Simon Willison on 'Feeling Sad About AI': The Despair Phase Passes

Every few months, a thread like this one reliably lights up Hacker News: someone describes real sadness, not hype-fatigue, about what AI coding agents are doing to the craft they built a career on. This time the trigger was an essay by Andy Balaam, and the resulting discussion ran the usual range — some readers called AI a superpower, others said it was quietly hollowing out the parts of the job they actually loved.

What made this round notable is who answered it. Simon Willison — a widely read voice on practical LLM tooling — posted a direct reply that skipped the philosophical debate and spoke from experience instead.

The claim: this is a phase, not a plateau

Willison's argument is narrow and specific, which is why it lands. He says he went through this exact moment of existential crisis himself, "a few years ago now," and that a lot of engineers who feel disheartened watching a coding agent finish a week's worth of work in an hour eventually come out the other side of it.

His framing of the underlying shift: "translating an exact specification into decent code isn't a unique skill any more." That's a real loss of exclusivity for a skill a lot of careers were built on. But he doesn't treat it as the end of the job — he treats it as the end of one particular bottleneck, freeing attention for the larger set of problems that were always there: what to build, whether it's correct, how it holds up, who it serves.

Why this matters beyond one blog post

None of this is a new pattern for software engineering specifically. The field has absorbed compilers, frameworks, cloud infrastructure, and open-source package ecosystems, each of which eliminated some previously-scarce skill. Willison's point is that AI coding agents are the same kind of upheaval, just compressed into a shorter timeline — which is exactly why the emotional reaction hits harder even when the structural pattern isn't new.

The uncomfortable part, visible in the HN thread itself, is that the timeline compression isn't cost-free. Faster upheaval means less runway to retrain, and it raises real questions about how newer engineers build the judgment that senior engineers are now told to lean on — judgment that, historically, came from years of writing the code that agents now write instead.

What to actually do with this

• Stop measuring your value by typing speed or lines shipped per day — that metric is being commoditized in real time.
• Invest deliberately in the parts of engineering an agent can't own outright: specifying the right problem, reviewing for correctness and security, and taking responsibility for outcomes in production.
• Treat agent output as a draft from a fast, uneven collaborator — verify it the way you'd verify code from a junior engineer you don't fully trust yet, not the way you'd trust a compiler.

That's also the practical shape of what people mean by 'vibecoding' when it's done seriously: not typing prompts and shipping whatever comes back, but directing, specifying, and reviewing AI-generated work with the same rigor — arguably more — than hand-written code ever got.