jj4211

Well a light refactor of deduplicating code would be a refinement that is generally easy to review and likely comfortably in the capability of the codegen.

Versus creating, which generally is harder to review and requires the submitter to explain the rationale behind the change.

Volume of code and nature of change.

Of that comes to pass, they can change it.

The codegen is a particularly big problem for famous open source projects. All of a sudden you have a flood of people trying to get their name on your project and flooding you with low quality stuff. Codegen enables folks without good ideas to generate submissions that demand a lot of review. Being about to shoot down obviously AI content without deeper review is pretty critical to keep the project viable. There may be good things tossed out with the bathwater, but it is a ton of bathwater. Besides, if there’s something real behind it, a good human is likely to emerge to take it forward more credibly.

Closed source or personal projects can enjoy a bit more judgement. You can evaluate the people independent of the submissions and may be able to have an even more nuanced take.

It’s not like it is a legally binding agreement that needs a high level of quality in adjucation. If things seem off then it can be cited and you don’t have to prove it. Conversely, if something passes human review despite being codegen, well then I guess the codegen isn’t that bad

I’ve seen a pull request that was ultimately nonsensical but plausible and short, and purported to be needed in a configuration none of the maintainers had access to. It was a pretty harmless change either way so they were inclined to accept it. But they sought a little clarification and the clearly LLM response in comment got it shut down. No human stepped up to actually explain an understanding and the LLM sounded like a TV show writer writing tech, sentences that sounded credible but to anyone with understanding was nonsensical bullshit.

LLM code can sometimes be hard to tell in a short code change, but the prose when things are anything vaguely needing explanation is just so very very obvious.

Yeah, but you’ll have to submit to every known provider and hope the user used one of the ones that provide this checking service, and didn’t do something like ask a local model to just randomize synonyms in a text.

You still have choices.

Variable names are pretty free form.

A switch statement or if/else might be a choice that achieves the same thing. Waffling between them would be highly suspect, since a person isn’t going to be so wishy washy. So you could use structural choices too.

AI code tends to look more obviously AI than prose anyway.

Well, they got the 1 off-site bit…

Hence why openAI seems now to want to stall IPO as long as they can…

And you are right, they singlehandedly are most of the problem despite not even being in the strongest position even if you are bullish on AI.

Volumes are going to be down, but they want the same revenue, so they increase unit price, comfy in the reality that it will still be, percentage wise, not worth it for most to care. Will it make it unaffordable for some? Probably not appreciably more than the hardware is doing all by itself.

So in short, it is just because they can.

In the same way it’s accurate to call the human brain a prediction machine

I have seen no evidience presented that the human brain is simply making statistical predictions.

This seems to be a relatively new thing some people are asserting, because they see the LLM behavior and have decided that absolutely it looks so human like it must act the same, and since we understand that LLM is fundamentally statistical prediction of a token, then obviously human consciousness must operate on the same principle.

Even the latest models generate the sort of mistakes that stem from the fundamental limitations. A lot has been done in some contexts for making that not matter so much (e.g. in software development, the strategy is that code may have some verifiable goal, and the models can let the mistakes fly, then take the feedback from the goal, and iterate more). So it still can be very useful, but clearly it isn’t human like because of some of the completely dumb behaviors that result from it not actually thinking about it. It is at it’s best when either mistakes don’t matter (particularly fiction) or at least first mistakes don’t matter and can be automatically reconciled with facts.

By nature, the meme examples will get better because everyone talks about it and suddenly having all that discussion in training data and in search results, the statistics fix. However the operating principle behind them remains the same.

LLM output makes everything stupider, even LLMs.

Sadly, often.

It is sticking the most common content into a blender and dispensing the result according to a few tweaks. I have seen a lot of people eat that up.

Same reason why there’s so much money in low-effort mobile gaming.

It’s one of the things that really sucks, people with creative talent that would otherwise be rewarded and have their vision shared for lack of too many other options being drowned out by a flood of no-effort slop that the masses are just fine with.

For reference, that was the year Vista came out with a minimum memory of 512MB, with 1GB suggested.

You have to remember that the current investment level assumes that somehow everyone is going to use the AI stuff that much. It won’t do if the only heavy usage is code generation.

This likely refers to the famous fictitious quote from Heny Ford that wanna-be self-important “innovators” like to trot out to show how they are smarter than the customer when they put out some unasked for bullshit.

"If I had asked people what they wanted, they would have said faster horses.”

People like to imagine the car as some miraculous thing no one but a genius could have imagined wanting and the genius descending from on high and delivering. Truth was people could easily imagine and want it, the genius was in knowing how to make what people knew they wanted.

Instead, Henry Ford said: "If there is any one secret of success, it lies in the ability to get the other person’s point of view and see things from that person’s angle as well as from your own”

So essentially the exact opposite of “screw the customers, we know better!”

It sounds like how students think professional sounds.

Well, also some executives that I can’t stand that, if I didn’t know better, would swear were ChatGPT for decades…

Though with the ability to not even have to think about synonyms and indirect references or foreign language or potentially if someone sent you a damn screenshot of a terminal output instead of the text…

I remember a news piece where they were interviewing someone about their renouned AI work that was not LLM related. The interviewer clearly struggled and while the person conducting himself well, you knew he had to be annoyed at having to explain his work had nothing to do with the chatbots, datacenter buildout, and related resource consumption.

As little as possible of my work is boilerplate, but some times there will be something like “I want to take width,height, and depth as arguments from the command as variables”, and poof, all the argv tedium is done.

This is a point that really sticks with me. Using it for the sometimes spot on cakewalk segments is a fairly productive win. By the time you stubbornly insist on driving it entirely blackbox with chat and trying to get the right results without actually touching code… Well, even when it works, it’s often more work than just doing it yourself.

Someone rebased a UI I worked on in a new version of the UI framework. As a result, there was this one odd gap in the UI in one specific place. A vibe coder spent 3 hours back and forth with the AI trying to get it to correct the gap and finally submitted their merge request. Hundreds and hundreds of lines of CSS. So I declined the merge request, open the gui, looked at the gap, hit f12, adjusted a single padding statement, and it was all good. People are struggling with defining all sorts of criteria and rigging it to let it try and try and try again and hopefully laid out every contingency, every corner case, and spent hours laying the ground work and could have done similar in a more straightforward way.

You are right, but right now we are living in a very messy reality where it’s hard to know who’s being stupid and who is using it well. One person I know that, well, I was never a huge fan of his work but at least it was somewhat serviceable is now all-in on AI and his code has been rewritten in a similar manner as the parent poster comments. He’s got no idea how it works or how it should worked, the AI decided to rewrite it in an entirely different language, and it’s a buggy mess and it never fixes the bugs without making new bugs. Then he hit his token quota 3 weeks early and basically said he was going to stop working on it because he no longer could manually work the codebase. He didn’t ask for a rewrite, but AI advised him that his language choice was a poor fit and reworked things in another language, one that none of us use to that level of seriousness. It also made it largely based on super convoluted regular expressions.

The problem is that the leadership is singing the praises of these people, they were on their ‘leaderboards’ of AI adoption and much like the craze of praising “lines of codes”, we are neck deep in the most stupid non-technical evaluation of technical work you could imagine.