The AI Race Is Really a Race for Your Habits


Follow the AI companies for one month and you hear two stories that do not fit together.

The first is a warning. Earlier this month, Anthropic's chief executive, Dario Amodei, published an essay titled "We Must Pace the Frontier." In it he wrote that "a race to the bottom, spurred by commercial incentives, can make these risks more acute," and that "we must slow the pace at which we improve the capabilities of AI models."

The second is a release calendar. On September 2, Google introduced Gemini 3.8 Flash, describing it as "our third Flash release in only six weeks." On September 8, Meta introduced Muse, a personal AI agent that works inside WhatsApp. On September 21, SpaceXAI released Grok 4.7. On September 22, Anthropic released Claude Opus 5.5, and its own announcement said plainly that it was "our first release since we called for pacing the frontier."

The warning and the shipping are happening at the same time, sometimes inside the same company.

It is tempting to read that as hypocrisy, or to read it as a simple contest to build the smartest model. Neither explains why these companies are spreading their products so widely, giving so much of it away, and working so hard to live inside other companies' software.

Here is the view of this post. Much of the AI race is not about who builds the best model. It is about whose product people learn to work in first. A better model can be copied within months. A habit cannot.

One Month of Releases, Read Differently

Look at what each of those launches is actually trying to do, and the benchmark scores start to look like the smaller part of the story.

Google said Gemini 3.8 Flash arrives "at the same speed and low cost" as the version before it. Speed matters most for the small request you make ten times a day, not the big one you make once a week.

Meta said Muse was designed around the way people already communicate, so that "talking to it works just like messaging another person, in the Muse app or directly in WhatsApp." It also said Muse was built for billions of people "so there's no learning curve."

OpenAI's model guide sorts its GPT-6 family by the kind of work you bring: GPT-6 Astra for complex reasoning and coding, GPT-6 Sol to balance intelligence and cost, GPT-6 Luna for cost-sensitive, high-volume workloads. Whatever the job, there is a door for it in the same building.

SpaceXAI said Grok 4.7 is "available today in Cursor," a coding editor made by another company, and through third-party coding harnesses and model routers.

Anthropic's announcement led with long, difficult work: one early tester completed a 680,000-line code migration in less than a day.

Five companies, five different bets on the same question. Not "which model is smartest," but "where does a person's daily work start, and how do we get there first."

What a Habit Does After the Comparison Ends

This blog has covered the general mechanics before: how a default wins without ever entering a competition (The Products You Never Chose), and how the price of leaving is the one nobody quotes (Switching Costs — The Price of Leaving Is the One Nobody Quotes). The short version is that people compare products when they choose them and then mostly stop comparing. Someone with a deadline opens yesterday's tool and continues yesterday's work.

Relationships make this stronger. When researchers tracked 136,009 users who moved from Twitter to Mastodon, they reported "the strong influence of the social network in platform migration." People did not move because a feature was better. They moved, or stayed, largely because of who else was there.

I write from Korea, where this lesson is part of everyday life. For most of my adult life, the search habit the people around me formed first was not Google. It was Naver, with its own blogs, forums, and question boards built around it. A global company with far more money spent years as the challenger here. Being used and being chosen are different things (Being Used Is Not the Same as Being Chosen), and a habit can protect the first long after the second has moved on.

What is new is what happens when the product is an AI.

Why AI Habits Go Deeper Than Most

A search engine or a social app teaches you where the buttons are. An AI tool you use for work teaches you a way of working.

At first, it is a text box. You type a question and read an answer. Comparing products at this stage is easy; you paste the same question into three of them. The better answer wins, and people move quickly.

Then the use changes. You learn what to tell it first. You learn how much of a task to hand over and how much to keep. You learn how to check what comes back. You save the instructions you repeat, set up projects, connect files and outside tools. None of that shows up on a leaderboard, and all of it has to be rebuilt if you switch.

Inside a company, it goes further. Someone decides what data can go in, who reviews the output, and which actions need a human sign-off. Staff get trained on one way of working. At that point AI switching costs are no longer about an app. They are about a workflow, a policy, and a team's shared habits.

So when a rival announces a better model, the real question for a working user is not "is it better?" It is "is it better by enough to relearn everything I just learned?" If the gap is large, people move. If the gap is small or hard to measure, they stay, and the first company keeps them without having won the comparison.



Who Actually Owns the Habit

There is a complication here that is easy to miss, and it matters for anyone trying to read which company is really ahead.

When Grok 4.7 runs inside Cursor, a developer may do a great deal of work with it. But the habit that developer forms is a Cursor habit. If Cursor swaps the model underneath next month, the developer keeps working in Cursor. The model company served the requests. The tool company kept the relationship.

It is a little like an index fund, where the fund holder owns the result but a committee decides what goes inside (Somebody Chooses the Index). In a tool that routes work to different models, someone other than the user is often choosing the model.

That means there are really two races running at once. One is to supply the best model. The other is to own the place where people use models. A company can win the first and still lose the second. When anyone tells you a model's usage has grown, the useful follow-up is: grown where, and chosen by whom? A user who keeps selecting your model by name is a different asset from a user whose tool happened to route a request your way.

The Loop That Makes Early Users Worth More

Repeated use gives a company something besides revenue.

Real users bring real problems. Test questions do not show you how a summary drops the one sentence a reader needed, or how working code quietly changes a feature nobody asked it to touch. People who use a product every day run into exactly those failures. Where they got stuck, what they had to correct, and which results they actually kept are clues about what to fix next.

So the accumulation runs in two directions. Users get more fluent with the product, and the company gets a clearer picture of what real work demands. If the improved product then gets used more, the head start compounds. A company that has to win every customer again with each new model is competing under different conditions from one whose customers are already working inside its product.

One limit belongs here, because this idea is often stretched too far. Heavy use does not automatically mean every conversation becomes training data. Remembering things for a user, analyzing how a service is used, and training a model are separate activities under separate policies. OpenAI, for instance, states that since March 1, 2023, data sent to its API "is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us)." Volume is also not the same as quality. The loop is an opportunity to improve, not a guarantee.

Why the Warnings and the Launches Coexist

Put the pieces together and the opening contradiction becomes easier to read.

If a company pauses to test further, its rivals keep shipping. The people using those rivals' products keep learning their habits, connecting their tools, and building their records. The longer the pause, the stronger the reason a future user would need in order to come back. Waiting costs a company more than this quarter's sales. It costs a seat at the table while everyone's working habits are being formed.

That pressure exists alongside whatever safety concerns are sincerely held. The second limit of this post is here: a new release is not, by itself, proof that a warning was insincere. At the AI Seoul Summit in 2024, the UK and South Korean governments announced that companies including Amazon, Anthropic, Google, Meta, Microsoft, OpenAI, and xAI, along with Korea's Naver and Samsung Electronics, had agreed to the Frontier AI Safety Commitments. Those commitments ask each company to set thresholds at which severe risks would be deemed intolerable, and, in the extreme, not to develop or deploy a model at all if risks cannot be kept below them. They do not ask anyone to stop releasing models.

So the useful question about any launch is not whether it happened. It is what the evaluation found and whether it could have changed the decision. Anthropic's Opus 5.5 announcement, for example, names outside evaluators, including Frontier Design and METR, that tested the model before release. That is the kind of detail worth looking for in every company's releases, because the competitive pressure described above is strongest precisely where those checks matter most.

Familiarity Gets You In. Trust Gets the Work.

The stakes rise as AI moves from answering to acting.

A wrong explanation can be checked. A wrong email or a wrong purchase leaves a mark. Meta's Muse, by the company's own description, comes back for approval "before it sends an email or makes a purchase." That design choice is an admission of where the risk lives.

Familiarity helps a product survive small annoyances. Trust is what gets it the important tasks. They are different, and the second can undo the first. A user who gets fast but unreliable answers learns a habit too: never use it for anything that matters. A product that repeatedly breaks something important gives its users exactly the reason they needed to pay the cost of leaving.

A company that wants a durable user base needs both. Getting there first earns the habit. Only reliability keeps it once the work gets serious.



What the Leaderboards Leave Out

Anthropic's own release contains a sentence worth noticing: "at these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences." If the builders themselves say the scoreboard is losing its signal, it is worth knowing what else to look at. None of the following requires a view on which company wins.

  • Do trial users come back? Sign-ups and monthly visitors say little about whether a product became part of someone's work. AI user retention, measured over months rather than days, says much more.
  • Does use move from questions to tasks? A product people ask things is different from a product people hand work to.
  • Who chooses the model? Usage inside another company's tool is worth less than usage from people who pick your model by name.
  • Do failures decline? "New features" matters less than whether tasks that used to fail now finish, and whether users spend less time checking.
  • Can the habit travel? If saved settings, history, and connected tools become easy to move between products, the advantage of being first shrinks, and users can respond to quality more quickly.

Public launch posts do not answer most of these. That is a reason to watch where habits are forming rather than to name a winner. The economics underneath, with AI costing more like a factory than like software (AI Looks Like Software. It Costs Like a Factory.), make those habits even more valuable to whoever holds them.

Before the Perfect Model Arrives

The AI companies are racing to build the most capable model someday. They are also racing over something much smaller, happening today: which program people open first thing in the morning, and how they have learned to hand it work. The two races feed each other. Today's use creates the chance to improve tomorrow's product, and a better product holds on to today's users.

The unsettling part of familiarity is that the choices keep adding up even after people stop making them. Every time someone opens yesterday's tool, a little more of their work settles there.

People are learning to work with AI long before any perfect AI exists. Whose product they learn it in will shape the market that follows, and that, more than any single benchmark, is why companies that warn about the risks still find it so hard to slow down.

Data as of September 25, 2026. Sources: Dario Amodei, "We Must Pace the Frontier" (September 2026); Anthropic, "Introducing Claude Opus 5.5" (September 22, 2026); Google, "Introducing Gemini 3.8 Flash and 3.8 Flash Cyber" (September 2, 2026); Meta, "Introducing Muse" (September 8, 2026); OpenAI API documentation, "Models" and "Data controls in the OpenAI platform"; SpaceXAI, "Introducing Grok 4.7" (September 21, 2026); UK Government, "Frontier AI Safety Commitments, AI Seoul Summit 2024"; "Flocking to Mastodon: Tracking the Great Twitter Migration" (arXiv:2302.14294, 2023).

Related Posts:

Visuals on this post are AI-generated. The author works with AI as a research and drafting assistant; topics, judgments, and final edits are the author's own. This post is observation, not investment advice. See full Disclaimer for details.

Popular posts from this blog

SK Hynix's ADR Listing — The Split That Isn't a Split

Image
A stock can be expensive in two very different ways. It can be expensive because the business is overvalued — the price has run ahead of what the company is worth. Or it can simply look expensive because a single share is a large object, a big ticket a smaller investor has to swallow whole. Those are not the same problem, and they do not have the same solution. SK Hynix is now handling the second kind in the United States, and the tool it is using looks, at a glance, like a stock split. It is not one. Understanding why is a small lesson in how one country's shares actually get plumbed into another's market. The reframe — the wrapper changed, not the share In Seoul, SK Hynix trades as common stock under the code 000660. According to the company's SEC filing, a single common share was referenced at 2,425,000 won — roughly $1,581 at the filing's exchange rate. That is a large amount to pay for one share, and it is exactly the kind of number that keeps smaller U.S. buyer...
Read more

Is the Petrodollar Dying? The "Iran War" and the Irony of the Dollar Index

Image
The headlines have been loud all year: "The Era of the Petroyuan is Here." As the conflict between the U.S., Israel, and Iran unfolded, Deutsche Bank and other Wall Street institutions warned that the decades-old dominance of the petrodollar was finally fracturing. The story is compelling, and parts of it are true. But the market data from the crisis told a stranger story than the headlines did. This post first ran in late March, at the height of the Iran conflict. Two months later, with a U.S.–Iran ceasefire reportedly in its final stages, the picture has shifted — and the central paradox it described has, if anything, become clearer. So this is an expansion, updated with where the dollar actually stands as of late May. The Petroyuan Reality Check It is true that sanctioned nations — Russia, Iran, Venezuela — have increasingly turned to the Chinese yuan (CNY) or the UAE dirham (AED) to bypass U.S. sanctions. Saudi Arabia has experimented with China's cross-border dig...
Read more

How to Protect Your Wealth in 2026: The Hidden Trap of Inflation

Image
You may have done everything right. Saved diligently. Lived below your means. Kept money in the bank where it felt safe. And still, the house you could have bought ten years ago has quietly moved out of reach. That is not a failure of effort. It is the hidden trap of inflation — and in 2026, the trap is open wider than it has been in years. This post first ran in March. Two months of data have since made its central point harder to argue with, not easier. So this is an expansion, updated with the numbers that have arrived since — including an inflation reading that surprised to the upside as recently as late May. Is Your Money Quietly Losing Ground? The honest framing is not dramatic. It is arithmetic. A dollar held as cash earns whatever your savings account pays. If inflation runs faster than that yield, the real value of the dollar shrinks even while the number in your account stays the same or grows slowly. During the post-COVID era of zero interest rates, this was easy to se...
Read more