Will there ever be a worse time to start a startup?

Today's frontier is tomorrow's tech debt.

If only he had waited a bit longer.

Deflation is an odd phenomenon. The problems associated with inflation are fairly intuitive—when prices go up, people can’t buy as much of the stuff they want or need. But deflation? People like lower prices! It’s a whole thing! If inflation is bad, shouldn’t its opposite be good?

Most economists say, emphatically, no. Deflation is not only bad; it’s often considered worse than inflation.¹

Because, when prices are falling and people expect things to get cheaper, they save their money instead of spending it. Moreover, borrowing—which fuels a lot of economic activity—is especially disincentivized, because if you borrow $400,000 to buy a house, the $400,000 principal you owe back to the bank will be more valuable than the $400,000 you borrowed.² Finally, to make up for the money they’re losing from falling prices, companies need to reduce wages or lower employees’ salaries. Though that’s technically possible, workers tend to “resist pay cuts for many reasons, most obviously because cuts lead to a lower standard of living, but also because they may be perceived as unfair or demeaning.” This makes cutting wages practically infeasible, so firms have to save money in other ways—by building less stuff, by reducing employee benefits, or by laying people off. The whole thing can spiral: People save more, borrow less, and spend less; firms invest less and fire people; this reduces economic activity further; as their finances tighten, people save more, borrow less, and spend less; down and down and down.

And the more severe the deflation, the more it compounds. If you think cars will cost 1 percent less in a year, you may still buy one today. But if they keep getting 10 percent cheaper every month, well. Imagine the car you could buy if you just wait a year.

Anyway. It would be very strange to say that right now is a bad time to start a company. Startups are growing faster than ever, money is coming out of venture capitalists’ ears, and everyone you know is a billionaire. The conventional wisdom is that now is a great time to start a company. There has never been a better time to start a startup. There’s never been a more amazing time to go create something totally new:

You have access to tools that can let you do what used to take teams of hundreds.

This was particularly evident last week, when Anthropic released Opus 4.5. As other people have said, Opus 4.5 was alarmingly good: “You can build astonishingly complex apps without looking at a single line of code.” Or, as I said in an earlier post:

The comfortable physics that I thought governed Silicon Valley—that stuff takes time to build; that products need to be designed before they can be created; that computers cannot assume intent or interpolate their way through incomplete ideas—broke, utterly. It all worked too well, too fast. I was staggered, drunk on the Kool-Aid and high on the pills, unwell and off-brand. I knew that anyone can now build vibe-coded toys; I did not know that people with a basic familiarity with code could go much, much further. … In 2013, it took us eight people, nine months, and hundreds of thousands of dollars to build something we could sell, and that was seen as reasonably efficient. Today, that feels possible to do with one person in two days.

Ah, no, wait. That quote isn’t about Opus 4.5; it’s from February, and is about Cursor and Anthropic’s Sonnet 3.5, a model that’s so outdated that it has already been deprecated.

But that’s the pattern now, isn’t it? A new model comes out; we declare the game changed. And we wonder how we got anything done with the trash we used to have:

Claude 3.5 Sonnet isn’t just an incremental update—it’s a quantum leap forward.

There was this moment—I think it was around when Sonnet 3.7 came out—where I used it and I was like, holy shit, this is a completely new paradigm.

Holy shit. I’ve used ChatGPT every day for 3 years. Just spent 2 hours on Gemini 3. I’m not going back. The leap is insane.

Opus 4.5 has left him feeling wonderstruck, excited.

In a sense, none of this is new; technology is always getting better. “You’re the oldest you’ve ever been and the youngest you’ll ever be” is both profound and trivial; so is “technology is the best it’s ever been, and the worst time it’ll ever be.” If all you think about is the tools that are available to you, then today is always a better time to start a company than yesterday, and today will always be worse than tomorrow. The cost of doing something with a computer goes one direction: Down.

But what if those costs are falling quickly? What if doing things gets 10 percent cheaper every month?³ Imagine what you could build if you just wait a year.

For example, I’ve been playing around with various AI coding tools to build a couple personal products or experimental ideas. Every couple months, when some new AI thing comes out—first Replit; then Cursor; then Lovable; then Cursor, with Gemini 2.5 Pro; then Claude Code; then Antigravity; then, last week, Claude Code and Opus 4.5—the same story repeats itself:

1. I try to use the new tool to improve the current version of my app, and it works ok.

2. I eventually give up, lob a short description of my app at the new tool, and tell it to build a fresh version from scratch.

3. Almost immediately, the new one is better. It’s more durable; it’s faster; it’s better designed.⁴

4. I think, ah, yes, this is the one.⁵

Part of the reason this seems to happen is because each successive model is, in effect, writing for a different—and potentially incompatible—audience. The early versions of Cursor finished the code that I started. Then, it was more of a coworker, pair programming alongside me. Now, why even look at an IDE anymore? Let them generate whatever code they want. As long as it works, we should be happy.

Earlier this year, I guessed that AI models would eventually drift towards writing code that was optimized for their computer comprehension over human comprehension:

Perhaps we should let the machine write for itself and its abilities. Let it be redundant. Let it ignore the frameworks; let it create explosively large codebases. Let it write 1,000 distinct pipelines in 1,000 distinct languages. Let it be offensively unaesthetic. Let it be radically simple.

But that theory has a corollary that I hadn’t considered: The machines aren’t static. Across vendors and release versions, models could develop their own habits. Just as Sonnet 3.5 might be better off unencumbered by us and our feeble reasoning, Opus 4.5 could be better off unencumbered by Sonnet 3.5. And if that’s true, it seems naive to assume that some future AI won’t want to undo the mess Opus 4.5 made. There is no fixed definition of tech debt—it is simply code that the current engineers would prefer to be written differently.

Of course, a side project is not a startup, and benn.chat is not a capital-P Product. But still—to rephrase the original question, if now is a great time to start a company, was 2023 a great time to do it? Well, yes, obviously—that is where all the billionaires came from. But if you started a company in 2023 and it didn’t take off—if you’re not a billionaire but are instead still pushing the boulder uphill—would you do things differently if you were starting the company again today? Do you wish you were starting it today?

After all, you can rename your company to keep up with the latest fads, but it’s not so easy to rebuild it.

1. Though not every economist agrees with this, because if you put two economists in a room, you’ll get three opinions.

2. Banks could, in theory, charge negative interest rates, where they loan out $400,000 and ask for only $390,000 back. Central banks have experimented with this, but it doesn’t really work for private lenders because they can just hold cash. Why would a bank lend someone $400,000 so that they can—probably!—get paid $390,000 later, when they could instead put $400,000 worth of cash in a vault that will—definitely! Or, probably?—be there in a year?

3. “The length of coding tasks frontier systems can complete is growing exponentially—doubling every 7 months,” which, very roughly, implies that they can complete 10 percent more every month.

4. Sure, they’re great at writing code, but, are they good, I am contractually obligated to ask you, at analyzing data?! Maybe! I don’t know! Let’s all find out together!

5. At the risk of falling into the exact trap this post is trying to highlight: Opus 4.5 is unsettlingly good. When I said this in April, I felt like I be losing myself a bit in artistic hyperbole:

The dominant conglomerates of the future won’t be the companies that build software with humanoid agents, but those that figure out how to run the computing machine at a massive scale. They will figure out how to put coding agents on a perpetual loop, in a factory that doesn’t have to sleep or take vacations. They will be the companies that industrialize the most, and optimize for ACPE—average compute per employee. They will be the ones that turn engineers into factory supervisors who watch the line, look for defects, and doze off to the dull hum of the machinery that replaced them.

Now, frankly, I’m not sure it goes far enough.