OpenAI's CEO Just Told Non-Coders They're the New Power Users

If you can't write a line of code but you know exactly what your customers actually need, this is the week to believe that matters more than it used to. It's also the week to notice the catch nobody's headline mentioned.

What changed

At the G20 Innovation Ministerial on September 7, OpenAI's Sam Altman said AI is producing "the revenge of the idea guys" — people with deep knowledge of their users and their industry, but no technical background, who can now build the product themselves. "For a long time, the most important ingredient was technical talent," he said. "Now people who just really deeply understand their users and can't code at all — I want to fund those people."

Four days later, developer Simon Willison responded to a post that had gone viral on Hacker News titled "Feeling sad about AI." His reframe: translating an exact specification into working code isn't a scarce skill anymore. What's still scarce is scoping the problem correctly, debugging when something goes wrong in a system nobody fully understands, and knowing which questions to ask before you start. Altman is describing the same shift from the funding side — that Willison is describing from the execution side.

The same week, Goldman Sachs executives raised a specific worry about OpenAI's new ChatGPT Financial Services workspace, which automates M&A modeling, buyer screens, and earnings analysis for banks. If a tool does the analysis a junior analyst used to do by hand, where does that analyst's judgment come from later, when they're the one signing off on a deal? The reps that used to build expertise are the same reps a tool like this removes.

The skill implication

Altman's claim and Willison's claim point the same direction: the bottleneck moved from "can you execute this task" to "do you know what a good outcome looks like." That's real, and it's good news if you've spent years in a specific industry without ever needing to code. But Goldman's warning is the part that doesn't fit neatly into a funding pitch: judgment isn't a fixed asset you already have. It's built by doing the work, including the parts that are tedious enough to delegate first.

Three things to do differently this week, not eventually:

1. Treat your domain knowledge as the asset it now is. If you understand your customers, your market, or your industry's actual constraints better than most engineers do, that's the input AI can't supply on its own — put it into the brief before you generate anything, not as an afterthought once you're reviewing the output.
2. Practice telling good output from plausible output. The skill Willison is describing — knowing whether an answer is actually right, not just fluent — doesn't come from reading about it. It comes from deliberately checking your own AI-generated work against a standard you set first.
3. Keep doing some of the reps by hand. If you're in a role where AI can now do the analysis you used to do yourself, don't let all of it go. The judgment that makes you useful when something goes wrong was built by the version of the work you're now tempted to skip entirely.

Worked example: the marketing manager

Renata runs marketing at a 40-person B2B SaaS company and doesn't write code. She asks an AI tool to draft a go-to-market plan for a new feature, and gets back something fluent, well-organized, and wrong about her actual buyer — it assumes a self-serve motion when her last six deals all closed through a two-person sales team. She catches it not because she prompted better, but because she's spent three years on calls with these exact buyers. She rewrites the targeting section herself, keeps the AI draft's structure, and ships a plan that reflects how her company actually sells. The code-writing was never the hard part. Knowing her buyer was.

Second example: the junior analyst

Dmitri is eighteen months into a research-analyst role at a regional bank that just adopted a ChatGPT-based financial workspace for earnings analysis. The tool cuts a task that used to take him a full afternoon down to twenty minutes, and his manager is thrilled. What worries Dmitri, a few months in, is that he can produce the analysis faster than he can explain why one comparable company was the right one to model against — he skipped the part of the work where that judgment used to form. He starts doing one analysis a week without the tool, comparing his manual read against what the model would have produced, specifically to keep that muscle from going quiet.

The takeaway

Domain expertise is worth more than it was two years ago, but only for the person who keeps building the judgment that makes it usable — not the person who lets AI do the reps that judgment came from.