Hacker News Reader

Comment by basilgohar | Hacker News Reader

I feel like this is mostly proofless assertion. I'm aware what you hint at is happening, but the conclusions you arrive at are far from proven or even reasonable at this stage.

For what it's worth, I think AI for code will arrive at a place like how other coding tools sit – hinting, intellisense, linting, maybe even static or dynamic analysis, but I doubt NOT using AI will be a critical asset to productivity.

Someone else in the thread already mentioned it's a bit of an amplifier. If you're good, it can make you better, but if you're bad it just spreads your poor skills like a robot vacuum spreads animal waste.

View on HN

I think that was his point, the project full of bad developers isn't the competition. It is a peer whose skill matches yours and uses agents on top of that. By myself I am no match for myself + cline.

That’s true in the short term. Longer term it’s questionable as using AI tools heavily means you don’t remember all the details creating a new form of technical debt.

Dude, have you ever looked at code you wrote 6 months ago and gone "What was the developer thinking?" ;-)

yes, constantly. I also don't remember much contextual domain info of a given section of code about 2 weeks into delving into some other part of the same app.

So-called AI makes this worse.

Let me remind you of gyms, now that humans have been saved of much manual activity...

>So-called AI makes this worse.

The AI tooling is also really, really good at being able to piece together the code, the contextual domain, the documentation, the tests, the related issues/tickets, it could even take the change history into account, and be able to help refresh your memory of unfamiliar code in the context of bugs or new changes you are looking at making.

Whether or not you go to the gym, you are probably going to want to use an excavator if you are going to dig a basement.

> So-called AI makes this worse.

I think that needs actual testing. At what time distances is there an effect, and how big is it? Even if there is an effect, it could be small enough that a mild productivity boost from AI is more important.

Exactly. Claude code can explain code I've written to me better than I could. I feel like people who don't see AI as a transformative element to programming probably didn't experience what it can do today as opposed to 6 months or a year ago. It's night and day difference. And it still was useful back then

Yeah 6 months ago Claude could make me a rust function that wouldn't compile but got me pointed in the right direction. Now it will make it correct with comments and unit tests with idiomatic style just using chat. But we don't have to use chat. Even open models today like devstral when combined with an agent can run cargo check and clippy and self prompt (with rusts great error messages) to fix everything. Prompting it with some unit test cases lets it iterate until those pass too. Software development has fundamentally changed. I still would advise developers who care about performance to be able to read asm. But just like I wouldn't write asm anymore, because the llvm optimiser is really good, we are going to get to a point where designing the test cases will be the same as developing the software.

I don't need to remember much, really. I have tools for that.

Really, really good tools.

IMO LLMs are best when used as locally-run offline search engines. This is a clear and obvious disruptive technology.

But we will need to get a lot better at finetuning first. People don't want generalist LLMs, they want "expert systems".

Speak for yourself, I prefer generalist LLMs. Also, the bitter lesson of ML applies.

> I prefer generalist LLMs.

No you don't. You personally probably don't need a Pokemon encyclopedia or Bengali spell checking in your daily LLM usage.

These are the kinds of things you're paying for when you're using the newer models with huge parameter counts.

while I probably won't need pokemon knowledge from an llm, I still think broad pre-training enables unpredictable cross-domain knowledge transfer - seemingly irrelevant data can contribute to reasoning abilities that improve specialized tasks. and most developers work across domains anyway, so better to start general and post-train than lose emergent capabilities from narrow foundations.

even within 'coding' there's huge breadth. you might need to understand database schemas, api documentation, domain-specific algorithms, legacy system constraints, or business logic from completley different industries. a narrowly-tuned model might excel at leetcode problems but struggle when you need to build a financial trading system or whatever.

We need 1000 different specialized LLMs, not one huge LLM with 1000 times more parameters.

Whoever figures this out from a technical point of view will probably win this AI bubble iteration.

Also, the big money is in local inference, not cloud chat. The pendulum will swing back towards personal computing. Whoever figures out the solution to this technical problem will also win.

(I have no use whatsoever for C# or Java knowledge. I would pay for LLM weights that can figure a complex Nix codebase, preferably my own. I want to buy the weights though, cloud is a complete non-starter here.)

Reply on news.ycombinator.com