IF (strong if) AI begins replacing coders, AI will code in it's own optimized language, I speculate something like minimized JS or WASM. It doesn't have to be human readable, it just needs to pass unit tests. So humans will probably only write unit tests, and AI will write code. So knowing programming will not be necessary anymore, only prompt-derived unit tests.
I tend to liken them to very drunken scholars. They know things, usually at about a Wikipedia level, and they’re cheery too. But they lack a capacity to doubt themselves; they often are confidently wrong.
Often the greatest help is understanding natural language, but given its hiccups… time using it is probably best spent using it as a drunken librarian – to teach how to phrase and fetch information
As one telling recent example, I tried using an LLM to help with some jq (with which I’m rusty); it got a few basics and then repetitively tripped over a syntax hiccup on loop, “correcting” itself to the same answer each time. A StackOverflow search or two, for comparison, answered my questions and taught some new syntax too. Probably took less time, but more critical thought.
That, coupled with the fact LLMs tend to give an answer and then also an unnecessary verbose step-by-step, means I tend to dislike them.
I also have a huge bugbear about “AI” as a term because it tells you very little. Plenty of applied statistics (markov chains, clustering algos, deep learning eg computer vidion; even SearchRank) are used heavily in research and other cases to do a lot of good. Even for the layman: the Seek app by iNaturalist is awesome for identifying common plant species; Stockfish is (now) a NN that dominates in chess.
But these are classifiers, not generators. By their very nature it is just statistics to evaluate a classifier on a test dataset. Generators, however, are far, far thornier to test, and seem a lot more prone to overfitting.
While I’m not familiar with a typical trained generator tensor, I imagine the optimal one will be surprisingly sparse, though not in a structured way - corresponding to a more clustered “small world” network, which IRL seem the most productive.
Only when they can also dismissively sneer at a marketing person will they be truly ready to replace programmers.
"coding" is like learning cursive and how to spell words, no matter how good you are at that it won't make you a NYT best seller.
I think a great many more experienced people forget just how hard it is to learn your first, just how many new unintuitive concepts appear no matter what language you choose. Many, many more people are now able to learn that without getting frustrated to the point of giving up, because they all have a digital Aristotle in their pocket to guide them.
We will probably also see renewed interest in the "out of the box" options available on common platforms like Debian. By my count those would be awk, sed, and perl. [1] There's just something very satisfying about being able to spin up a bare server and get useful work done with just the essentials!
[1]: https://hiandrewquinn.github.io/til-site/posts/what-programm...
I don't need to Google search a particular issue in the hope that someone else had the same issue and someone had a solution. I don't need to wade through unprofessional comments on forums.
Having an LLM is almost like having a tutor I can ask questions, even silly questions. I can ask it to make inferences between particular mistakes and I can dive deeper. I can also supplement the LLM with Google searches. I can ask it to reframe code if I want it to or help me fix certain mistakes, or understand particular concepts better.
I'd be excited if I was going to University and had coding subjects knowing I had an LLM I could ask for help.
I very much doubt we're going to see a massive shift where everyone becomes a system analyst or service designer and we just punch in business requirements out comes a ready-to-release system.
I can see automated ui testing tools becoming truly amazing if AI Agents are even half of what they're hyped up to be. At the moment they kind of work, but also a bit of a headache.
There are also tools like https://rockyai.me/ that are making LLMs accessible within your browser so you don't even have to context switch between chat gpt and the tab you are on to ask LLMs questions. I use this all the time when reading engineering blog and I want to dig more into a specific section that I don't understand
Even when the AI was technically right it often didn’t solve my problem because I didn’t understand how to ask the right question. I didn’t know exactly how to step back and ask at the right level of specificity. “Solve this tiny narrow problem” doesn’t help me and I often omit broader context that an expert would have added for this situation. Going too broad doesn’t actually get to the specifics of what I need and misses important details.
The trial and error of formulating a meaningful question, rethinking questions, working through hallucinations, and digging into fundamentals from authoritative sources all worked well together.
For some, they never have to learn. For others, it makes everything click. Maybe they didn't have a good teacher or book. But now they can experiment and play around with different variables and wonder. Then other things start to click - angles, friction, and so on.
I think it should accelerate things to either the point where people say good riddance or where they can build more things on top of it.
It’s the world’s greatest rubber duck and even if that’s all it ever is, it’s a game changer for me.
I'm currently teaching a "Programming with AI" course where we're using Cursor with no restrictions. I now think "learning programming" as we've traditionally conceived it is toast. Core CS projects, the kind that would have been at the heart of the curriculum, take students maybe 30 minutes to do with modern tools.
Previously, the hard part of learning programming was developing the skill of putting code statements in the correct logical structure, and building that up from small programs of a few lines, to functions/classes, and eventually to larger programs. That happened in parallel with building the knowledge background of systems, libraries, algorithms, databases, etc. that you had to draw on to write complex applications.
The core work of the middle part of the CS major - building skill by writing progressively more complex functions and classes - can be largely automated at this point. You've still got to learn things, but the emphasis shifts. Details about libraries and frameworks - implementation at the level of functions and classes - are deemphasized. The scope of what my students can do is bigger, so they're dealing with the challenges of designing and debugging larger applications.
For new programmers the most important thing to learn is a mental model of how programs execute: Can you look at a small program and understand what it does? This is a prerequisite to generating bigger programs with AI. Students also need to learn to use AI collaboratively, to think through a problem like a pair programmer, rather than expecting to get complete one-shot generations.
I redesigned our intro course to move most of our core skills practice (what would previously have been homework problems) into class time. The out-of-class projects are now bigger and more ambitious, but also give students specific steps and prompts for using AI. A major point of the projects is agency. I like projects where there isn't one right answer, but instead students have to set a vision and then iterate on it.
Beyond the first course, we now need to spend more time on "software engineering" and craftsmanship:
- Clarifying requirements
- Reasoning about the design of a larger program with many parts
- Communication between parts of the application: DB design, APIs
- Debugging intuition, why is this thing breaking?
- Testing
- Working with bigger codebases
These things have always been around, but we often didn't teach them until upper-level courses. Early-level projects weren't complex enough to require serious craftsmanship. AI lets us do more ambitious things earlier in the curriculum - we should be looking for ways to raise our standards and continue challenging students.
For new programmers, I would encourage learning how to write standard single-file programs (with variables, loops, functions, lists, etc.) without using AI. Then bring in AI as a partner to write larger applications with specific libraries. Focus on raising the scale of your programs to the point where you can't immediately one-shot them, and then use the difficulties you encounter to think about design issues. Gradually build your knowledge base of systems and algorithms, but don't obsess about memorizing implementation details right away. Build end-to-end applications that do things; LeetCode problems have always been terrible.
For my fellow professors: You should start talking about how to redesign your courses and curriculum. That requires taking some guardrails off so you can see what students can really do with modern AI tools.
