[torben-friis]

>But all now seems like a transition phase. Transition to f-ing what though?

It feels like being in the middle of a tornado. But I think it helps to turn off screens, sit in a desk, and calmly remember first principles and consider them slowly.

Quoting obama, "reality has a way of catching up with you".

I see a lot of talk, but iOS is not delivering a decade of features and fixes on each yearly release. Literally no one does, if anything people are complaining that existing functionality is breaking down. So it can't be true that we're at 10x productivity, and this fact will eventually catch up with us.

Let's be human, and remember that many people are emotionally invested. Juniors want this to be a chance to shine in a market that otherwise rejected them. CEOs placed their bet on AI and don't want to walk that back. Seniors want to signal that they are not obsolete. AI companies will poison discourse. But all this smoke will eventually clear.

[Folcon]

I've been thinking a fair bit about what I'm seeing in terms of the output I experience

It's quite hard to quantify, but I think it's one shot nature really makes it hard to gauge it's capability

Friends have spoken of good days and bad coding days with me, and I find it odd nodding along, it's a strange new normal

At times it feels like we're just coding with one-armed bandits, trying to carefully line them up for a jackpot and just discarding and retrying if we don't hit

I think about some of the more complex systems I've built and I wonder how well we can build them like this

And over engineering, there seems to be over engineering everywhere, and yet, more fragility to our systems

It's all a little surreal

[customguy]

I imagine this stuff is probably really good at iterative changes to improve objective benchmarks like CPU or RAM use. I'm thinking of little contained optimizations that you can understand in and of themselves, that you maybe have done yourself before in other places, found really quickly and applied uniformly. Stuff you can confirm to be what you expect it to be by mostly just scanning. Low hanging fruit for sure, but something where you can actually know what is a fruit and what is crap, and only keep the fruit, and develop some confidence in the process, if that makes sense?

Or trying to reduce complexity, increasing readability and coherence of variable names (the opposite of code golf if you will), while staying within a certain limit of performance regression (e.g. "make this code as nice as you can while making it at most 0.5% slower").

Making the stuff millions of people have been using for decades better, in a way that also makes it better for humans when they read the code. Surely that's possible, some people are probably doing it but it doesn't go viral as much, because it's too mundane.

And of course, making new stuff is more exciting. I mean, you could hit on something with a vibe coded thing, and then know it's now worth to make a non-sloppy version, but you won't get much fame for making ffmpeg twice as fast by prompting an LLM. Though on the other hand, it's like a safe investment (if not in "fame", then in "improving the stuff we all have to use daily"), because you know ffmpeg and many many other things will still be around, whereas a vibe coded thing that wasn't special will be 100% forgotten the next day, or have just the one user forever.

[patates]

> Juniors want this to be a chance to shine in a market that otherwise rejected them.

I actually am training 2 trainees (Azubi in German) and 1 working student. All three are somewhat anxious about the future but also all are learning in a significantly increased pace, compared to the ones I worked with 1.5 years ago.

They don't have to wait for random senior to answer questions, so they get stuck way less often. They aren't allowed to use AI to generate code though, so not sure how that'd look like learning-wise if we/they went all-in on AI.

[leononame]

Is that a good thing? I think getting stuck is an intrinsic part of the learning process and sometimes it's good that there isn't an immediate answer from a senior. For some things you'd never have solved yourself, sure. But pain and suffering is a big and important chunk of learning and I fear we just throw it all out of the window with asking AI.

[runarberg]

I am skeptical. This is a testable hypothesis, and I know there have been tons of research done trying to provide evidence of claims like these. However I have yet to see a convincing evidence. If this statement was indeed true, I think experimental data backing it up would be old news at this point, it however is not, and I suspect that is because junior developers are by and large in fact not learning faster or more with the help of AI.

[kibwen]

> But all this smoke will eventually clear.

I wish I could be so optimistic. Our lives are ruled by distorted, irrational, inefficient, failed markets, and the markets can remain distorted, irrational, inefficient, and failed for longer than we as individuals can remain solvent. "In the long term the market is a weighing machine", for term lengths that include the heat death of the universe.

[sph]

> It feels like being in the middle of a tornado. But I think it helps to turn off screens, sit in a desk, and calmly remember first principles and consider them slowly.

It’s been enormously alienating talking to laypeople about the apocalyptic atmosphere in the tech world, while for them ChatGPT is a cool piece of software but the world still hasn’t changed very much. They look at me bug eyed when I tell them how dramatically software engineering has changed in the matter of a couple years. We are in a bubble, or we’re just, as you say, in the eye of the cyclone which still has to hit the rest of the world.

[abustamam]

> iOS is not delivering a decade of features and fixes on each yearly release

This is insinuating that code was the bottleneck in the first place, or that every line of code is to build a new feature and not fixing existing bugs, or that apple didn't lay off enough engineers and reallocate resources to other departments to make up for the productivity boost.

I do think that companies with poor AI practices will eventually pay the piper in the form of technical debt or debilitating bugs. But let's not equate a productivity boost with a boost in releasing features, because there's plenty of business reasons to not release thousands of new features every year.

I agree with you on the rest of your points. Eventually the smoke will clear. What awaits to be seen is who is left standing when it does? I don't think I like the answer to that question.

[thedevilslawyer]

I sense your comment as saying: "AI is hype, and reality will catch-up.".

But the simple fact is there's massive evidence that in skilled hands 10x or 100x engineers are possible. We're seeing evidence of it across major open source project as well. And definitely behind closed doors across companies.

Reality will catch-up with that too, once the other smoke clears.

[shinryuu]

100x is achieving something in two days, what it took an entire year before. I strongly doubt that is happening for an individual.

[mistersquid]

> 100x is achieving something in two days, what it took an entire year before.

Reduce your scale: "100x achieves in 1 hour what used to take 1 week."

One year of work could require levels of complexity and human judgement that can't be accelerated past a certain point.

1 week of work can be reduced to an hour and some change.

[shinryuu]

That requires that you not only one-shot the thing. You will also not have the time to verify the solution yourself in that time period.

Besides 1h what used to to take 1 week is basically 40x given a workweek is 40h.

[freeopinion]

Ignoring the obvious math problem here, I think the broader scope is important. Some of the point from preceding comments is that overall longterm output is not changing. So if somebody is oneshotting a week of work in an hour, but has the same annual output... where are you losing all the normal productivity that should have happened the rest of the week/year?

If one week of work can be reduced to an hour, then you should be able to complete a year's worth of work in 50 hours. If you break that into two 25-hour weeks (because a 40x dev earns the right to loaf?), what is that dev doing for the other 50 weeks in the year? What is making them so incredibly unproductive 50 out of 52 weeks in a year?

[fauigerzigerk]

>One year of work could require levels of complexity and human judgement that can't be accelerated past a certain point.

In other words, if you include everything that's required to create useful software then 100x turns out to be a fantasy.

[ruszki]

That’s also not possible in “skilled” hands. My output is roughly the same. It does the scaffolding, but I need to rewrite almost every line, because it introduces footguns around that often. And before I had about 20000 LOC it failed even with scaffolding, ie architecture. And it wasn’t taste, just footguns all around, architecture ones. Nowadays for example still introduce mutability or completely unnecessary complexity where it shouldn’t, even when the example code which does almost the same is pristine. Many times it’s like StackOverflow, when a question doesn’t need 90% of the accepted answer, but people happily copy it brainlessly.

This is especially bad with new, or quickly improving frameworks, like Android Compose. LLMs use completely outdated, deprecated APIs all the time, when they are not completely supervised. Or at least, I hope so that the framework causes it. Because if that’s not the case, then your products are fucked.

Also even with the best prompts it could never produce more working code in an hour than what I can produce in a day. Regardless of quality, just “working somehow”. Not even with an uninterrupted session. If that’s the case for some, then there is definitely also a developer skill issue. And so would definitely not trust anything coming out of their “supervision” of an LLM.

[Kudos]

The baseline engineer in that case must really be something incompetent.

[kibwen]

Ironic. The best way to turn 1x engineers into 100x engineers is to reduce the median skill level by 100x.

[vanviegen]

> But the simple fact is there's massive evidence that in skilled hands 10x or 100x engineers are possible. We're seeing evidence of it across major open source project as well. And definitely behind closed doors across companies.

Each of these three sentences are in need of some evidence. I'm not actually seing any signs of software velocity notably increasing anywhere. Except perhaps in the AI-reseller sphere, but that seems mostly due to throwing huge amounts of VC money at it and a lack of quality control.

[koonsolo]

Name a single project that has a 10x increase. I mean real production ready code, not some single person hobby project.

[f17428d27584]

AI is turning 1x engineers into 0.3x engineers claiming to be 100x engineers.

[realusername]

> And definitely behind closed doors across companies.

At my large org (+100 engineers), I'd say it's a mixed bag and the overall impact of AI rollout looks to be slightly negative productivity.

They probably won't say it publicly though.

It's not because some people are more productive with it that all of them are and it certainly doesn't mean that the company itself is more productive either as you have other things than code to take into account.

[pianopatrick]

I can believe that the difference between the slowest programmer in the world and the fastest AI aided programmer in the world is now 100x in terms of lines of code output. Like I can imagine a programmer writing 250 lines of code per week by hand and I can imagine an AI powered person writing 25,000 lines of code per week.

[nunez]

But we spent decades as an industry trying to dispel of the notion that SLOC/KLOC does not matter.

I still believe that. 250 lines of tight code that solves a specific problem in a way that others can maintain will always be better than 25k lines of code that's difficult to review and consume (and, therefore, becoming a liability).

[torben-friis]

>and I can imagine an AI powered person writing 25,000 lines of code per week.

A year of that is 1.3M code: the size of systemd, or postgres.

Can you imagine a single person writing systemd (not a POC, the current version feature complete and battle tested) from scratch in a year? If so can you point me to any such project?

[pianopatrick]

Well, as a thought experiment, I'd say that depends on if the measure is "1.3M lines of production code" or "1.3M lines of code (test code, comments and production code combined)".

I don't think you could get to 1.3M lines of production code, but people say AI agents are good at writing tests. I could imagine that if you had an unlimited budget you could set up agents to generate lots and lots of tests and comments, inflating your weekly total. Like, maybe you can set up an agent to loop against a code coverage tool trying to generate more and more tests to hit MC / DC levels of testing.

In the extreme / absurd case, if you could hit the SQLite ratio of 590x lines of test code vs real code then 25,000 lines of code per week could be 43 lines of production code and 24,958 lines of test code.

You'd be a "100x programmer" in terms of lines of code output, but that would not get you to SystemD levels of production code in a year.

I can't point you to a project taking that route, and I don't have the budget to try it, but I can imagine someone hitting 25k lines of code per week with lots of tests and comments padding the numbers.

I am not sure software written that way would be any good though.

[sph]

I have seen people with that kind of AI assisted productivity-by-volume, building eldritch abominations like Gas Town.

[afdbcreid]

10x? The most generous studies I know give up to a 50% improvement.

[torben-friis]

I'm not say it's pure hype. Not in the sense of previous hype waves like the metaverse or NFCs. It clearly has uses.

I do think it is hype as a killer of knowledge work. It can certainly remove a lot of friction in the kind of borderline mechanical work that you'd formerly outsource to the lowest priced denominator, serve as an idea bouncer, remove friction for bug tracing, etc.

Attempts to cross the next line ("no need for architecture discussions, ai plans", "no need to read the code, ai reviews", and so on), nope.

As someone else mentioned, 100x is a couple days producing the outcomes (remember, not output) of a year. Or for a team, iOS delivering in a single year ten times as many features as its entire previous existence. It's not something that doesn't get noticed.