To code and to not code

Hey friends,

This week I went to OpenAI's fourth annual DevDay conference in San Francisco. It's been a while since I went to a big tech conference, and I've never had the privilege(?) of going to one where the company is not just hot, but at the center of every economic, cultural, and political discourse.

On my way into Fort Mason, anti-AI protesters were waving signs and singing slogans. The tech stream/podcast TBPN had set up a booth to report and comment on DevDay live, as if it were the Olympics. As I walked into the main building to find a seat for the keynote, Sam Altman was being interviewed against the (amazingly gorgeous) backdrop of the Bay, surrounded by a ring of lights, camera equipment, and security.

I’ll say this for Altman and OpenAI — they are much better at reading a room than Anthropic. For starters, even though the week started with the company announcing they were delaying their next high-end model, GPT-6.1 Astra, while they figure out how to address numerous security and safety concerns, no one at DevDay was talking about whether AI models were going to kill us. For that matter, no one was saying coding is "solved" or that software-making had become a low-skill, fully automated job.

Instead, there were tons of nods to creators, including/especially software developers, and how ChatGPT and Codex enhance their creative process. For instance, a gorgeously produced video and accompanying physical installation called Codex Originals showcased individuals who are using OpenAI’s products to do amazing things, from work toward literally curing cancer to building platforms to help people get better access to education.

Now, there is a gigantic extent to which, when OpenAI says "dev," they're really saying "build." While several of the newly announced products were APIs for (traditional) developers, mostly the emphasis was on ways you can create apps or games with only prompting. At the closing Q&A, Tibo — the OpenAI engineer in charge of the Codex coding agent — mused aloud they might rename next year's event "BuildDay."

As much as Big Tech loves to test engineering candidates on algorithms, data structures, and other CS fundamentals, programming at scale has always been as much about reading the docs or Googling "[specific problem] site:stackoverflow.com" as knowing what code to write.

Earlier this year, a study by ETH Zurich found that prior knowledge of computer science was the single biggest predictor of success when "vibe coding" — basically, you get better results from AI coding agents if you know what you're asking for but just want the robot to look at the docs and type it out for you.

Additionally, better general writing skills were correlated with success at getting good results prompting — this makes sense, because prompting is writing, and clarity is important when your audience is an AI model that has no world knowledge and has never been alive. A lot of communication between people can afford to be "lossy" because we can count on the other person to either know certain things, ask follow-up questions, or selectively ignore parts of what we're saying that aren't relevant. Good communication with models presumes less and specifies more.

In the best cases, even though coding with AI can rob builders of a deep feeling of satisfaction from figuring out the solution to a coding problem — seeing an error-free UI appear in the browser, seeing a test suite flip from red to green — in return we get speed, momentum, and (probably most important) freedom to focus on whether we built the right thing, not just something that compiles.

Ideally, using AI allows us to move up the hierarchy of building, focusing less on how to build software and more on what to build. "What", in this context, has a couple of layers. At the moment, most AI models are competent coders but bad architects; per the ETH study, you get better results if you know how to architect an app and are simply asking the AI to fill in the blanks with code.

If the day should come when AI gets good at software architecture, it'll still be necessary to deeply understand what problem you're trying to solve — what people like me call product sense — and, if your goal is to build a business and make money, how to talk about it with other people (i.e. sales & marketing).

Only the bottom two layers of the pyramid above are "building" in the sense of creating software or digital systems — the rest is all about people.

Links worth reading

🤖 The dot and the swarm — Ethan Mollick on "the Bitter Lesson" — that many problems we assume require human taste & judgment can actually be solved by throwing enough AI at them — and how it applies to swarms, like the one OpenAI used to solve Navier-Stokes, and personal bots like Muse, Grok Bot, and OpenAI's just-released dots.

🏇 What happens when you can build it all yourself? — Renato Valdés-Olmos grapples with questions in the same vein as my essay above—what does it mean when building becomes fast and easy?

🐸 The HuggingFace AI attack explained as a Frog & Toad picture book — Before you ask, while it's about AI, all the writing and artwork were 100% human-created.

📰 EmDash's 1.0 release — EmDash, announced as a "spiritual successor to WordPress" earlier this year, has reached a stable 1.0 release. It's based on Astro, and (at least by default) works best when deployed to Cloudflare, which acquired Astro at the start of the year and employs the folks working on it.

Heeeeere's Johnny

After I got back home to New Jersey on Thursday night, Johnny felt it necessary to keep me from leaving again by planting himself on my legs and sleeping there. (He's doing it again as I write this on Friday morning.)

See y'all next week.