The number of times I have received an un-proofread two sentence email is too damn high.

Just a few trillion more dollars, bro. We’re almost there. Bro, if you give up a few showers, the AI datacenter will be able to work perfectly.

Bro.

My jr developer will eventually be familiar with the entire codebase and can make decisions with that in mind without me reminding them about details at every turn.

LLMs would need massive context windows and/or custom training to compete with that. I'm sure we'll get there eventually, but for now it seems far off. I think this bubble will have to burst and let hardware catch up with our ambitions. It'll take a couple of decades.

AI is more likely to generate code that’s hard to follow and therefore harder to check.

Sure. It's making the errors faster and at a far higher volume than any team of humans could do in twice the time. The technology behind inference is literally an iterative process of turning gibberish into something that resembles human text. So its sort of a speed run from baby babble into college level software design by trial, evaluation, and correction over and over and over again.

But because the baseline comparison code is, itself, full of errors, the estimation you get at the end of the process is going to be scattering errant semicolons (and far more esoteric coding errors) through the body of the program at a frequency equivalent to humans making similar errors over a much longer timeline.

Also seems like it'd be a lot harder to modify or extend later

The way it was described to me by some academics is that it’s useful…but only as a “research assistant” to bounce ideas off of and bring in arcane or tertiary concepts you might not have considered (after you vet them thoroughly, of course).

The danger, as described by the same academics, is that it can act as a “buddy” who confirms you biases. It can generate truly plausible bullshit to support deeply flawed hypotheses, for example. Their main concern is it “learning” to stroke the egos of the people using it so it creates a feedback loop and it’s own bubbles of bullshit.

That's not a bad idea. I'm already downloading lots of human knowledge and media that I want backed up because I can't trust humanity anymore to have it available anymore

Yeah what you say makes sense to me. Having it make a "wrong start" in something new is useful, as it gives you a lot of the typical structure, introduces the terminology, maybe something sorta moving that you can see working before messing with it, etc.

It is great for boilerplate code. It can also explain code for you, or help with an unfamiliar library. It's even helped me be productive when my brain wasn't ready to really engage with the code.

But here's the real danger: because I've got AI to do it for me, my brain doesn't have to engage fully with the code anymore. I don't really get into the flow where code just flows out of your hands like I used to. It's becoming a barrier between me and the real magic of coding. And that sucks, because that's what I love about this work. Instead, I'm becoming the AI's manager. I never asked for that.

You want to see someone using say, VS Code to write something using say, Claude Code?

There's probably a thousand videos of that.

More interesting: I watched someone who was super cheap trying to use multiple AIs to code a project because he kept running out of free credits. Every now and again he'd switch accounts and use up those free credits.

That was an amazing dance, let me tell ya! Glorious!

I asked him which one he'd pay for if he had unlimited money and he said Claude Code. He has the $20/month plan but only uses it in special situations because he'll run out of credits too fast. $20 really doesn't get you much with Anthropic 🤷

That inspired me to try out all the code assist AIs and their respective plugins/CLI tools. He's right: Claude Code was the best by a HUGE margin.

Gemini 3.0 is supposed to be nearly as good but I haven't tried it yet so I dunno.

Now that I've said all that: I am severely disappointed in this article because it doesn't say which AI models were used. In fact, the study authors don't even know what AI models were used. So it's 430 pull requests of random origin, made at some point in 2025.

For all we know, half of those could've been made with the Copilot gpt5-mini that everyone gets for free when they install the Copilot extension in VS Code.

What about my comment made you believe I was using gut feelings to judge anything? My ticket completion rate, number of tickets, story points, and number of projects completed all point to massive productivity gains.

It depends on the task. As an extreme example, I can get AI to create a complete application in a language I don’t know. There’s no way that’s not more productive than me first learning the language to a point where I can make apps in it. Just have to pick something simple enough for the AI.

Of course the opposite extreme also exists. I’ve found that when I demand something impossible, AI will often just try to implement it anyway. It can easily get into an endless cycle where it keeps optimistically declaring that it identified the issue and fixed it with a small change, over and over again. This includes cases where there’s a bug in the underlying OS or similar. You can waste a huge amount of time going down an entirely wrong path if you don’t realize that an idea doesn’t work.

In my real work neither of these really happen. So the actual impact is much less. A lot of my work is not coding in the first place. And I’ve been writing code since I was a little kid, for almost 40 years now. So even the fast scaffolding I can do with AI is not that exciting. I can do that pretty quickly without AI too. When AI coding tools appeared my bosses started asking if I was fast because I was using one. No, I’m fast because some people ask for a new demo every week. Causes the same problems later too.

But I also do think that we all still need to learn how to use AI properly. This applies to all tools, but I think it’s more difficult than with other tools. If I try to use a hammer on something other than a nail, it will not enthusiastically tell me it can do it with just one more small change. AI tools absolutely will though, and it’s easy to just let them try because it’s just a few seconds to see what they come up with. But that’s a trap that leads to those productivity wasting spirals. Especially if the result actually somehow still works at first, so we have to fix it half a year later instead of right away.

At my work there are some other things that I feel limit the productivity potential of AI tools. First of all we’re only allowed to use a very limited number of tools, some of them made in-house. Then we’re not really allowed to integrate them into our workflows other than the part where we write code. E.g. I could trivially write an mcp server that interacts with our (custom in-house) ci system and actually increases my productivity because I could save a small number of seconds very often if I could tell an AI to find builds for me for integration or QA work. But it’s not allowed. We’re all being pushed to use AI but the company makes it really difficult at the same time.

So when I play around with AI on my spare time I do actually feel like I’m getting a huge boost. Not just because I can use a claude model instead of the ones I can use at work, but also just basic things like e.g. being able to turn on AI in Xcode at all when working on software for Apple platforms. On my work Macbook I can’t turn on any Apple AI features at all so even tab completion is worse. Or in other words, those realities of working on serious projects at a serious company with serious security policies can also kill any potential productivity boost from AI. They basically expect us to be productive with only those features the non-developer CEO likes, who also doesn’t have to follow any of our development processes…

Because it has been like this before vibe coding existed...

That's a valid question, especially with AI coding being so prevalent.

It's been tested a lot and the results are that it can't be trusted at all unless you are already an expert in the thing you're asking it to "help" you with so you can correct the many mistakes it will make, but it's slower and, again, is **guaranteed **to make mistakes (hallucinations are built into what techbros are insisting on labeling as "AI", no matter how many resources you throw at it).

All of this at great environmental and human cost too.

I think his point is that this is less "news", and more "well, duh".

The Turing test says more about the side administering the test than the side trying to pass it

Just because something can mimic text sufficiently enough to trick someone else doesn't mean it is capable of anything more than that

The Turing Test has shown its weakness.

Yeah, good point. Was hoping to avoid downloading another random app, but at this stage, I guess It's something I should look into.