The developer’s melancholy
Hi everyone, I’m back in Paris just in time for this warm, summer-like weekend ahead. I’d like to thank Box once again for supporting Hypertext as its very first sponsor. Feel free to share this email with a friend, it really helps 💌 Sign up here.
Developers as train dispatchers
There’s an old saying that technological shifts happen gradually, then suddenly. And it feels particularly apt when talking with software developers. Since the release of Anthropic’s Claude Opus 4.5 model in November 2025, many developers feel like they got a new job.
Instead of writing code directly, they increasingly write instructions for coding agents that generate the code for them. If coding agents are the trains, developers have become the train dispatchers. They assign tasks, monitor progress, fix failures, and coordinate dozens of parallel workflows.
And don’t just take my word for it. This week, Niklas Gustavsson, Spotify’s Chief Architect and VP of Engineering, gave an interesting talk at an Anthropic event reviewing the company’s shifting development strategy. “AI tools in general have gone completely bananas. And today more than 99% of our engineers use AI coding tools every week,” he said.
Quick aside: I’ve always paid attention to Spotify’s architecture work because they’ve regularly been extremely innovative. For instance, when Spotify was still a small Swedish company, they used peer-to-peer transfers to optimize server bandwidth. When a user hit play on a popular song, Spotify would try to download the track from someone else’s computer before downloading it from the company’s servers.
They didn’t talk about it too much because peer-to-peer transfers had a bad reputation due to Napster, Kazaa, LimeWire, eDonkey2000 and (eventually) BitTorrent. That way, they didn’t have to constantly explain: “No but this is legal peer-to-peer!”
Today, Spotify ships 76% more updates to the code base (pull requests, or PRs for short) compared to last year. “Now, by far, most of the PRs that we ship are authored by an AI agent together with the developer,” he said. Every month, AI agents generate over a thousand pull requests.
Right now, a human developer looks at the agent’s work and validates the pull request. But this is Sisyphean work. AI never gets tired. Humans do.
“One of our most frequent feedbacks at the moment is there’s just too many freaking PRs to review. So we need to figure out where we apply humans to review those PRs where it matters the most,” Niklas Gustavsson said. “We’re already auto-approving some PRs that we think are safe enough to merge without human review and then we try to focus the human review where it really matters.”
And this may be the most important sentence in Gustavsson’s entire talk: “We'll figure out over time where we need the human judgment to be applied.”
I used Spotify as an example, but I could just as easily have referenced Mistral’s Arthur Mensch testimony at the French National Assembly (Chris O’Brien has a good write-up of the testimony if you want more on this). “Today, engineers at Mistral no longer write lines of code,” the CEO of Mistral said.
“In the past, [software development] was more of a craft if you were an individual contributor. That is to say, you wrote your own code. In fact, people enjoyed that craftsmanship,” he added. “Today, you’re no longer a craftsperson; you’re a manager, and you ask agents to write the code for you. You provide the specifications, you’re the one giving orders, and that’s a pretty profound change.”
These days, a lot of developers talk about potential burnout from AI tools for a combination of reasons:
- The obvious part: they now have so many agents to manage and so much code to review.
- The pernicious vibe: they feel like they have to outperform their peers to remain relevant as big companies like Meta and Amazon are laying thousands of employees (approximately 8,000 jobs for Meta this week alone). Developers often tell me that the current AI shift is “a rat race.”
- The side effect: due to those huge productivity jumps, more faulty code gets pushed into production. When AI breaks things, you call a human for an emergency code review.
But one thing that comes up less often, and may ultimately be even more heartbreaking, is what I’d call developer melancholy.
Many developers loved getting in the zone and writing hundreds of lines of code in a single session to solve a Very Hard Problem. When you’re spending your days talking with agents, it’s just not the same.
Huge amounts of business content remain ungoverned. Box brings structure, security, and control to your unstructured data, enabling secure, enterprise-ready AI.
Tokens as the universal currency
A few years ago, I wrote an AI glossary with my friend (and former colleague) Natasha. Here’s how we described tokens: they are the basic building blocks of human-AI communication, representing discrete segments of data that have been processed or produced by an LLM. They are created through a process called tokenization, which breaks down raw text into bite-sized units a language model can digest, similar to how a compiler translates human language into binary code a computer can understand. In enterprise settings, tokens also determine cost — most AI companies charge for LLM usage on a per-token basis, meaning the more a business uses, the more it pays.
What I didn’t include in that definition is that tokens are increasingly becoming a global currency with geopolitical consequences. Mistral CEO Arthur Mensch articulated that idea pretty well in his testimony at the French National Assembly.
“We’re in a business where we convert electricity into token generation. And fundamentally, this is something that needs to be treated more or less like a natural resource,” he said. “In other words, we need to think about intelligence the same way we think about energy. So our goal is to provide intelligence that is affordable, secure in terms of supply — meaning we don’t have to source it from the U.S., for example — and that is also sustainable because it transforms energy with a lower carbon footprint than elsewhere, since we use French energy in particular.”
Later in the conversation, he even went one step further and roughly translated a gigawatt of electricity into a corresponding amount of tokens. Those tokens can then be monetized (with a margin, hopefully) through different business models:
- Hosted AI inference APIs (Anthropic’s strategy with Claude Code).
- Raw GPU capacity (SpaceX’s strategy with Colossus 1).
- Enterprise AI transformation projects (Mistral’s strategy with its services-led enterprise customers).
But here’s a fourth idea few people had considered. OpenAI CEO Sam Altman is offering $2 million worth of OpenAI tokens to every startup in the current Y Combinator batch in exchange for equity in the startups (through an uncapped SAFE agreement that would convert into equity at the Series A round).
Yes, you read that right. OpenAI could end up owning stakes in hundreds of buzzy startups without spending a dollar from its cash balance.
Taken together with OpenAI’s convoluted deals with Microsoft, Nvidia, AMD, Oracle and others, this suggests Sam Altman may be an AI innovator, but he is also proving to be an even more important innovator… in startup financing.
Have a good day ☀️
Romain