whstl
People are producing output that is worse than "doing nothing".
Ego and agendas multiplied by 1000.
Code was never the bottleneck.
I've worked with plenty of teams that did the "in prod the same day" trick without AI.
The problem was just fucking ego. Designer has an ego, product owner has an ego, QA has an ego, stakeholder has an ego, platform teams have massive egos, developers have egos.
Unrelated but: the only reason people are shipping more with AI is because more people are ok with slop now.
I guess it was slow-boiling, but somewhere between 2010-2015 was the turning point.
It got dominated by brogrammers who judged you by whether you used vim or not, teams of 20 people producing the same output as a team of two, designers doing mockups for absolutely everything and trying to boss the programmers, product managers making mountains of useless documents for everything and trying to boss the programmers, some rando had Scrummaster certification (who doesn't) and wanted to boss the programmers.
It was also when everyone started copying Google and Spotify and etc and a meeting was required for every decision, RFCs or other documents that would be produced but wouldn't survive reality, the software would have more trackers than features, and subscription was the default software model.
People keep saying "you only need 1 or 2 people to make a product now".
Well that was already the reality for me from 2000-2016.
Although I'm sure someone is bolder than me is running a fleet of Mac Studios somewhere!
"One cool trick" I like for running those kinds of models very quickly on macOS is converting them to CoreML so they run in the Neural Engine.
I had amazing success with Spleeter, Demucs, MVSEP-MDX23-music-separation, DeepFilterNet, Whisper, Hunyuan3D for 3D mesh generation. Plus a lot of others for both work and leisure.
I also convert the code to Swift without dependencies, but that's a personal preference.
The runtime on some of them went sometimes from 60s to 10s. And since it's just Swift, I can just give them to non-technical co-workers and they don't have to setup a whole Python environment.
If you don't wanna do this by hand, this is something that existing LLM models are good at. Even Luna can do it without much fuss.
Even in the 90s, the internet and the web dramatically lowered the costs of distribution, and YC itself was built on top of that (ViaWeb).
Then there has been so many frameworks, libraries, tools, services and APIs to pick and choose from that most of the coding work has been about organization, selling an idea and doing a lot of testing. Which are still unsolved problems.
The only new idea here is the idea that Software Developers are disposable.
This was brought by the same people who did RTO and layoffs post-covid, but now with support of non-tech workers who want a piece of that (supposedly) sweet money.
Seriously now, Marx says that industrialization and other developments were essential for having a socialist society able to sustain itself. Russia, China and Cuba did it way too soon if you go by the book.
So yeah, there's that.
But I feel like it's a minority that really know how the software they're writing really works. And it's 10x worse for older legacy software.
Mostly because the people doing it don't care. Not judging, just reporting I guess. I know because I used to work in quite a few companies where the "money maker" app was on maintenance for years and nobody was able to answer questions or didn't even know most corners of it still existed.
This desperation mindset also often correlates with people writing code without care for the next maintainer.
And the scary part is that this also correlates with advocacy for going all-AI.
Even the article itself: they care about "having a job", not having that specific job. And they care about it in order to "make a living".
The solution in those languages is either to use something like serde validator, garde (for Rust), or go all the way to protobufs... or write your own data validation/parsing by hand.
Rust and C++ don't even have a Javascript deserializer.
The guy who has some power about who to fire, but still flips burgers or mans the cash register when everyone else is busy.
Not saying it's not a noble job, but it's very far from executive.
But we can still already replace 99% of C-levels with a $20 subscription to Claude and a few MCP integrations.
Anything that needs managers to actually go and check.
It's not that they trust the employee to work with zero supervision, it's just that most managers and execs don't want to do any kind of work.
The webcam checks are pure projection.
I saw a situation just like the one you mention recently: guy burned out, broken inside, working way too much, but not being noticed at all.
Execs were basically claiming he didn't do a certain long-term task and that "nothing was delivered".
The problem is, there was a public Slack group for visibility with evidence of the task, artifacts of the task, plus relentless requests by a certain person that he deletes the data in there. This person is the same one who went around claiming he didn't deliver anything, a narrative that was bought by execs.
After presenting the evidence and being shown the artifacts in a meeting, no exec accepted reality, no exec walked back, and just shrugged off with "it's too hard for you to understand".
Half of this business is not really doing much work, and a big chunk of those are psychopaths that depend on throwing people under the bus to survive.
Humans, especially ones under abusive managers, often have a lot of responsibility but no authority that affects the outcome.
In an ideal situation you either have both, or none. You make the decision and answer for it, or you follow decisions but don't get blamed when it's wrong.
LLMs managed to slide into a space where they get to make authoritative decisions but if something fails we don't really blame them.
I'm not saying LLMs should be held responsible, but it's funny how people claim they replace humans, but are giving them an extremely easy, kid-gloves, success target.
It's "free speech", not "full agreement".
People run a stupid amount of expensive queries that end up costing way too much because they're asking Claude the wrong query.
Not to mention people running wrong queries, using the result as gospel, and then the result has to be sent to a data analyst to be reverse-engineered so the numbers make sense.
I have seen the Cursor leaderboard on my company and the vibe coders consume about 5x more tokens than the developers. They and other office workers also have Claude and their limits are often over around Wednesday.
People are using millions of tokens to do very simple HTML reports. I have seen someone asking the LLM to download the entire data into the context and asking it to sort.
Those usage patterns don't correlate to output.