Steve Krouse, co-founder of collaborative coding platform Val Town, published an essay in March 2026 pushing back against the growing consensus that AI is rendering code obsolete. Drawing on a quote widely attributed to Bertrand Russell — "everything is vague to a degree you do not realize till you have tried to make it precise" — Krouse argues that vibe coding, staying at the level of English-language intent while AI generates running code, produces a dangerous illusion of precision. He illustrates the failure mode with entrepreneur Dan Shipper's experience: a vibe-coded collaborative text editor went viral, then collapsed, because the apparent simplicity of the English specification "live collaboration" masked genuinely hard engineering problems that only reveal themselves under real-world conditions. This mirrors the broader problem of <a href="/news/2026-03-16-coding-agents-suck-at-the-xy-problem-llms-never-question-user-intent">agents that never question user intent</a>. The essay is available at stevekrouse.com/precision.

The core of Krouse's argument is a reframe: code is not merely a means to produce software, it is itself a central artifact. The fundamental tool for managing complexity, he contends, is abstraction — the ability to compress multiple concepts into a single, precisely defined unit, and to do so recursively without limit. He cites Sophie Alpert's elegant refactoring of Slack's notoriously complex notification logic as a concrete example, and invokes Dijkstra's definition that abstraction exists not to be vague, but "to create a new semantic level in which one can be absolutely precise." Krouse argues that as AI improves toward AGI, developers will rationally direct that intelligence toward crafting better abstractions and making sound <a href="/news/2026-03-14-ai-is-great-at-writing-code-terrible-at-making-engineering-decisions">engineering decisions</a>, not generating more low-quality output. He offers a personal data point: using Claude Opus 4.6, he generated a full-stack React framework called vtrr for Val Town in a single session — what practitioners call one-shotting a project — framing it as AI enabling higher-order creative work rather than displacing it.

The Hacker News discussion adds empirical weight to Krouse's thesis from an unexpected direction. Commenter lateforwork highlighted a detailed February 2026 review by Chris Lattner — inventor of Swift and LLVM, and founder of AI infrastructure company Modular — of Anthropic's Claude C Compiler. Lattner called the project a genuine industry milestone, marking AI's crossing from local snippet generation into participation in large, multi-subsystem engineering. His verdict on the design itself was pointed: the compiler was "LLVM-like," its architecture shaped by decades of training data rather than representing novel insight. AI systems at their current level conform to the consensus in their training distribution and cannot independently advance the state of the art — a finding that reinforces Krouse's argument that human direction and genuine invention remain indispensable.

Not everyone on Hacker News bought the full argument. A handful of commenters called Krouse's AGI confidence "arrogant and incurious" — the premise that machine intelligence will inevitably match human cognition is doing real work in his optimistic scenario, and he doesn't fully address it. But the core claim doesn't depend on AGI arriving on schedule. Lattner's compiler review already shows where the ceiling is: an AI trained on decades of compiler literature reproduces that literature faithfully, but the design choices nobody has written down yet remain out of reach. Lattner is working on those problems at Modular. So, for different reasons, is Krouse at Val Town.