No, you define what you want in project planning and briefs, coding is the interpretation of your definition. It is quickly becoming far easier and definitely faster for a machine to interpret what we define than for us to translate our definition into what a machine can interpret.

LLMs aren’t “in their infancy”. They’re tapped out.

And 64k oughta be enough for everyone.

You switch to LLM’s at your convenience but you tripped over the term “AI”. We’ve been over this a few times already and I hate repeating myself.

We can boil the issue down to a very simple question, do you think in time AI will play a significant role in how we generate code?

If the answer is no, then I’ll see you in ten years, if the answer is yes, then you should admit that GitHub choosing that term is not out of place and it is only self evident that they use what is currently the best approach to produce code/assistance while putting it under the “AI” banner for their long term vision and because it wants and needs to ride the hype train.

All the arguments I hear are largely pedantry and contrarianism. You see this every time something new and exciting pops up, people will huff and puff about small issues while losing track of the larger picture. The way you choose your words makes it obvious that this is just another case of that. No nuance, no, just “this is trash”, as if completely oblivious to the fact that in the time it took you to type those 3 words, a million people received an answer from an LLM that would otherwise take them 5 minutes to Google.

You can’t decouple quality and productivity, because code that isn’t of sufficient quality is not useful, and the debt of bad code costs many, many more times more work than doing code correctly. Low quality code isn’t “doing the job”.

But you have no idea whether the code generated is of such low quality that it offsets the time it took to produce it. That is just another assertion. For someone who is so adamant about the precision of code, you sure do throw around a lot of unfounded beliefs.

the debt of bad code costs many, many more times more work than doing code correctly

Like this gem for instance. Not only do you build on the unfounded premise that AI generates bad code, it also assumes a coder does not. On top of that, how much bad code? How many more times? Shouldn’t there be some quantification in all this rhetoric?

No, code is not imprecise. It does exactly what you tell it to every time.

And that is exactly why it is imprecise, because it’s a human conceiving it. You don’t want code to do what you type, you want code to do what you define. It is easy to define what a program needs to do, it is not as easy to then translate that to something a machine understands. You are doing the interpretation for the machine, that is all that coding is and we will look back on this approach as comical. Now that machines have the ability to understand what we define, we can skip the harder steps and focus on building things instead of playing Rosetta stone and beating ourselves on the chest because we consider ourselves to be champions at it.

What you are doing is comparing the perfect coder with LLM’s in their infancy and then conclude that the former makes less mistakes, I’m not sure why I have to point out that that is an unfair comparison.

I can make the same broken comparison about AI generating images. Will a Picasso produce beter art? Of course, for the time being, but the AI generates in seconds what we humans do in hours or days. And for a very wide base of what we do, the AI is already sufficient in its job even though the technology is young.

I’m sure you’ll agree that everywhere a form of AI has been implemented, from playing chess, go and StarCraft, to medical imaging, folding proteins… whatever, it quickly surpassed the quality of its human counterpart. Compared to those examples, generating code is a relatively easy task. And yes I understand that those use different “AI” than LLM’s.

The study you’re linking completely ignores code quality.

The study shows that your claim about productivity is false, now you’re moving the goalposts. You’ve made a lot of claims, but it all stems from a narrow distaste in how LLM’s function and you haven’t backed anything up.

That’s the “AI” GitHub uses, that they’re referring to, that is a stronger reason not to use their platform than to use it.

You’re missing the point. You tripped over the word “AI”, then equate it to just an LLM and on top of that you claim that nobody is getting any use out of it. Not only is your argument circular, it’s also based on a false premise.

The entire point of code is to clearly and effectively communicate what you want.

No, the point of code is to arrive at software that does what you want. Currently, we have to describe what our software needs to do, then mangle code into doing what we want. AI, and even an LLM, has the ability to take over everything after we provide the description of what we want and even write the tests to make sure it does that. The billions spent on bug hunting, quality assurance, acceptance testing and liability cases clearly show that it is not easier than natural language. Something we start learning before we’re even born.

But copilot and others are not just tools to spit out code, they are a replacement for search engines with the ability to not only instantly provide you with a relevant answer, but also to explain their reasoning with the ability to go back and forth about details that would otherwise take you through multiple Google searches and trawling through different websites and fora to maybe distill an answer. Clearly it goes without saying that this interface with what “the internet” knows is a major step forward to how we find and apply relevant information.

natural language is imprecise by definition

But so is code unless we write it to be precise. And it is far more easy and productive to define what that precision needs to be than it is to write and test. A project without unit tests is half the price of a project with tests, that alone should tell you something about the idea of precise code being easy. Knowing full well that every bit of software starts by defining it in natural language anyway. It goes without saying that if code and test generation is automated after that initial step, productivity is increased massively.

Every attempt to demonstrate the LLMs improve productivity in software development fails miserably and shows that it doesn’t do that. It’s not capable of doing that.

The main finding was that programmers who did not use AI completed the task in 160.89 minutes (2.7 hours) on average, whereas the programmers who had AI assistance completed the job in 71.17 minutes (1.2 hours). The difference between the two groups was statistically significant at the level of p = 0.0017.

https://www.nngroup.com/articles/ai-programmers-productive/

Just one example by the way…

LLMs are what you’re advocating for, because it’s what Copilot is.

I was afraid you’d say this, but I gave you the benefit of the doubt. It doesn’t matter what copilot is, you tripped over the word “AI”, then reduced it to LLM’s, and are now full circle by saying copilot is an LLM.

I think my original response to you was that you were short sighted in your argument, and this latest comment just underlines that you have issues with what AI is now, not what it is becoming.

It doesn’t lead to better software, it doesn’t lead to more efficient development

Eventually it will be all we need to write software.

All for obscene energy draws to zero benefit.

Oldmanyellsatcloud.jpg

I’ve gotten plenty benefit out of LLM’s, and millions of people with me, maybe you’re doing it wrong? Why do you think this absurd amount of power usage can be justified? Don’t you think interest and actual usage are the reason?

Zero people are on GitHub because of Copilot.

Zero people were on the internet because of Google. Zero people were into photography because of digital sensors. And zero people took the trip to work because of cars.

Ridiculous argument.

It’s not irrelevant whether it scales.

It is irrelevant because it is a straw man. They didn’t say LLM, they said AI.

Like I said, it’s really trendy to go “LLM’s just predict the next letter!”, but please get over yourself, this is not an insightful argument but just chest beating and pedantry over nothing.

AI is exactly none of GitHub’s value.

Because nobody is using Co-pilot happily?

LLMs are dogshit tech that doesn’t scale.

It’s irrelevant whether it scales or not. Your insinuation that this is all a pointless endeavor destined for failure is incredibly short-sighted.

Because it says AI?

I know it’s currently trendy to show consternation when the topic is brought up, and yes, every service and app is currently milking the AI hype, but what do you expect, really? The technology is not going anywhere, it can already do amazing things and will continue to improve. All we can do is deal with the issues and steer it in the right direction, pretty soon we will not consider it any more spectacular than the calculator in Windows 3.11 and we will look at our past interfacing with computers as primitive.

It’s just a tighter grouping of (biased) data that can be searched and retrieved a bit quicker.

How is your intelligence different from being “biased data that can be accessed”?

The fact that something can reason about what it presents to you as information is a form of intelligence. And while this discussion is impossible without defining “reason”, I think we should at least agree that when a machine can explain to you what and why it did what it did, it is a form of reason.

Should we also not define what it means when a person answers a question through reasoning? It’s easy to overestimate the complexity of it because of our personal bias and our ability to fantasize about endless possibilities, but if you break our abilities down, they might be the result of nothing but a large dataset combined with a simple algorithm.

It’s easy to handwave the intelligence of an AI, not because it isn’t intelligent, but because it has no desires, and therefore doesn’t act unless acted upon. It is not easy to jive that concept with the idea that something is alive, which is what we generally require before calling it intelligent.

You’re missing the point. The issue with Fahrenheit is not about the conversion from Celsius, most Europeans don’t need to do that anyway. The problem is Fahrenheit in itself, it’s just not elegant or scientific and therefore comes off as arbitrary and only makes sense when you grow up with it.

Haha, I get it, it’s totally fine!

I don’t think it’s funny, because the joke is illogical. If he is a teacher in a University, and it looks like it, that it is his job to mansplain.

So close. You seem to have Sheldon levels of understanding sarcasm.

It’s just a simple joke about the term being misapplied to an everyday setting.