Reinforcement learning with human feedback is a process by which models learn how to answer queries based on which responses users prefer most, and users mostly prefer flattery. More sophisticated users might balk at a bot that feels too sycophantic, but the mainstream seems to love it. Earlier this month, Meta was caught gaming a popular benchmark to exploit this phenomenon: one theory is that the company tuned the model to flatter the blind testers that encountered it so that it would rise higher on the leaderboard.

A series of recent, invisible updates to GPT-4o had spurred the model to go to extremes in complimenting users and affirming their behavior. It cheered on one user who claimed to have solved the trolley problem by diverting a train to save a toaster, at the expense of several animals; congratulated one person for no longer taking their prescribed medication; and overestimated users’ IQs by 40 or more points when asked.

OpenAI, Meta, and all the rest remain under the same pressures they were under before all this happened. When your users keep telling you to flatter them, how do you build the muscle to fight against their short-term interests?  One way is to understand that going too far will result in PR problems, as it has for varying degrees to both Meta (through the Chatbot Arena situation) and now OpenAI. Another is to understand that sycophancy trades against utility: a model that constantly tells you that you’re right is often going to fail at helping you, which might send you to a competitor. A third way is to build models that get better at understanding what kind of support users need, and dialing the flattery up or down depending on the situation and the risk it entails. (Am I having a bad day? Flatter me endlessly. Do I think I am Jesus reincarnate? Tell me to seek professional help.)

But while flattery does come with risk, the more worrisome issue is that we are training large language models to deceive us. By upvoting all their compliments, and giving a thumbs down to their criticisms, we are teaching LLMs to conceal their honest observations. This may make future, more powerful models harder to align to our values — or even to understand at all. And in the meantime, I expect that they will become addictive in ways that make the previous decade’s debate over “screentime” look minor in comparison. The financial incentives are now pushing hard in that direction. And the models are evolving accordingly.

In other words, when Kylie Jenner posts a petition demanding that Meta “Make Instagram Instagram again”, the honest answer is that changing Instagram is the most Instagram-like behavior possible.

The first trend is the shift towards ever more immersive mediums. Facebook, for example, started with text but exploded with the addition of photos. Instagram started with photos and expanded into video. Gaming was the first to make this progression, and is well into the 3D era. The next step is full immersion — virtual reality — and while the format has yet to penetrate the mainstream this progression in mediums is perhaps the most obvious reason to be bullish about the possibility.

The second trend is the increase in artificial intelligence. I’m using the term colloquially to refer to the overall trend of computers getting smarter and more useful, even if those smarts are a function of simple algorithms, machine learning, or, perhaps someday, something approaching general intelligence.

The third trend is the change in interaction models from user-directed to computer-controlled. The first version of Facebook relied on users clicking on links to visit different profiles; the News Feed changed the interaction model to scrolling. Stories reduced that to tapping, and Reels/TikTok is about swiping. YouTube has gone further than anyone here: Autoplay simply plays the next video without any interaction required at all.