“Financially, what will take me to 1 billion?”
—OpenAI co-founder Greg Brockman’s diary entry from 2017, which surfaced in discovery as part of the epic Musk v. Altman litigation. The trial will commence in April; Musk is seeking up to $110 billion in damages.

And now for the main event…

Though there are dozens of A.I. coding tools out there, Anthropic’s Claude Code seems to exist at the center of what feels like an actual revolution—one of the first real examples of the radical impact that A.I. is sure to have on software development, and the subject of fawning headlines like “Move Over, ChatGPT” (The Atlantic) and “Even Non-Nerds Are Blown Away” (The Journal). It’s a boon, of course, for Anthropic, and a sign that its focus on coding and enterprise applications is paying off. Anthropic’s Boris Cherny claimed that his team built the recently released Claude Cowork software in under two weeks using Claude Code.

Last March, Anthropic C.E.O. Dario Amodei predicted that 90 percent of all code would be written by A.I. by the middle of last year. That didn’t quite pan out, but we are entering an era where amateurs with little to no software development expertise are, well, developing software. One woman documented how she used Claude, along with ChatGPT and v0 by Vercel, to build a restaurant-selection app in a week. Others are using it to build personalized news feed apps, or email organization workflows, or custom mobile games for the family to play over the holiday.

Recently, I gave it a try myself. I decided to start slow, with GitHub Codespaces and a boring idea offered up by ChatGPT: “Build a simple but powerful browser-first web app that automatically saves copied content with context (source, time, surrounding text) and makes it searchable and reusable later.” After a few minutes, that 23-word prompt became a modest but functional web app. A quick highlight and a right click, and I can save quotes—complete with surrounding context and a link to the original source, be it tweet or article—on an endless clipboard attached to my browser. Useful, sure, but not a game-changer.

Then I downloaded Claude Code, and asked it to make me a Star Wars–themed online tower defense game, a request it obliged. In less than a half-hour, I was fighting the Empire (I managed to win after 15 rounds), although Claude Code failed to honor my request for a background that looked like planets from Star Wars. We had a bit of a (polite) back-and-forth, but I gave up after a few attempts. Still, I had a workable game after $20 and 20 minutes.

A good start, so I decided to up the ante. I’ve been intent on mastering conversational Spanish for a long time, and have the lapsed Duolingo account and dusty textbooks to prove it. So I asked Claude to build me a “highly competitive, production-level app designed to teach non-Spanish speakers total conversational Spanish.” Claude got to work, building me a web app it called FluenteSpanol, which included vocab practice and scenario-dependent conversational training. This was by far the most complex thing I attempted to build with Claude Code, and it was genuinely remarkable.

And yet, there were immediate issues around context awareness during the conversational practice, as well as some broken links. And thus began my days-long back-and-forth with Claude: I would ask the bot to fix these issues; it would “identify” and “fix” the problem in the app’s codebase; and then… the issues persisted. I went through six rounds of this before giving up. Yesterday, I tried it twice more in a fresh chat, but to no avail.

As I explained to Dr. Armando Solar-Lezama, an MIT computing professor, my amazement at the speed with which Claude built my app didn’t diminish my skepticism of the approach. As a non-coder, I assumed that problems on the front end implied that there would be some on the back end, too, but I’d need a professional to dogfood it for me. (Also, I probably needed a Spanish speaker to vet the results that Claude was spitting out.) This, Solar-Lezama told me, was consistent with his own experience: “At some point, you need to know how to program in order to actually get what you want,” he said. Part of the challenge is simply in identifying where the coding tool went wrong and understanding how to fix it—an insurmountable challenge for most amateurs. “If you are not writing the code, some basic understanding of coding is still essential to write anything that is not trivial,” Solar-Lezama continued. “All the people I know who are leveraging these tools to do useful things know how to code, even if they are not software developers.”

So, A.I. has not made professional developers obsolete, at least not yet. Itamar Friedman, the C.E.O. of code quality and review startup Qodo, told me that the tools are not at a point where non-developers can truly contribute to software development. “With all due respect to everything you’re seeing on Twitter and everyone saying, Wow, Claude Code created this amazing thing, 99 percent of what you’re using right now on your phone and in your computer is basically code that was written four years ago,” he said.

The journey to that stage, he added, would involve developing A.I. systems that have reliable contextual awareness of the individual quirks, standards, and best practices of a given company’s codebase, while also training the underlying models to both generate and review code. Non-developers, he said, can easily create a bug without knowing what they’ve done. (By the way, he fully expects a cyberattack resulting from shoddy, A.I.-generated code, à la the 2024 CrowdStrike outage, to occur soon.) “Let’s admit, it’s still in the early stages,” he noted. But he thinks we’ll get there eventually. Friedman said he’s aware of engineers at enterprises who are now managing sizable teams of A.I. coding agents.

As code review and verification systems improve—and as A.I. coding tools help non-developers act more and more like developers—Friedman envisions a barbell effect, wherein the number of junior and senior developers increases while midlevel developers decline. Despite Amodei’s premonition and increasingly noisy anecdotes about A.I.-related job replacement, the number of software developers around the world has increased dramatically since 2022, according to market research firm SlashData, from 31 million to 47.2 million in 2025. That’s not to mention the persistent lack of clear evidence, pointed out by the Oxford Economics group and others, that causally ties A.I. to job losses in general. And even as A.I. adoption increases, many developers still don’t trust the accuracy of its outputs. Only 15 percent say they perceive A.I. as a threat to their jobs, according to Stack Overflow’s 2025 developer survey.

Of course, there’s little doubt that L.L.M.s are changing the practice of coding itself. But that doesn’t mean the field of software engineering will simply disappear. Instead, said Dr. Fabian Braesemann, a researcher at the Oxford Internet Institute, coding will just become more accessible to the masses—similar to how, in the early days of the internet, anyone with a bit of HTML knowledge could slap together a simple blog or website. “Back then, the internet did not lead to the disappearance of software developers, but instead to the creation of a whole new occupation—web developer—and ever more appealing websites,” he said, adding that he expects the same cycle to repeat itself now. “Instead of developers being pushed off the job market, there is even more demand because many companies in non-tech industries can now afford to also do their fair share of web and software development in-house.”

As for the value of today’s personalized A.I. vibe-coded apps, Braesemann is skeptical. “We will see a very long tail of not-useful tools that are not professional and that will disappear,” he predicted. “Over time, however, new industries and big corporations that leverage A.I. and data to produce novel software products will undoubtedly appear.”

That’s all for today. I’ll see you on Thursday.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006