Expert Beginners and Lone Wolves will dominate this early LLM era - Jeff Geerling

Jeff Geerling, a veteran technical architect with extensive experience building Drupal sites, has published a thought-provoking analysis of how AI coding tools are reshaping the software development landscape. In his latest blog post, Geerling examines the unintended consequences of LLM-assisted development, particularly how it's creating a bifurcation in developer skill sets and eliminating crucial mentorship opportunities.

The AI-Assisted Development Experience

Geerling's insights come from firsthand experience using local LLMs to migrate 13,000 comments from Drupal to Hugo during a recent site migration. He tested models like GPT-OSS 20B and Qwen3 Coder 30B through Ollama on his Mac, finding them comparable to junior-to-midrange developers on a small team.

The speed difference was striking. Where a traditional development team might take days or even weeks to deliver working code for a feature, these LLMs could produce runnable code within 1-2 minutes. "Maybe not code that works, but it would be in close proximity to the code that meets all my requirements," Geerling notes.

This rapid iteration allowed him to complete the entire comment migration in just a few evenings—a task that would typically require significant developer time and coordination.

The Four Persistent Problems

Despite the impressive speed, Geerling observed that LLMs still struggle with the same fundamental issues that plague human developers:

Missed Requirements: LLMs often implement features that technically work but miss critical business requirements or edge cases that only become apparent through stakeholder review.

Suboptimal Implementations: Without the accumulated experience of building hundreds of sites, LLMs frequently choose approaches that work but aren't maintainable or efficient in the long term.

Premature Optimizations: LLMs sometimes spend computational effort optimizing code that doesn't need optimization, like spending hours refining a one-time migration script.

Burnout Management: While not applicable in the traditional sense, LLMs face their own version of burnout when context windows grow too large, requiring resets and restarts.

The New Developer Archetypes

Geerling identifies two emerging categories of developers in the AI-assisted era:

Expert Beginners: Junior developers who feel empowered by AI tools to tackle any programming challenge. They can produce working code quickly but lack awareness of the hidden pitfalls and footguns in their implementations.

Lone Wolf Developers: Experienced developers who leverage AI tools to work faster and more independently than ever before. These developers can orchestrate LLMs to build decent software quickly, but now have little incentive to work on teams or mentor junior developers.

The Missing Middle Class

The most concerning aspect of this shift, according to Geerling, is the disappearance of what he calls the "middle class" of developers—those who want stable careers building useful software without necessarily becoming technical architects or team leads.

These developers traditionally benefited from mentorship and the occasional "consternation" of senior developers who had already made and learned from the common mistakes. "You don't get that for free," Geerling emphasizes, noting that the best projects require people who've experienced catastrophic failures like bringing down entire sites with bad queries or deployments.

The QA Challenge

Geerling also highlights a critical gap in the AI development workflow: quality assurance. "Sycophant LLMs are not a substitute for senior devs," he writes, and they're particularly ill-suited for QA work. This is especially problematic given that many companies have already diminished the role of QA professionals, who often serve as the last line of defense against major project failures.

The Broader Implications

The post concludes with a sobering observation: there's now less of a clear path from Expert Beginner to Lone Wolf Developer. The traditional progression that involved learning from mistakes, receiving mentorship, and gradually building expertise is being disrupted.

Geerling questions whether aspiring to become a Lone Wolf Developer is even desirable, arguing that the software industry needs developers who can work collaboratively, mentor others, and build sustainable, maintainable systems rather than just shipping code quickly.

His analysis suggests we're entering an era where the economics of software development will favor speed and individual productivity over team collaboration and knowledge transfer—with potentially significant long-term consequences for software quality and developer career development.

The full post, including the original comment migration excerpt, can be found on Jeff Geerling's blog, where he promises to explore this topic further as the implications of AI-assisted development continue to unfold.