In his conversation with Zach Bruggeman, Principal Software Engineer at Ramp and lead of Inspect, Silver.dev founder Gabriel Benmergui describes an uncomfortable incentive: AI makes second-rate work almost free, while excellent work still costs money and attention.
He puts the concern bluntly:
"And so that very scarily points me to a dystopia where everything is a slop because the slop is free."
If you're building a startup, this affects what you ship and who you hire. You can produce more code or fill more pages without making the product any better. You still have to decide whether the extra output helps a customer enough to justify keeping it.
Software had a quality problem before AI
Gabriel brings up Jonathan Blow's view that software is getting worse. He thinks Blow is too strict, but recognizes the complaint: developers could produce plenty of junk long before coding agents arrived.
AI makes that habit easier to indulge. A team can generate an implementation before it has agreed on the behavior users need. Reviewing it takes effort, and deleting something that looks finished can feel wasteful. So the team keeps code it would have questioned if someone had proposed spending a week writing it.
That is a risk to manage, not evidence that every AI-assisted project is bad. The useful question for a review is whether this particular change deserves to exist. A working demo alone cannot answer that.
A checkout page can earn its copywriter's fee
Gabriel uses copywriting to explain the difference between acceptable output and work worth paying for. Even a shoe checkout page involves choices about language that can affect whether someone buys.
A capable copywriter considers what the buyer needs to know at that moment. A sentence might clear up a concern about the order. Another might distract from completing it. Generating fluent alternatives gives you material to edit; choosing among them requires an understanding of the customer.
Gabriel describes the price gap:
"second-rate copy is free, virtually free, and first-rate copy is very expensive."
His point is about the temptation to accept adequate writing because improving it requires a budget. The cost of producing a draft and the value of the finished page are different things.
Skills and templates still need an editor
Gabriel includes guidelines, skills and templates in his argument. You can give a model better instructions and get something that passes a basic quality check. You can also follow those instructions and publish a page that sounds interchangeable with your competitors.
A template can specify the sections. Someone on the team has to decide whether the argument is useful and whether you can support its claims. Adding another instruction to remove generic language won't supply a customer insight you never had.
For code, make that responsibility explicit in review. Ask the author to explain a design choice and show how they checked the behavior. Give them room to discard the generated approach. If accepting the agent's first plausible answer is the easiest way to finish a ticket, expect engineers to do it under pressure.
Hire for the decisions after generation
If you're hiring engineers in Latin America for a US startup, let candidates use the tools they would use on the job. Review the result together.
Ask them to explain an assumption the agent made. Change a requirement and watch whether they understand the implementation well enough to adapt it. A candidate who notices an unnecessary dependency or questions the requested feature gives you evidence of judgment that a polished output alone cannot provide.
Silver.dev's AI adoption interview guide, in Spanish, is a related resource for candidates preparing to discuss how they work with these tools. Founders should also decide who owns the quality bar after the interview. An engineer needs permission to reject unnecessary work, including work a founder requested.
For your next AI-assisted change, put a reviewer on the finished result and ask what the customer gains. If you cannot answer, reduce the scope before adding more output. Watch the full conversation with Zach Bruggeman for the discussion behind the clip.