The Problem Is Not AI-Written Code, but the Knowledge Teams Have Lost

Simon Späti, a data-engineering writer who publishes his notes at ssp.sh, argues in a note dated 26 September 2026 that the real problem with AI-generated code is not the code itself but the knowledge that disappears with it. Teams, he writes, no longer know how their own system works or why particular technical decisions were made.

The argument: average code, missing intent

In his assessment, AI writes roughly average code, so a codebase that was below average can be lifted up to average fairly easily. He does not treat that as the main gain: "The problem is not the AI code, but that nobody knows anything, and everyone just asks Claude." What is left behind, he says, is a team with no plan at all.

He adds that engineers who worked before AI tools had to learn a domain in depth, while those tools have made that knowledge look "seemingly obsolete", so someone starting today in an unfamiliar field never builds that foundation.

What engineers report from large companies

He quotes a post on X by the user Voxium, who describes moving into a large company: after half a month he found that specifications, code, tests, PRDs, tickets, ticket resolutions and reports were all produced with Claude Code. According to that account, nobody on the team likes it, middle management keeps pushing for faster shipping, and people work 12 to 13 hours a day.

Späti presents the post as an illustration of his thesis rather than as measured evidence, and the note collected 214 points on Hacker News by 28 September.

Data teams, product managers and the fundamentals

A counterpoint he quotes comes from Hoyt Emerson, who argues that data engineering is different: data people have always had to understand the product and the business from day one, so AI mainly removes friction for them. Späti partly agrees, but notes that this advantage rests on earlier experience.

He also cites Sean Behan's remark that people who cannot code but know exactly what they want have always been valuable. Späti's caveat is that such a person can now build a product, but without coding skill the foundation is weak, especially if the wrong language or the wrong mental model is chosen at the start.

Maintenance is still the final boss

His conclusion is that maintenance stays the hardest part. The easier it becomes to generate a pipeline, an app or a dashboard, the more there is to maintain, and that becomes genuinely difficult when nobody understands the system. AI cannot prompt itself, he notes, which is why intent, taste, design and architecture still matter.

Späti calls the situation partly self-inflicted and repeats an argument he read in the discussion: if companies still hired junior engineers, the knowledge gap would not grow this fast. He concedes that this is easier said than done.