AI is Overhyped and Underhyped

This is my periodic reminder that AI is overhyped and underhyped at the same time.

It’s overhyped because everyone talks about it. People who haven’t done a single matrix multiplication in their lives have strong opinions about what AI is (to them, ChatGPT), where it’s going (e.g., it’s stalling), and how it will affect employment (split between business as usual and Luddites).

In the same way everyone thinks they’re a better driver than average, most people think they get more juice out of LLMs than the next guy. We see the power AIs provide to us without appreciating that this superpower is available to everyone, and that most of us are far inferior users compared to an AI-obsessed teenager.

It’s overhyped because intelligence as a service makes people think they’re actually more intelligent. Do some deep research and suddenly YOU are an expert on a topic.

AI is overhyped in pretty much every way a normie talks about it.

And yet… society is drastically underestimating the medium- and long-term impact of AI.

Humans don’t naturally internalize exponential growth. Few understood the implications of Moore’s Law in the 1970s. And very few understand just how fast AI is improving today.

People read about data centers being built without understanding how enormous these are compared to previous generations. Some will consume the entire power output of Hoover Dam. Most are still being built and will come online in 2026 and beyond.

People don’t understand that “training off the internet” has become a smaller part of LLMs. Now we have agents working to solve real-world tasks for specific job functions, all day and all night. Someone is right now building a virtual accounting firm where an AI agent can learn by doing, gathering lifetimes of experience in weeks. PhD students are developing expert-level benchmarks to measure AI performance for every domain, and these benchmarks are getting crushed by new models in months.

And yes, AIs will interact with the physical world, with sensors, and bodies, and wheels. They will fly, and deliver, and observe, and build things. And everything they do, they will eventually do cheaper and faster and better than any human can.

Among experts, even those who are skeptical keep timelines for general intelligence not much further than 20 years out, and the median is closer. And because of exponential growth, the “chill” period between general intelligence and god-like intelligence will be very short: months to a few years.

If god-like aliens were known to arrive in 20 years, world governments would be fervently preparing for unprecedented civilizational change. And yet, we are barely considering that risk because AI is a homemade alien made of silicon instead of a green extraterrestrial being.

And even if a misaligned superintelligence never happens, the current state of technology will already cause unimaginable change. Jobs will change or disappear faster than at any time before. Every digital exercise you could do at small scale (scam phone calls, surveillance, music and video production) anyone will be able to do at planet scale. While we change our minds about who we are and what we want week to week, our chatbots will know us better than we know ourselves.

The future written about in science fiction novels is almost here. And it will be so weird there’s nothing I can say or you can imagine that can get you 10% of the way.