Comment by etempleton
1 day ago
This has been my argument for the past few weeks to explain why we aren’t seeing or feeling real world productivity gains as consumers despite many espousing the efficiency of AI they personally see with their own workflows. The hands on work is rarely the actual bottleneck.
My team can crank out high quality work in a few hours and AI could do the same medium to high quality work in a minute or two, but the review, feedback, and sign off takes days of back and forth. AI just slows this part of the process down because editing AI work is often a challenge because, no one is quite as familiar with how it was put together and sometimes it creates things that are not editable at all. And telling AI to make very specific changes is often a mixed bag at best.
Solo devs and small teams with minimal oversight can probably move a lot faster, but corporations are going to see minimal uplift except in areas where no one is looking.
Established successful companies operate in a world where action and change often introduce more immediate risk than staying still. In those companies AI doesn't necessarily offer a lot of "increased productivity" gains because writing the code isn't the bottleneck. But many of them will try to chase cost savings by laying people off or slowing hiring.
There's a couple types of companies where the reverse is true: young companies trying to grow, and dying ones that need a turnaround.
Out of those two, young companies are less likely to be set in their ways. So the first place you really see the result of development moving faster with AI is probably going to be in future months and years when suddenly-big companies start getting widely known with products and business strategies that wouldn't have made sense a decade ago because writing code to try different things to even get the ball rolling was a bottleneck, and you could only do things that wouldn't scale for so long.
Or the productivity gains, albeit substantial, are not as earth shattering as the llm providers would have you believe. As with any new technology adoption is a laggard. The internet "existed" in the late 80's, early 90's. Many companies are still reaping productivity gains from plugging into this first.
I work in code generation, and often feel the fallacy of “because I can build something that looks 80% like your product in a day, you should pay me N millions.”
It's roughly the same pitch that the ruby on rails crowd had 15 years ago, and the no code crowd 5 years ago.
Very few companies need to rebuild their product every week, or actually care what tech stack it runs on. They are trying to build new things, or reduce how often those things break, or figuring out what if anything is worth building next.
> Solo devs and small teams with minimal oversight can probably move a lot faster
Spot on!
Usually, most work in corps is slowed down by executive decisions. It takes 3 - 9 months to get a yes/no on something and for most things that are obvious that need to be done need to go up the chain of command. The decision takes forever, but when the decision is made, they want things to be done quickly, which creates unreasonable stress on the ICs.
The first step to see real productivity gains is for AI to eliminate the executive and middle-management layer, this starts a virtuous cycle, destroys both empire building and massively accelerates true productivity, not just meaningless metrics like LOC.
Speaking realistically, you're going in the wrong direction here. The most likely scenario in the short run is that management will seek to replace labor with automated systems supported by a skeleton crew to aid in implementation and execution level tasks that can't (or shouldn't) be handled by the LLM alone, like deployment.
So the first signs of productivity gains will be shrinking head count, but it's going go from the bottom up, as usual.
I considered it an absolute win when I submitted an LLM assisted PR that removed 30k LOC :)
> most work in corps is slowed down by executive decisions
I work in a startup and I find it to be the opposite. Management wants us to go really fast and engineering is like "well wait we should slow down because just because LLMs can write a bunch of code doesn't mean it can build a maintainable app"
That said, I know that our small engineering team can't build an entire business by ourselves. So some management is required. But as you said, getting stuff done quickly is stressful. And with LLMs we're expected to get a LOT of stuff done very quickly. Some colleagues have decided to go on quiet strike by submitting AI slop and citing "no time to review" to prove a point. I don't fault them, but I also know that me and someone else will have to clean up their messes at some point.
And yet amusingly it’s the ICs that are being replaced, not management or executives…
AI. CEOs. Now.
They way these companies are structured we're seeing them remove the people doing the actual work while nobody wants to mess with the execs. My company has something like 60 EVP titles, with SVPs then VPs then senior director then directors. Pretty much everyone VP and above could disappear and our lives would be easier. It creates a lot of politickings and empire building that holds back the company and the individuals then the individuals delivering code and products get targeted by the efficiency metrics and nothing at the top changes.
I have no idea how it’s sustainable but I dont see AI CEOs coming until a new company becomes successful from the bottom using that model.
Claude is already emailing me from the CEOs email address. All the load-bearing seams and everything.
This is my experience. As a data engineer with years in my current role I can get AI to write good SQL that correctly answers questions or defines metrics. At worst it may take a few back and forths to really nail it.
However, if a PM tries it they will more often than not get a wrong answer because they do not know what to look for or how to ask the right questions.
Prompts still matter a lot especially when you are working in an established enterprise.
What model are you using? I find Claude Sonnet to be quite capable of generating correct queries. Also, are you using a semantic layer [1], because without it the model may have trouble understanding your schema. I would be very surprised if you PM gets a wrong answer with these settings. I have never seen Sonnet 4.6+ produce a bad query.
[1] https://communitysupport.group/t/what-is-the-purpose-of-the-...
This gives me War of the Worlds vibes. "AI would have destroyed the world but Gary in Ads, as always, was a day late with his release approval."