1000x Users

I realized something about AI usage.

I probably use AI manually two hundred times a day, maybe more. If you include API calls from agents running in the background, the number is probably closer to a thousand. I am sure there are other tokenmaxers out there who do even more.

Then I saw some relatives in my hometown using AI. They use it maybe four or five times. A question here, a translation there, maybe help writing something. Useful, but not central to their day.

When you spend your time on X/Twitter, you do not realize this gap. But in the real world, for most people, AI is just a replacement for their daily Google searches.

This made me think about ChatGPT / Gemini announcements about, say, 900M weekly active users. Well, if one weekly active user has 3 searches, and another has 10,000, that is clearly not a good comparison. If anything, comparison should be based on weekly tokens spent.

It is true that tokens are not a direct measurement of productivity either. But the active user concept is going to become less valuable as time goes by, because one active user could have 100 agents.

This brings up another important point. Previously, there was a limit to how much leverage a user could have. I might do 100 Google searches, and someone else might do 5. I am 20x him.

Now with AI agents, I think I may be 1000x him. Do you see where I am going? AI and agents are adding incredible leverage to heavy users.

Who are the heavy users? Coders are one example. Then I thought about my surroundings and who the heaviest users are. They are the people with the most activity in their lives.

By activity, I mean people who are trying to get things done. Business owners, managers, perhaps busy mothers. The more things going on in your life, the more you leverage AI, because it helps you start a thread, an initial motion, on a dozen projects you have on your list.

Personally, I am currently on a Delta flight, writing this. And I still have to write this myself, because I just cannot like what AI does. Yet in parallel, I have an open-source project for my Google Chrome workflow, a deck for my company, some research on investing and Microsoft, and pretty much 20 other things.

So if you are someone who is trying to do a lot, you have infinite opportunities to use AI. And guess what, the start is just a prompt. It is very low effort.

I think this has interesting implications for markets, users, and labs. First, people who were already more agentic will use AI more. They simply have more needs for it.

Second, they will get more leverage from this technology, which will make them more productive and rich. We have already seen this with Anthropic. Technology does seem to increase the overall pie, while also allocating a bigger portion of it to a few individuals.

Third, weekly or monthly active users do not matter as much as weekly token spend or tool calls, because those are better metrics for understanding true usage. A single user may be worth 1000x another.

In that sense, you can see that Anthropic was so much smarter than everyone else. They bet on coding before anyone, and they are reaping the benefits. OpenAI realized this, and now with Codex they are close followers too.

Google, unfortunately, is still facing the innovator’s dilemma. They need cheap models for their billions of users, but those billions of users may soon become less relevant if another set of 10 million users has demand equal to 10 billion people.

Another point about heavy users is the type of queries they ask. Most people who use single-shot AI are replacing simple Google queries. In fact, for those, I go to Google too. AI Overviews are super fast and nice. But they are just informational questions, and I do not add much value to Google’s underlying models.

Whereas with Codex and Claude, I am trying a dozen different ideas. I scream at it, I swear at it, I show where I am unhappy, and I give so much data about what I am trying to achieve. All of this is going through the API. OpenAI / Anthropic are seeing what I, and a million other heavy users, are trying to achieve. They see where the models fail.

Then they take that and improve the next iteration. And so it goes, and so the gap widens. I want to be wrong about this, but perhaps this is why Google could not catch OpenAI / Anthropic for a while. Their coding models are simply not as good.

Hence the point of this write-up. We are going into a world where 1000x users, 10x engineers, or 100x salespeople will matter more and more. You have to build for those people because they can increase aggregate demand by creating more agents every day.