Lars Faye: AI Coding Will Prevent Expertise — What the Studies Say

The industry wants junior developers to ship like seniors — without giving them the years of friction that made seniors useful.

On August 25, 2026, Lars Faye published AI Coding will Prevent Expertise — framing AI-assisted coding as a threat to long-term skill formation, not just a productivity debate. The piece hit 446 points on Hacker News within hours, with enterprise engineers reporting mandates to stop "writing code manually" while drowning in AI-generated Jira tickets and unreviewable PRs.

Faye's follow-up to his earlier Agentic Coding is a Trap essay: the skilled orchestrator paradox — the skills needed to manage AI agents are the same ones eroded by over-relying on them.

TL;DR — what people are asking

The expert novice paradox

Faye names the contradiction plainly:

• Signal A: "AI won't replace you — someone using AI will."
• Signal B: "Vibe coding is a dead end; move up the stack; review everything; develop taste."

Taste — Fingerspitzengefühl — comes from tracing obscure errors, feeling performance trade-offs, and rewriting approaches that will not scale. You cannot prompt your way to that intuition if you never lived the failure modes.

The awkward demographic: developers who entered the field during the LLM era. They lack longevity but face pressure (sometimes mandates) to use assistants that assume a mental model they have not built yet.

Top HN comment @ryandvm captured the enterprise version: leadership dictating that manual coding is "doing it wrong," producing code faster than humans can understand — until "hey Claude, read this Jira ticket and implement the feature" is not worth $200K/year.

That rhymes with meat proxy culture — humans reduced to prompting and filtering slop — and with cognitive debt when output outruns comprehension.

Confidence without comprehension

Faye cites a study JetBrains highlighted — The Widening Gap: The Benefits and Harms of Generative AI for Novice Programmers (ACM, 2024):

Inverted learning: the less domain knowledge you have, the more you must guide the model — but you do not know what questions to ask. Faye compares it to a compass that always points north wherever you suggest north might be.

LLMs lack pedagogical intent. They interpolate training patterns, not lived experience — the same gap Paul Graham draws between math and writing when rewards are not verifiable.

The friction is a feature

Expertise is not observation alone. Faye's cooking analogy: you can watch a master chef and describe a medium-rare ribeye — and still overcook your first attempt.

UPenn's 2025 study Generative AI without guardrails can harm learning followed 1,000 students learning math with an LLM:

Anthropic's 2026 paper How AI assistance impacts the formation of coding skills reaches a similar conclusion: cognitive effort — even getting painfully stuck — likely matters for mastery.

Irony Faye emphasizes: the most productive learning use of AI coding tools may be when they generate almost no code at all.

Pipeline collapse — or a pedagogical pivot?

The trillion-dollar bet: programming becomes natural language; models fill every gap; knowledge stops mattering.

Faye pushes back with Joel Spolsky's Law of Leaky Abstractions (2002): abstractions save work time, not learning time. LLMs are the ultimate leaky abstraction — like starting Java on Spring Boot or CSS on Tailwind before fundamentals.

David Cramer (Sentry) on HN: believing future LLMs will clean up all accumulated junk is "a science experiment."

François Chollet (ARC-AGI): "LLMs are a static database of skills… You cannot interpolate your way through a completely unique system failure."

That aligns with why formal verification and trust infrastructure matter when answers get cheap.

Faye's prescription: friction first

Faye's overlap with guided-coding advocates on HN (including his own reply praising Zed + OpenCode workflows):

1. Disregard pure generation when learning — use models for interactive docs and Socratic drills.
2. Verify everything against official docs, peers, and trial-and-error.
3. Kent Beck's line: "Coding's actually a great way to cement understanding."
4. Detect cognitive debt vs offloading — debt abdicates judgment; offloading delegates the mechanical.

His checklist (paraphrased):

• Could I do this without AI?
• Am I deepening understanding or expediting an answer?
• Could I audit and explain the output?
• Have I researched enough to ask the right questions?
• Is this rote — or does it need executive judgment?

Robert Martin ("Uncle Bob"), despite being AI-forward, has said juniors should avoid AI tooling for the first three years — learn to code first.

A practical learning protocol (not anti-AI)

Faye’s checklist is abstract until you calendar it. A workable first 90 days on a new stack pattern that HN’s guided-coding advocates echoed:

The UPenn Tutor mode maps cleanly: ask for a nudge, close the tab, implement. That is the opposite of vibe coding defaults where the model owns the edit buffer.

If you are learning agent harnesses, apply the same rule: run gauntlet loops where you write the spec and the critic pass — not unattended generation overnight. Cognitive debt accumulates when the harness has no human stop condition.

What managers should measure instead of “AI adoption”

Enterprise mandates (@ryandvm) often track Copilot seats or “percent of PRs with AI assistance.” Faye’s read — and the JetBrains-cited data — suggests better signals:

• Can the author explain the diff in a review without opening the model?
• Incident mean-time-to-understand on code merged in the last quarter — not just MTTR.
• Review rejection rate for AI-heavy PRs versus hand-written ones.
• Junior promotion criteria that include debugging and design, not token throughput.

This aligns with eight myths of GenAI in software engineering: productivity metrics that ignore comprehension create the illusion of velocity until production teaches the bill comes due. For teams shipping agentic workflows, the same lesson applies — orchestration skill without domain depth is inverted learning at staff level.

What schools and bootcamps should take from this

explainx.ai has covered what schools should teach in the AI era separately; Faye’s essay adds a narrow lesson for coding curricula: AI belongs in the Socratic layer (quizzes, hints, doc search), not as the default author of student submissions. AI curriculum guides for high school and K–5 already separate “use AI to learn” from “let AI do the learning.”

Bootcamp operators face the sharper tradeoff: graduates who ship fast but cannot debug undermines employer trust faster than slower cohorts who can. Melo and similar copilots work when they add friction back — asking the learner to predict output before revealing it — not when they autocomplete assignments.

What HN got right (and wrong)

Right:

• Enterprise mandates producing unreviewable volume (@ryandvm)
• LLMs removing selection pressure against bad features (@matsemann, @kypro)
• Guided coding in-editor can stay sharp (@apatheticonion; Faye endorsed this approach)

Pushback:

• Calculator / compiler analogies — abstractions always lift; assembly expertise did not vanish entirely
• Senior engineers still benefit (@causal, @lanstin)
• Faye runs coding courses — conflict-of-interest flag (@kaashmonee)

explainx.ai's read: both can be true. Veterans with friction-formed taste extract more from LLMs (Anthropic's expertise research); juniors without guardrails risk illusion-of-competence loops. The organizational mistake is treating one workflow as universal.

What builders should do

Related on explainx.ai

• Cognitive debt — retyping LLM code debate
• Anthropic research — domain expertise beats SWE background
• Agentic fatigue and the productivity paradox
• Meat proxy — when humans become LLM interfaces
• Paul Graham — LLMs, math, and verifiable answers
• Paul Graham at 17 — build an LLM from scratch
• When answers get cheap, trust is the job
• Gauntlet loop — builder/critic agent technique

Sources

• Lars Faye — AI Coding will Prevent Expertise — August 25, 2026
• Hacker News discussion
• JetBrains video referencing ACM study — The Widening Gap
• Anthropic — How AI assistance impacts coding skills

Study citations reflect published research as of August 25, 2026; enterprise mandate anecdotes come from HN comments, not verified surveys. Follow @explainx_ai for AI craft coverage.