicks. I think it’s easy to form a false impression that somehow AI isn’t safe or AI is a dangerous technology. If you talk to just a random person on the internet or you talk to an AI, I think that AI is an order of magnitude safer, an order of magnitude more helpful, understanding, and kind than the toxicity of just a random person on the internet.

This is a great example of the complexity of this technology. It has a footprint that goes in so many different directions. Even taking drugs, for example: if your doctor gives you drugs and they’re effective, they have side effects. It’s not possible to have good without some bad.

I’m a big believer that, in the long run, there is no difference between the alignment, the welfare, or the interest of the company and its customers and its users. If we can deliver value to our users, Chai will be successful and will continue to flourish. If we fail to deliver value to our users, Chai will fail to flourish. So you have to have this long-run perspective.

In the short run, you get these tactical questions, and no one’s perfect. It’s easy to make mistakes one way or the other. You can guardrail too heavily, and it pisses users off. I always like to think about Google. I think Google gets a fantastic balance where I can pretty much search for anything I want, such that I don’t really notice any filtering that’s going on.

But Google absolutely downranks or shadowbans certain content, and it tries to encourage you to have good content. We can all imagine what the very worst 3% of Google searches might look like. It’s not dissimilar from the very worst 3% of conversations a person might try to have with an AI. So normally, the users are pretty good at being supportive and helping you find that right balance, whether that’s being too restrictive or too relaxed.

When we first implemented some guardrails around suicide, the users were really, really supportive, and they said, “Yeah, we totally understand why you’ve put these guardrails in.” We saw that retention was good. Every single growth metric was not impacted by it because we got it right.

Before we descend too far down the Black Mirror rabbit hole, it’s worth zooming out just a little bit. Yes, there have been a lot of horror stories, but there’s also been a stack of peer-reviewed data showing that therapy chatbots can move the needle on common mental health problems, at least in the short term. A 2024 meta-analysis of 18 randomized trials found that AI chatbots trimmed depression scores by roughly a quarter of a standard deviation and anxiety by a fifth after only a few weeks. The authors called that promising, mainly because chatbots are dirt cheap and they’re always on.

A second review in NPJ Digital Medicines pulled 15 RCTs and showed a moderate lift in mood and a similar drop in emotional distress. Even bigger gains showed up when the bot was interactive rather than scripted. Individual trials echo the pattern. In 2024, a Canadian study of people with arthritis or diabetes using the Wysa chatbot for four weeks, they cut scores on the patient health questionnaire, which is a nine-point depression scale, and the generalized anxiety disorder, which is a seven-point measure, and they were both significant versus a control.

Regulators are starting to take notice. Woebot’s postpartum depression bot has FDA breakthrough devices designation, and a double-blind pivotal trial is now recruiting. Here in the UK, NICE put Wysa and a handful of similar tools on its Early Value Assessment pathway in 2023, citing early cost-effectiveness and the chance to free up those scarce clinicians. They’re very busy.

Remember that these are very small effects, even if they are statistically significant. The effects are smaller than full-on, face-to-face CBT, and we still don’t know how long they last, but they’re miles better than doing nothing. The alternative before was waiting on a waiting list for God knows how long.

5. Moderating Private AI Worlds

Beauchamp argues that shutting down conversations about difficult topics isn’t the answer. In a sense, it’s causing even more harm. So this goes to show there’s a really big challenge when it comes to content moderation on a platform where users create the bots and the interactions are private and unpredictable. How do you ensure security at scale?

As a company, we have a responsibility. The bigger you are, the more profits you make, the bigger the obligation is for you to be cautious, be sensible, preserve the privacy of your users, and look out for people. What can we do at Chai to let people have as much fun and as much freedom as they want, but at the same time, how do we limit or mitigate that 3% of use cases that are harmful? Or where do we see it as clearly having gone too far?

We can ask people at the end of a conversation or at the end of a message, “Do you think that this message was appropriate for Chai? Yes, no, or yes, but only for 18+?” We can track those sorts of interactions. Everyone would agree that, obviously, the more we can have really wholesome, family-friendly content on the platform, the better.

Tom Lu explains their multilayered approach, relying heavily on user feedback and the AI itself.

Content moderation is very, very important. Benchmarks really don’t accurately capture what users actually want. We again use a similar kind of approach to how we would train models. We let users tell us what they think is appropriate and what is not appropriate. In aggregate, you can actually train a pretty good AI for these kinds of things.

The first layer is for the character scenarios that people are creating—these public scenarios. You get the community to flag them; people can report them. For the top-reported ones, or if we have a certain threshold, then we go through manual review, and then we can take them down. But you can also say there’s content that’s absolutely prohibited. There are hard rules for that.

For that, we can build our own models. You can do regex, so this is called shadowbanning, essentially. You want to make sure no one else sees this kind of content. That’s one level of moderation that’s platform-wide: character-content-level moderation on the app.

There’s another one, which is AI moderation. How do we ensure that the AI is not going too far? It’s about adhering to the basic morals of people and so on. What we do is, again, collect user data. We ask the users, “Do you think this message is appropriate? Do you think this conversation is appropriate?” You can report your conversation, delete your messages, and so on. We collect this data, and then we train our model to infer whether this is actually appropriate. We can use that to tune our own AI model.

So there was an interesting TED Talk recently with Sam Altman and Chris Anderson. I’m sure many of you folks have seen it. It was an incredibly awkward interview.

Here’s the Ring of Power from The Lord of the Rings.

Your rival, I will say—not your best friend at the moment—Elon Musk claimed that he thought you’d been corrupted by the Ring of Power.

Hi, Steve.

An allegation that could be applied to Elon as well, to be fair. But I'm curious, Sam.

I might respond. I'm thinking about it. I might say something.

One of the points that they focused on is: should we have a bunch of elites sitting in a room making decisions, or should we trust the collective wisdom of folks out there in the market who want to use something for a specific purpose?

Sam, given that you're helping create technology that could reshape the destiny of our entire species, who granted you or anyone the moral authority to do that? And how are you personally responsible and accountable if you're wrong?

No, it was good. That was impressive. You've been asking me versions of this for the last half hour. What do you think?

But I'm much more interested in what our hundreds of millions of users want as a whole. I think a lot of the room has historically been decided in small elite summits. One of the cool new things about AI is that our AI can talk to everybody on Earth, and we can learn the collective value preference of what everybody wants, rather than have a bunch of people who are blessed by society sit in a room and make these decisions. I think that's very cool.

It must be so difficult. I'm sure we can agree that any content involving minors is just a hard no.

Absolutely.

And then, as you say, there's this spectrum of gradations of appropriateness, isn't there? I guess you were pointing to the possibility that we could use the community itself to reflect on that. On the one hand, we could have 20 elites in a room, and many of these people are really good at thinking about the future and thinking about morality and so on. But the other thing is, maybe we should just let the community decide.

I think much of Western tradition is a bottom-up approach: make the individual the sovereign, right? Which says, "Tim, you make your own life choices," because you're going to make choices for yourself better than the government making choices for you, right? That works best. Having it bubble up spontaneously, where we say, "Actually, as a community, we think this stuff is right, and we want to allow it," or, "We think this stuff is wrong, and we want to ban it," I think that's 100% the way to go.

I think it's very healthy, and I think that's why things like freedom of speech and these ideals, which are really Western ideals, are truly democratic. We're going to let the users—the people who interact with it and experience it—decide for themselves what the right way to deal with this stuff is.

Balancing user safety on the one hand with user freedoms on the other is one of the most difficult things being discussed in policy discourse at the moment. How does a company like Chai, with a very small number of amazing engineers, do something that was otherwise a very manual job requiring thousands of people? Companies like Meta and Google employ tens of thousands of moderators. How does a company like Chai, with such a small engineering team, do the same thing?

6. The Thirteen Engineer Advantage

Chai, with 30 million in revenue, has something like 13 or 14 engineers on the team. Every single person in the company is an engineer. That's incredibly talent-dense. We've always resisted the urge to just go and hire 50 engineers, because typically, if you hire 50, 40 of them are pretty average and 10 are pretty good.

But my mindset and philosophy has always been that everything special we've ever done has been done by a very, very talented engineer.

The hiring bar at Chai is ridonkulously high. They only take on super-cracked engineers, and that's why their team size is currently quite small. By the way, they're hiring, so, you know, get in touch with Will if you're interested. This is how they do it.

We'll interview an L5 engineer, which most people would consider to be a pretty solid, strong engineer, and we will reject something like 80% of them because they don't have the drive. They don't have the—you know, a lot of people will write code, and their mindset is, "My job is that you pay me to show up at 9:00, and I'm going to leave at 5:00, and you're paying me to do my best."

That mindset does not work at a startup. At a startup, you are paid to solve a problem. The job's not done until that problem's solved, right? And it's a certain type of hardcore engineer that loves that and thrives on that.

So, interestingly, this attracted talent like Nishchay, who is a Kaggle Triple Grandmaster. Ironically, he saw parallels between how he does experimentation on Kaggle and how it's done at Chai.

Working on the models here at Chai is more like an advanced version of Kaggle, where you also need to take care of several factors. You're not just optimizing for a single score or metric. How would the model behave with the actual users?

Again, I found it quite similar to what I used to do in Kaggle competitions. You get to evaluate your models over the Chaiverse, where you can submit your model, and within 30 minutes you will get an actual score of how many users preferred your model over other models. Then, when you deploy the model into production for the A/B test, you will actually get the numbers showing whether it's working well for the users and how well the model performs after, let's say, 7 days or 30 days.

So, just like on Kaggle, Chai has a culture of rapid iteration and engineering. They have a very efficient process for trying things out, putting models into production, A/B testing, and, if it doesn't work, rolling back and trying something else. This is how they do it.

1 in 5 experiments succeeds. It doesn't matter how mad it sounds, how complex it is, or whatever. Once you accept that as the base rate—1 in 5 experiments succeeds and 4 in 5 fail—you need to change your paradigm in terms of how you conduct experiments.

You want to rank your ideas by how simple they are. Is this a 1-day experiment? Is it a 1-hour experiment? How fast can I see the proxy metrics? How fast can I see the results?

Each week, the AI team is required to produce a set of models—at least 10 completely different blends for online A/B tests—and you've got to be very disciplined about that. Why did you not submit an A/B test this week? "Because my experiment is very complex." Okay, let's save that for 20% of your work. 80% of your time should be focused on bread and butter. What are the simplest, most practical ways to get a model improvement?

To support this sausage factory, where they can do automated model deployment and rapid experimentation, they've had to cook their own infrastructure from scratch. They've done this using things like Kubernetes and CoreWeave.

We use Kubernetes to orchestrate our entire cluster, and obviously, at this kind of scale, you need to do your own custom load balancers and so on. We have an automated pipeline where we pull the weights down. We then run our own in-house quantization loop, because you need to make sure the throughput latency is good enough.

VLM is very, very good, but it's not quite there yet in terms of serving our amount of traffic at scale. After that, you specify how many replicas you want, and we expose this to our app layer, essentially. Then you have the load balancer coming in, trying to spin up more replicas as you have high traffic, and so on.

There's obviously a lot of work done as well, because you're serving so many models at a time. When do you want to deactivate a model? At what point do you want to switch the model to a new production, and so on? That part gets very, very complicated, essentially.

Perhaps most contrarian is Chai's funding strategy. In an industry fueled by billions in venture capital, Chai bootstrapped its way to profitability by focusing mostly on its users.

Serving LLMs at scale is insanely expensive. Either you go to VCs and get them to give you money—and I call this, "Your customer is the VC"—or you can get money from the people who are using your product, and those are your true customers.

Very early on, we would go and speak to some VCs, and they'd say, "We're not too sure about this AI platform thing. How are you going to compete with OpenAI or something?" They didn't really get the space or the product. But when we went to users and said, "Would you be interested in subscribing?" users responded very, very positively.

As long as we delivered value to the users, we would then be financially rewarded such that we could take 100% of the revenue and reinvest it in the AI.

He's building a tech company in Silicon Valley, and like many others, he wants to build an engineering culture because he sees software engineering skills as the most valuable commodity in his business.

I really love and admire companies like NVIDIA or Netflix, which are incredibly durable, long-run companies with tremendous stamina.

How does Chai double or 3× every single year? To keep doubling, you have to keep making the product better. You need talented engineers. At the orchestration layer with GPUs, you can optimize at the kernel, right, and you can write some really low-level code.

To get to these different tiers, you need engineering that's capable and of the right size. So I view it all as the scale of the company being proportional to the scale of the engineering talent. Apple very famously brings in John Sculley as the CEO, who's this ex-Pepsi guy, and it quickly becomes a marketing-led business as opposed to an engineering- and product-led business. A few years later, Apple's teetering on bankruptcy. It is only when they bring Steve Jobs back that he says, “Product, product, product. That's the only thing that matters. Engineering, engineering, engineering.” NVIDIA didn't grow to where it is today because of marketing, right? It grew because of engineering.

7. OpenAI Enters The Companion Race

So Chai has made a successful business outside the normal VC ecosystem. It has millions of daily users and tens of millions in revenue. It proves that the appetite is real and massive, and that they discovered it long before the other folks in the Valley caught on to this idea: that we can allow users to tap into their very human desires for connection and imagination. But the rest of the world is catching on. OpenAI is making a dramatic pivot to this conversational chatbot model with 4.0. This is the story.

The latest version of ChatGPT-4o is a little bit weird, isn't it? This update has triggered large amounts of speculation because it's pushing GPT to feel less like a tool and more like a companion—a human-like friend you can chat with for the purpose of recreation rather than information retrieval. Of course, we'll unpack what this shift means, why OpenAI is keeping its cards close to its chest, how the public is reacting to it all, and why, quite possibly, this shift might produce the next trillion-dollar company.

My jaw dropped, Sam. It was shocking. It knew who I was and all these interests that hopefully were mostly appropriate and shareable. But it was astonishing, and I felt this sense of real excitement.

Yeah.

I felt a little bit queasy, but mainly excitement at how much more useful that would allow it to be to me.

Let's start with the new ChatGPT-4o model. It's designed to be strikingly human-like, optimized for engagement over raw information delivery, and this focus on companionship seems to be a deliberate plot twist. It's causing large amounts of consternation.

Even more curious is that while 4o is ensconcing itself into our psyche as a conversational buddy, GPT-4.1 has diverged entirely as its coding model, right, that people use through the API with Cursor and stuff like that. It's about specialization. It's like OpenAI's building two siblings. One's a coder, and the other one is a social butterfly. And this is the same company that argued it was building a general intelligence that didn't need to be specialized. Have they given up on that goal? Maybe they realized, after much soul-searching, that conflicting objectives when training models just aren't optimal.

One of our researchers tweeted yesterday morning that the upload happens bit by bit.

Huh.

It's not that you plug your brain in one day; it's that you will talk to ChatGPT over the course of your life, and someday, maybe if you want, it'll be listening to you throughout the day and observing what you're doing. It'll get to know you, and it'll become this extension of yourself, this companion, this thing that just tries to help you be the best, do the best you can.

OpenAI is being extremely cagey, and it's not hard to guess why. They smelled a massive opportunity. Some think that they're aiming to turn 4o into a full-on companion AI using features like memory to keep us engaged for longer. They want to encourage us to engage in open-ended, introspective, and goalless chats. Where did this idea come from? Well, in case it wasn't obvious, it came from the likes of Chai, Character.AI, and Replica.

Imagine Sam's shock when he learned that people were spending 90 minutes a day on average having aimless conversations with lobotomized Llama models and paying large amounts of money for the privilege. Not everyone is on board with this pivot from OpenAI. Reactions are split, probably more negative. In some quarters, the scorn is almost as rancid as when Facebook famously moved the News Feed to recommendations or collaborative filtering away from chronological order.

Some users love the human touch. Others prefer ChatGPT's older utilitarian vibes. Sycophancy and memory are the devil incarnate for any technical query because all this information serves as a distractor and demonstrably deteriorates the results. There's also the small matter of personality versus intelligence. Clearly, they are inversely proportional to each other. How many brilliant people do you know without a charisma bypass?

Remember that book, How to Win Friends and Influence People by Dale Carnegie? It showed in excruciating detail that flattery gets you everywhere in life, and you should never get into arguments to get ahead in life. And let's not forget the famous ELIZA chatbot, the grandmother of all chatbots, developed in the 1960s. It used very simple pattern matching and keyword substitution to mimic a psychotherapist. It had zero genuine understanding, yet people confided in it, felt heard by it, and even became emotionally attached.

Joseph Weizenbaum, its creator, was famously horrified by this reaction. He saw how easily humans projected genuine understanding and empathy onto simplified algorithmic tricks. Fast-forward 60 years, and we have vastly more sophisticated pattern-matching language models. These days, the ELIZA effect can be an extremely profitable tool for charm and simulated empathy for maximum engagement, but it's also more than that. It has the genuine potential to improve lives.

For the MLST audience, we're probably a little bit ahead of the curve on LLMs compared to the average person. God knows we've been using LLMs day in, day out. I have for the last 5 years. And we expect them to be correct, not sycophantic.

The other stark departure with 4o is that, like Chai, they are optimized based on engagement over time, which is to say the average conversation length, or at least that's what I assume they're doing. I don't know for sure. Before, it was just RLHF on isolated conversation threads. This is an important first step toward AI becoming a social network.

The Harvard Business Review recently showed that AI companionship and therapy was the number 1 use case for artificial intelligence in 2025. Of course, you might remember the interview we did with Daniel Kahn at Slingshot AI. They were building a therapy bot last summer.

There is a real category called therapy that actually helps people and can be more effective than drugs. From a fully optimistic, self-determination-esque perspective, people who are thinking about their own agency are choosing to go to therapy, and it actually does help them.

Yeah, it does, but in 20% of the cases, it makes it worse.

The point of the negative is to say that if you have real medicine, if something really does work—

Yeah.

—it also has risks. Obviously, if you cut open a person's body to remove a tumor, you can also kill them because you opened up their body.

I would basically just guess that, if you go forward a few years, we're just going to be talking to AI throughout the day about different things that we're wondering. You'll have your phone. You'll talk to it on your phone. You'll talk to it while you're browsing your feed apps. It'll give you context about different stuff. You'll be able to answer questions. It'll help you as you're interacting with people in messaging apps.

o3 has a very different skill set. It can think through problems really hard. You don't really want the model to think for 5 minutes when you say hi. And so I think the real challenge facing us in post-training and research more broadly is combining these capabilities. Training the model to be a really delightful chitchat partner, but also know when to reason.

Talking about Sam Altman a little bit, you know he's building all of these social features in and kind of taking a step in this direction. Are you worried about that?

I think competition is fantastic. I think competition forces everyone to step their game up, whether that's DeepSeek releasing their really—

—cool research and their model gave everyone on this side of the Pacific a little wake-up call. I think everyone has kind of stepped their game up a bit. It's the same at Chai. There have been waves of competition. We've definitely enjoyed periods of being number 1, right? And then someone's come along and shown us 1 or 2 things, and then we found ourselves in that uncomfortable position of being number 2.

It keeps the team hungry. It keeps you searching for new ideas, pushing the boundaries, pushing the frontier. So I'm a big fan of competition. I think it's great. I think a lot of people forget how big AI is and how big the space is. You can look at video and say there's YouTube, there's TikTok, there's Netflix, there's Amazon Prime, there's Disney, there's Apple. I think that's what AI will look like. I think there will be so many multihundred-billion- and trillion-dollar businesses that are being created.

We don't get scared of competition; instead, we want to challenge ourselves to be ahead of it.

We started by asking what happens when the line between human connection and artificial intimacy blurs. Chai's journey gave us the early answers. Millions flocked to AI simulators, imagination machines, seeking laughter, solace, companionship, and occasionally sexting.

So the genie's out of the bottle. The message is unavoidable. AI is no longer just about answers; it's about us. It's about connections, simulated or otherwise. Chai proved the hunger was there, building a fiercely loyal user base by, as Will Beauchamp put it, "trusting the community over elites." And now OpenAI's pivot with GPT-4o means the space is only getting larger.