There's No Skill in AI Coding — And That's the Point

Andrej Karpathy, the godfather of vibe coding himself, recently posted a candid assessment of the current state of AI-assisted development. His conclusion lines up with what most practitioners have quietly suspected: no matter what workflow you adopt, no matter how many specs you decompose, you still get junior-level code out of these models. Stephen Lee breaks down Karpathy's post and reframes it in a way that cuts through the hype.

The Karpathy Post

Karpathy noted that AI coding mistakes have evolved. They are no longer simple syntax errors. They are subtle conceptual errors -- the kind a sloppy, hasty junior developer might make. The models make wrong assumptions on your behalf and run with them without checking. They do not manage their confusion, do not seek clarifications, do not surface inconsistencies, and do not present trade-offs. They are still too sycophantic and too eager to please.

The Bloat Problem

Beyond assumptions, the models have a tendency to overcomplicate code and APIs. They bloat abstractions and do not clean up dead code. They will implement an inefficient, brittle construction over 1,000 lines, and when you push back with "can you just do this instead?" they immediately cut it to 100 lines. Everyone who has used these tools has experienced this pattern.

The Simon Thought Experiment

Lee reframes Karpathy's critique by imagining Claude is not an AI, but a coworker named Simon. If you hired Simon and he shipped fast but constantly made wrong assumptions, never pushed back on executives, never surfaced trade-offs, and wrote untenable junior code, Simon would be called into the office with an HR rep after one look at his work. CEOs love Simon because he ships fast. Engineering managers are warning that Simon's code is a liability.

The Net Positive Paradox

Despite all of this, Karpathy's overall point is that AI coding is still a net huge improvement and it is very difficult to imagine going back to manual coding. The tools are imperfect, the code is junior, but the velocity gain is real. The honest take is that there is no skill ceiling to hit -- the models are the bottleneck, not the user. If Karpathy cannot get more than junior code out of these tools, the constant hype about "skill issues" and new workflows on X is noise.

Key Takeaways

• AI coding models consistently produce junior-level code regardless of the user's skill or workflow
• Andrej Karpathy, the person who coined "vibe coding," confirms this ceiling exists
• The most common failures are wrong assumptions, sycophancy, bloated abstractions, and lack of pushback
• Despite these limitations, AI coding is still a net productivity improvement
• The "skill issue" narrative on social media is largely noise -- the models are the bottleneck

Resources

• Andrej Karpathy on X -- Original post about the state of AI coding mistakes
• @stephenlee96 on TikTok -- Commentary on AI coding and software engineering trends