The amount of context switching in my day-to-day work has become insane. There's this culture of “everyone should be able to do everything” (within reason, sure), but in practice it means a data scientist is expected to touch infra code if needed.

Underneath it all is an unspoken assumption that people will just lean on LLMs to make this work.

They can give plausible architecture but most of the time it’s not usable if you’re starting from scratch.

When you design the system, you’re an architect not a coder, so I see no difference between handing the design to agents or other developers, you’ve done the heavy lifting.

In that perspective, I find LLMs quite useful for learning. But instead of coding, I find myself in long sessions back and forth to ask questions, requesting examples, sequence diagrams .. etc to visualise the final product.

My first thought was that I can abstract what I wrote yesterday, which was a variation of what I built over the previous week. My second thought was a physiological response of fear that today is going to be a hard hyper focus day full of frustration, and that the coding agents that built this will not be able to build a modular, clean abstraction. That was followed by weighing whether it is better to have multiple one off solutions, or to manually create the abstraction myself.

I agree with you 100 percent that the poor performance of models like GPT 4 introduced some kind of regularization in the human in loop coding process.

Nonetheless, we live in a world of competition, and the people who develop techniques that give them an edge will succeed. There is a video about the evolution of technique in the high jump, the Western Roll, the Straddle Technique, and finally the Fosbury Flop. Using coding agents will be like this too.

I am working with 150 GB of time series data. There are certain pain points that need to be mitigated. For example, a different LLM model has to be coerced into analyzing or working with the data from a completely different approach in order to validate. That means instead of being 4x faster, each iteration is 4x faster, and it needs to be done twice, so it still is only 2x faster. I burned $400 in tokens in January. This cannot be good for the environment.

Timezone handling always has to be validated manually. Every exploration of the data is a train and test split. Here is the thing that hurts the most. The AI coding agents always show the top test results, not the test results of the top train results. Rather than tell me a model has no significant results, it will hide that and only present the winning outliers, which is misleading and, like the OP research suggests, very dangerous.

A lot of people are going to get burned before the techniques to mitigate this are developed.

Overfitting has always been a problem when working with data. Just because the barrier of entry for time series work is much lower does not mean that people developing the skill, whether using old school tools like ARIMA manually or having AI do the work, escape the problem of overfitting. The models will always show the happy, successful looking results.

Just like calculators are used when teaching higher math at the secondary level so basic arithmetic does not slow the process of learning math skills, AI will be used in teaching too. What we are doing is confusing techniques that have not been developed yet with not being able to acquire skills. I wrack and challenge my brain every day solving these problems. As millions of other software engineers do as well, the patterns will emerge and later become the skills taught in schools.

For hobby projects though, it's awesome. It just really struggles to do things right in the big codebase at work.

And then you find out someone else had already solved it. So might as well use the Google 2.0 aka ChatGPT.

> "We collect self-reported familiarity with AI coding tools, but we do not actually measure differences in prompting techniques."

Many people drive cars without being able to explain how cars work. Or use devices like that. Or interact with people who's thinking they can't explain. Society works like that, it is functional, does not work by full understanding. We need to develop the functional part not the full understanding part. We can write C without knowing the machine code.

You can often recognize a wrong note without being able to play the piece, spot a logical fallacy without being able to construct the valid argument yourself, catch a translation error with much less fluency than producing the translation would require. We need discriminative competence, not generative.

For years I maintained a library for formatting dates and numbers (prices, ints, ids, phones), it was a pile of regex but I maintained hundreds of test cases for each type of parsing. And as new edge cases appeared, I added them to my tests, and iterated to keep the score high. I don't fully understand my own library, it emerged by scar accumulation. I mean, yes I can explain any line, but why these regexes in this order is a data dependent explanation I don't have anymore, all my edits run in loop with tests and my PRs are sent only when the score is good.

Correctness was never grounded in understanding the implementation. Correctness was grounded in the test suite.

> AI assistance produces significant productivity gains across professional domains, particularly for novice workers. Yet how this assistance affects the development of skills required to effectively supervise AI remains unclear. Novice workers who rely heavily on AI to complete unfamiliar tasks may compromise their own skill acquisition in the process. We conduct randomized experiments to study how developers gained mastery of a new asynchronous programming library with and without the assistance of AI. We find that AI use impairs conceptual understanding, code reading, and debugging abilities, without delivering significant efficiency gains on average. Participants who fully delegated coding tasks showed some productivity improvements, but at the cost of learning the library. We identify six distinct AI interaction patterns, three of which involve cognitive engagement and preserve learning outcomes even when participants receive AI assistance. Our findings suggest that AI-enhanced productivity is not a shortcut to competence and AI assistance should be carefully adopted into workflows to preserve skill formation -- particularly in safety-critical domains.

But that's what "impairs learning" means.

you can also read human code and empathise what were they thinking while writing it

AI code is not for humans, it is just a stream of tokens that do something, you need to build skills to empirically verify that it does what you think it does, but it is pointless to "reason" about it.

Yes, we can use it 10,000 times to refine our recipes, but "did we learn from it"? I am doubtful about that, given that even after running with the same prompt 10 times, it will give different answers in 8/10 responses.

But I am very confident that I can learn by iterating and printing designs on a 3D printer.

I’m starting to believe that people who think AI-generated code is garbage actually don’t know how to code.

I hit about 10 years of coding experience right before AI hit the scene, which I guess makes me lucky. I know, with high confidence, what I want my code to look like, and I make the AI do it. And it does it damn well and damn fast.

I think I sit at a unique point for leveraging AI best. Too junior and you create “working monsters.” Meanwhile, Engineering Managers and Directors treat it like humans, but it’s not AGI yet.