Happy Friday everyone. In this issue: Nepal elected its newest PM in a gaming chat app, we briefly had a new richest guy, and the blurred lines between AIs that want to save you time and AIs that want to date you.
The Spectrum of “AI Companion” Companies
This segment is more thinkboi than is typical for the letter, so you’re welcome/I’m sorry for all of the inside baseball.
I spent much of Tuesday’s letter talking about Friend, the wearable AI companion that wants to be a somewhat confused combination of your bff and your sigma grindset contrarian life coach. This week also saw the launch of another AI companion, Poke, that you communicate back and forth with via text message (there’s no hardware device involved). Here’s their launch video:
It’s not bad! The AI is portrayed mostly as a handy sidekick, and there’s a lot of emphasis on how it's present right when it’s needed or helpful, and otherwise invisible. The story is about how the product enables human experience and connection. No psychosis here. Until the end of the video. The climax shows the AI explicitly disobeying the user in an attempt to convince them that their new girlfriend is worth… staying in France forever? This is meant to be humorous, but to me it highlights an increasingly common pitfall for companies building products with AI personalities attached to them: confusing usefulness with companionship.
When looking at products like Poke, Friend, or even Claude Code, I visualize them on a spectrum: at one end, you have a pure “productivity assistant” that’s focused on doing work for you, and at the other you have a companion whose main purpose is to be your always-on best friend/confidante/boyfriend/girlfriend. The former is usually billed as some sort of “superhuman executive assistant” that reads your email, calendar, and text messages and can do everything from calling you an Uber to planning an entire itinerary for your next vacation (booking flights and AirBnbs, deciding where you should eat, ordering some new beach fits, etc.) without requiring you to do more than occasionally thumbs up a message. The goal here is “giving you time back” and thus increasing your productivity (it’s debatable whether removing friction is always a positive, but it sounds nice). The latter category—AI friends—is much more likely to pull you away from other activities, including human socializing, because the success of the product is determined by how much time you spend with it. More is better, and those minutes have to come from somewhere.
Here’s the Context Rot AI Companion Spectrum Informational Graphic:
My working theory is that products at the far ends of the spectrum will be the ones that succeed. A “super assistant” that can handle tedious things like making DMV appointments for you is something many people would pay for, and you don’t need to have a personal relationship with it for it to be successful. Even better if it’s confined to your work life—there’s plenty of drudgery to automate away at most jobs, and the framing of treating an AI agent like a piece of work software (raise your hand if you want to be friends with Microsoft Teams) should help mitigate the more harmful effects of humanizing the assistants. No one I know is trying to date Claude Code.
Products on the other (“friend you are probably dating”) end of the spectrum are already pretty successful. I generally don’t think they're a net positive for society, mostly for the above reasons that they succeed by taking your time and attention away from human relationships, and can lead to weird expectations around constant affirmation and emotional coddling (not to mention the weird romantic shit). But people keep building them, and people keep using them, so they’re probably not going anywhere.
This brings us to the middle: the Bog of Mixed Signals. This is the trap Poke is close to falling into. I think this is a no man’s land for AI products. No one is really asking for their email software to start giving them unsolicited life coaching, and attempts to occupy that space in someone’s life feel both misguided (it will turn off more users than it attracts) and also psychologically unhealthy. We’re already seeing that humans naturally anthropomorphize AI, which means we subconsciously pull all of these tools towards the right end of the spectrum over time (see: all the people in a relationship with ChatGPT). The companies building it shouldn’t encourage that.
This brings us to my second theory: AI personality is just UX. If your AI product has a good personality, people will enjoy using it more, just like how people are more likely to stick with a well-designed and fun-to-use mobile app than a clunky or aesthetically unappealing one. But you have to be selling something. Products on the left are selling efficiency—they’re creating code, searching the web, booking appointments, organizing and sending emails, the list goes on. They’re helping people do things that humans already do, just faster and with substantially less effort. Products on the right are selling companionship. It’s easy to let the smoke of a computer-that-is-maybe-a-person get in your eyes, but at the end of the day, the product is not the personality. And if it feels like it is, you’re probably selling companionship.1
Right now, the more you try and appeal to both ends of the spectrum, the more confused your product will feel. We may get to a point where something resembling AGI shows up that can occupy the entire line (that’s where the AI in Her would fall), but until then, if you spend time trying to make your AI email assistant horny or pretending that your AI human friend replacement is “useful,” it will only make your product worse. The personality of an AI agent is worth improving in the same way that making an app feel nice to use is worthwhile, but UX is not a moat, and usefulness tends to win out in the end.
Nepal Used Discord to Choose Its New Government
Nepal seems to have successfully overthrown its government, and it did so in a very online way. Young Nepalis (mostly Gen Z) began protesting the government earlier this year, and because they were organizing on social media, the government responded by trying to ban it. This went about as poorly as you’d expect, and quickly led to violent clashes between protestors and police. The unrest culminated with the Parliament building being lit on fire, the prime minister ousted, and more than 30 people dead.
It was at this point that the activists increasingly began using a public Discord server to make political decisions. Discord is a platform predominantly used by gamers and Very Online Communities—it was hugely popular with NFT collectors—and is known for turning into kind of a chaotic mess whenever a community grows past a few thousand members (everyone is basically just in a big group chat together). It’s certainly not a place you’d expect to host a national convention.
That, however, is exactly what’s going on in Nepal now. The server has over 100,000 members, and was at various points streamed live on national news sites. The Nepali army recognized its legitimacy, and said that they would meet with whoever the Discord members nominated as an interim leader. After a series of public polls (voting is usually done via emoji reactions), Sushila Karki, the former chief justice, was selected as the new government’s representative.
Discord is definitely not a sustainable way to govern, well, anything, but it’s fascinating to see it successfully used as a platform for a democratic political process (even if imperfect). Hopefully this leads to peaceful, official, elections in Nepal, and the return of Discord to what it’s best used for: being the place you meet up with friends online to hurl obscenities at 12-year-olds while playing Call of Duty.
Doomscrolls
Anthropic is having a good week. Microsoft, who has invested billions in OpenAI and can use their technology for free, found that Anthropic’s Claude models perform better at accomplishing tasks in Excel and PowerPoint—so much better that Microsoft is paying to use them instead. Anthropic is also rolling out the ability for Claude to create and edit files. It can generate and revise PDFs, slide decks, and other documents. Apparently it’s pretty good.
Tucker Carlson interviewed Sam Altman. This interview is truly wild—Tucker spends about seven minutes pressing Sam on the death of an OpenAI whistleblower. Overall, Sam remains quite composed and fairly candid in the face of some of the most direct and difficult questions I’ve seen him asked in an interview setting.
OpenAI signed a $300 billion contract with Oracle. The deal goes into effect in 2027 and will last for five years. Notably, OpenAI made $10 billion last year and is estimated to pull in around $20 billion this year, so those numbers better start going up if Oracle expects the checks to clear. The deal caused a massive jump in Oracle’s stock price and briefly made Larry Ellison the world’s richest person before Elon reclaimed the spot a few hours later.
Paramount is looking to buy Warner Bros. Discovery. Speaking of Ellisons (The CEO of Paramount Skydance is David Ellison, Larry Ellison’s son), Paramount is preparing a “majority cash” bid for Warner Bros. Discovery that would cover all of its media properties (HBO, Warner Bros. Pictures, CNN, roughly 1 million cable channels). The Context Rot Editorial Board would love to see David Zaslav in charge of fewer things (ideally zero things) and therefore endorses the deal.
Meta’s most elite AI researchers aren’t even on the internal org chart. They also sit in a part of the company’s HQ near Zuck’s desk that requires special badge access, and all of this is making other Meta employees feel Weird. The entire story around this latest iteration of the company’s AI efforts really makes it seem like Meta is facing a hugely uphill battle, both technically and culturally.
Robinhood is launching a social trading feed next year. Users will be able to share trades and post about their strategies and predictions. The UI in the teaser focuses on a “For You” feed, and it’s striking to me that “adding social” to an app no longer means seeing what friends are doing, but rather that a dopamine-maxing influencer-engagement-rewarding slot machine algorithm is coming. Casino culture rules everything around me.
Lossless audio is coming to Spotify premium. Spotify claimed it was “coming later this year” back in 2021, but this time they’re really truly for real they promise.
This Happened to My Buddy Eric
All em dashes inserted organically by me. All typos were chosen by a democratic process conducted within the Context Rot Discord server.
Thanks for reading. If you enjoyed this post, please share it with a friend.
There’s a caveat here that you may also be selling entertainment, but most of these products are entirely interacted with one-on-one in private spaces, which feels more like companionship than media or entertainment to me, at least for now.