AI Talk: Code Assist – The Death of the Programmer?

As someone who has taught programming and AI for decades and worked as an industry researcher for just as long, the latest AI coding capabilities are both exhilarating and terrifying. How should programmers navigate these tools? How should computer-science curricula adapt? This post shares my experiences and reflections.

My experience with pair-programming

Almost four decades ago, using Symbolics AI (Lisp) machines at the Boeing Advanced Technology Center, two of my colleagues and myself experimented with pair-programming (or three some!) . One person would type and the other two would kibitz and comment! Was a fascinating experience, as all of us were trying to learn. And that we did.

Gemini Generated Image - Pair Programming

My Journey with Code Assistants

I used GitHub Copilot in Visual Studio Code for about two years. It spared me the hassle of remembering every Python function signature: I’d write a comment prompt and Copilot filled in the boilerplate. If the suggestion didn’t fit, I deleted it and wrote my own.

Over the past six months I’ve switched to Cline, an IDE extension that lets you plug in different LLMs. At first I still used OpenAI models behind the scenes—and was thrilled. I could prompt for a function, watch it create a Python script, run it, and—if it failed—see a side-by-side diff of the fixes before rerunning. A clear step forward.

Then came Claude 4, and everything felt turbo-charged. Now I can describe a goal—say, “Add a voice interface using the latest open-source speech-recognition model”—and Claude:

1. Reviews my current modules (I click Approve).

2. Proposes new modules (Approve).

3. Generates a requirements.txt (Approve).

4. Writes the code (Save).

5. Runs it (Run).

6. It works.

My entire cognitive input? One paragraph of description. Claude did the rest. I was flabbergasted—and uneasy. So much code, so little understanding of what it actually did.

Experimenting with Prompts

To regain control, I tried several strategies:

• Generate a high-level plan first.

• Ask for alternative algorithms or security reviews.

• Implement one step at a time, so I can inspect each piece.

• Request detailed READMEs explaining the code.

• Generate tests for every step to catch errors early.

All of this helps, but the uneasy feeling remains. Programming has changed irrevocably. Pandora’s box is open.

The Academic Perspective

I’ve long run graduate courses in AI and NLP—modeled loosely on Stanford’s deep-learning/NLP offerings. Students propose projects, then implement them solo or in teams. I also have them write research briefs.

That assessment model is now shaky. Traditional plagiarism detectors struggle with AI-written text, and with modern code assistants a project that once took weeks can be finished in a day. How do we grade comprehension rather than button-clicking?

What Students (and Developers) Must Still Master

If someone leans entirely on automation, their comprehension will be low—I’ve lived that firsthand. Yet the primary goal of a CS degree is still mastery of concepts:

• Inspect and debug generated code.

• Spot algorithmic weaknesses and security holes.

• Write thorough tests and identify edge cases.

• Improve and refactor the AI’s first draft.

A Pragmatic Workflow

1. State the goal clearly.

2. Ask the AI to draft a detailed plan.

3. Have it critique that plan. Explore algorithmic and security options.

4. Pick a plan and implement one step at a time.

5. Understand each step before moving on:

   • Read the code.

   • Ask the AI to explain it.

   • Demand documentation.

   • Write and run tests.

6. Proceed only when you’re confident. My mantra remains: “Don’t trust—and verify.”

The Assessment Dilemma

How should faculty evaluate students now? I’m honestly not sure. Slide-deck project reviews might still work—LLMs aren’t (yet) great at polished, context-rich PowerPoints, and live Q&A reveals genuine understanding. I’ll be experimenting the next time I teach.

Closing Thoughts

Whether you’re in a CS classroom or an industry team, the steps above can help you learn while coding. AI should be your partner, not your replacement. Supercharged programmers still need sharp minds.

Acknowledgement

This blog is entirely mine. No AI was used for the text — other than a proof reading pass - and that even surprises me!