The Vision vs. The Reality
In a new interview, Andrej Karpathy painted a provocative picture of AI's future: autonomous agents conducting their own research, designing experiments, and improving themselves without human intervention. His AutoResearch project represents the cutting edge of this vision—AI systems that can close the loop on research entirely.
Yet while industry leaders race toward this autonomous future, individual developers are grappling with more immediate concerns. A computer science student's question on Reddit crystallized the tension: "Are there people here who still write the code by themselves?"
This isn't just generational anxiety—it's a fundamental question about the future of software development.
The Speed Problem Reveals Deeper Issues
The practical challenges of agentic coding are already surfacing. One developer running Qwen3-Coder-Next locally reported that while the model excels at chatting, it takes "over 5 minutes" to make specific codebase changes, spending excessive time searching individual functions.
"Isn't there like some system which scans your codebase and it can use it as an index... so the AI knows already where to look for specific stuff so it's faster?" they asked.
This performance gap between promise and practice highlights a critical issue: current AI coding agents often lack the contextual awareness and efficiency that human developers take for granted. The irony is palpable—tools meant to accelerate development can actually slow it down.
The Learning Dilemma
Perhaps the most thought-provoking angle comes from that computer science student's concern about maintaining fundamental skills. They want to "utilize AI" while still "typing codes, having an idea and see if it works, debugging etc."
This tension echoes our recent coverage of how code evaluation has become more valuable than generation. But it goes deeper—if students skip the fundamentals and jump straight to AI assistance, will they develop the judgment needed to evaluate AI output?
Karpathy addressed this indirectly when discussing "Relevant Skills in the AI Era" at the 28-minute mark of his interview. The implication is clear: the skills that matter are changing, but we're not sure what to preserve and what to abandon.
Advanced Features, Basic Questions
Meanwhile, developers are discovering sophisticated capabilities in existing tools. Google's Gemini CLI now offers advanced features like skills, hooks, and plan mode. Claude Code users are finding "fully customizable status bars"—small quality-of-life improvements that suggest these tools are maturing rapidly.
Yet the fundamental question remains: Are we building tools that augment human capability or replace human understanding?
What's Next
The trajectory is clear but conflicted. On one hand, we're moving toward Karpathy's vision of autonomous AI agents that can handle entire research workflows. On the other, individual developers are struggling with basic questions about skill development and tool efficiency.
This isn't just about technology—it's about the future relationship between humans and AI in creative work. As one Reddit commenter put it, the challenge is "pushing myself to build things from scratch" while staying competitive with those who "use AI heavily."
The answer may lie not in choosing sides, but in developing new frameworks for human-AI collaboration that preserve essential skills while embracing acceleration. Until then, the paradox deepens: the more capable our AI tools become, the more we question what capabilities we should retain.