Ask HN: Where are the AI-driven profits or promotions? I see so many comments on the huge productivity gains of coding with AI but it mostly seems to just be perceived productivity But does anyone have stories of the actual results or impact? If the productivity gains are so huge, shouldn’t we see those engineers getting promoted super fast or getting rich from building a product extremely fast? Where are those stories? |
The target for a great startup was 7% growth week-on-week, and now it's 10%. It means back then if you were making $1k/week in Jan 1, 2024, you'd expect to make $33k/week a year later. But thanks to AI, the standard is closer to $142k/week.
Back then, it was normal to raise $75k or so to build a prototype, now you can build a prototype in a couple of weekends. Back then, you would polish your pitch to get VC funding. Now you build, get the customers, then show what the customers are buying and why, and ask for funding to get 1000x the number of customers.
A lot of the risk factors have been removed. Building a legacy app no longer take multiple sync calls and weeks, they can be resolved in hours and maybe one sync call. Yes, I have a story of AI migrating a whole DB table when I asked it to move a button to another page. But it gets the tables right about 7 out of 10 times if you give it the right instructions. I dislike ORMs; they're practically a code smell. And AI works around that layer.
The story doesn’t state much about source but is it literally just an editorialised press release?
I'm leaving that utm_source in the link for irony, lol. Couldn't find it from the first page of Google, but AI found it in seconds.
Is that 53,000, i.e., a misplaced comma, or is it 5,300, i.e., an extra zero
I've seen no actual improvement in the development speed of our product. Same pace. Same endless discussions. Same meetings. It hasn't changed the pace nor it has lead to an explosion of ideas, implementations or discussions. It did lead to an increase in code and deployed microservices, though, but that is not really a good thing in itself.
It's like making a commitment to eat healthy, diversified food. And, as we did in the '90 in the post-communist country I'm living, we've let the doors wide open for all the food to enter. Junk or otherwise, without, from our part, any knowledge of what is healthy and what not. Tremendous amount of choices, little, close to zero knowledge about what is good and what is not.
Teams are ranked by AI usage now, making me ask claude code how it slept every morning.
You better brace yourself, because as such a company we will run circles around you.
Chasing a tail that seems to remain just slightly out of reach!
I think that's why we see small teams and indie hackers really utilize it. They aren't bogged down by the beauracracy of code changed in a large org.
Big tech is looking for a miracle cure without looking at the root problem (too much process, overhead, cruft). Current AI tools are also far less useful in large codebases than they are in greenfield projects.
If that’s the case the industry hasn’t considered how it will adapt to that. It’s not even in the conversation.
love this analogy, thank you !
Company offered Copilot subscriptions org-wide early, as well as Augment, and Slack integration for LLM's with internal docs and discussion contexts.
Copilot has been useful (I was part of the early access and used it +1 year prior to public launch, though) but I think Augment has really seen the biggest impact in our org.
Several small features and a lot of refactors have been done entirely through Augment.
Personally, I'm a bigger fan of Claude Code, but it's not covered by our org so I only get my $20/mo use out of it.
Honestly, the productivity gains feel like a mirage sometimes. Yes, I can prototype features in days instead of weeks now. But getting those prototypes to production quality? Still takes the same amount of time. Real-time voice processing can't have the hallucinations or edge case failures that AI-generated code often has.
The weirdest part is how expectations shifted. When I delivered a feature in 2 weeks before, that was good. Now if I use AI and deliver in 1 week, suddenly 1 week becomes the new baseline for everything. You're not getting ahead - you're just keeping up with inflated expectations.
I've seen the real AI profits go to people who can build complete products solo, not just code faster at their day job. A few engineers I know left big tech to build AI-powered micro-SaaS. They're making $50K/month instead of waiting years for senior staff promotions that might bump them $20K.
But here's the catch - if starting an AI product is 10x easier, so is your competition. We're seeing way more AI startups, but the success rate is about the same. The bottleneck was never "can you build it fast enough" - it was always "do people actually want this."
This is the productivity-pay gap, in place since 1980 as real wages have barely budged while productivity has continued to skyrocket, the value from the accruing to a small wealthy minority.
Besides causing social ills such as the concomitant de-democratization of society we are witnessing around us as a direct result of this, the productivity pay gap also undermines the fundamental fallacy of the AI hype crowd - higher productivity doesn’t translate into broader benefits for society, only for the business owning class.
Only external pressure can force business owners to part with gains from rising productivity. With the productivity-pay gap of the last 55 years wealth is so concentrated that the government today is entirely captured by business owners. Billionaires wave their hands and select government officials, unmake the constitution, and form new political parties. Can our government be relied on to pressure business owners into parting with their growing profits and raise wages? Our government has failed to do this for 55 years.
If engineers want higher wages for their increased productivity - which has multiplied since 1980 with little of that going to workers - then we will have to pressure business owners ourselves. The best way of doing that is unionizing.
All code is reviewed by senior devs. We use Rails so that probably helps get good results.
No promotions here but our clients love the results we’re delivering (they did before AI too).
Not a 10x story but we’re definitely getting a lot more done than we did before and we’re able to maintain a level of quality we just couldn’t before. TDD + AI refactoring leads to unbelievably clean codebases.
This is an entirely different question in itself, and isn't really related to AI, but more between productivity and corporations.
Corporations are a pipeline. You have a certain amount of input and a certain amount of output. There is a queue of input stuff. If that queue runs dry, then you have wasted productivity. The pipe manager is being negligent somewhere - either we can cut down on costs by removing the cogs in the machine, or someone further up the pipe isn't producing enough output to satisfy the inputs.
If FE has too much free time, then your options are: 1) pull FE into BE, 2) pull work from BE to FE, 3) get FE to do product stuff like A/B testing or talking to customers, 4) make more FE teams or even have one person in two teams, 5) turn the site into an app or the app into a site, 6) slow them down by making them return to office, 7) fire someone
A Chinese company might have a couple dudes dedicated to answering all questions on APIs and the systems. A Western company will simply not document things or overdocument it to the point that it's too obtuse to read and yet doesn't answer the question. The worst cases I've seen is developing a whole new programming language, which becomes a white elephant maintained by interns.
But the point is, there's mechanisms that kick in and slow down productivity so nobody needs to be laid off. There's also additional mechanisms like engineers being sick of working in 3 teams at once with no documentation, and so they leave.
The most likely path is to enable a million independent projects to flourish and to find unserved niches that lead to a good, but not exorbitant, income, at least for a time.
Sure there are workers that get bonuses, but your bonus for a worker or lower level manager typically is pre-determined so if you overwork and contribute 50% more productivity enhancing output, you don’t get 50% more in your bonus.
One path could be to build your own thing and become a business owner but that is much more work than just building software.
The "10x" programmer in your org is getting a pat on the back and a 2.1% raise instead of the 2% raise everyone else is getting.
The company I work for is improving their software at a much faster rate. We've solved extremely tricky problems plaguing our reliability and usability using AI. The error rates for key systems are approaching 0%.
> But does anyone have stories of the actual results or impact?
Will this work at every org? Probably not. Will it work for some? It seems reasonable to me, though we can't plot the long-tail of these sorts of efforts yet, I think.
Without it, the site would suffer a slow and painful death in the SERPs and would lead to about 10MM annual loss for the company.
Starting from scratch with a proper, qualified team was not possible for political reasons.
So, being able to do it as a single person, with heavy AI assistance, is a huge win.
Where I work there probably wouldn't even be a promotion on the line, not in the same year at least.
If a new, wonder drug or food or medicinal plant would appear on the radar with claims of "it increases your life expectancy by x years, your happiness and well-being by y amounts, etc, etc" would you say that those claims are based on fact or are just marketing?
But I’ll grant you this, having said the above, I have no way of actually objectively proving love.
I need to get back to him and ask him why he isn‘t making more money than before…
It'll be interesting to see if that becomes a trend. Just what are people supposed to do with vibe coded codebases...
I use ChatGPT every day, to make myself more efficient (I get things done more quickly and with higher quality). Nobody really knows I use it.
I dunno about you, but in almost every one of these threads the exact same two classes of people show up: the ones championing the insane multiples they get from their AI, and the people who say "it doesn't work for me" whereupon the first class of people say "well, then you're just holding it wrong"
So, no, it's not the same rate because the non-deterministic behavior means that one cannot provide a shell script that person A and person B can run on their machine and get the same outcome like in the good ol' deterministic software tooling days
1. We’ve got a smart senior dev who’s very excited about AI. His core work hasn’t changed much, but he has generated a lot of AI driven unit tests that mock so much that they’re worse than no unit tests. I approached him about this and he kinda shrugged - said it was good test coverage went up. He’s much more attracted to building new things than making things well, which I think is the source of the issue.
2. We have a low performing dev who has spit out not great code (I think a result of being unable to evaluate what code is good) at a much higher rate. The PR reviews are a massive increase in time wasted.
Both of these devs have personally stated AI has made them much more productive. It’s greatly tempered how I view AI productivity claims I see online.
Personally I use it everyday and am willing to pay for it, but I’ve seen limited uses where it’s 10x (translations, areas I’m a novice).
Also I've experienced quite a few times where a dev I respect is quite vocally pro-AI in public but in private is very neutral or negative about it
I'm a very straightforward person, so this behavior from other people I respect is really causing me a lot of difficulty
- they started to realise they didn't have some of the domain knowledge (I specialise in maps/GIS), so couldn't steer it effectively
- they said that the changes it made started to become unstable: while making one change, it would break other things. It got harder and harder to make progress.
I suspect the second problem wouldn't happen as soon (or maybe not at all) with an experienced dev running the process.
A product person can quickly validate an idea and, once the project is a bit more concrete you can bring in the engineering team and start caring about security, maintenability, scaling and so. A rewrite is almost always the best thing, you can start with a solid foundation instead of spaghetti vibe coded stuff
That doesn't make him wrong though.
But I doubt there will be proper data either way. Whatever model is used in research will be deprecated by the time something is peer reviewed. If a model hallucinates, can't do X, is boring, has so and so security holes, the report is out of date by the time it's released.
But the next best we have is that people who make these claims are throwing money at it. I would be highly suspicious if people made these claims and then didn't double down on their investments. Sundar says AI is bigger than fire which is highly sus, but the way he's been reorganizing one of the world's most stable companies to focus totally on AI suggests that he believes it.