Bill Gates famously said something along the lines of "Measuring programming progress by lines of code is like measuring aircraft building progress by weight". Now we're measuring AI effectiveness by the number of tokens generated? Plus ça change...
And yeah, what is it with C level executives and AI? I'm pretty much getting it forced on me by the higher ups. Management has insisted all employees come up with ideas on how to "leverage AI most effectively for the business". And now we're hiring an AI expert... For the small internal IT team of a publishing company...
At one point, I spent so much time fixing the disastrous results of people using ChatGPT to generate oodles of CSS that they would put in the WordPress Customizer that I had to resort to disabling the custom CSS functionality entirely. All because they wanted a link to be a different colour.
My assumption is that they are extremely excited about AI, because they are also extremely excited about being able to reduce their workforce while expecting more output from smaller teams.
I think that most of them are fundamentally bullshitters. Not all, but bullshitting while not knowing is what allows you to raise money and look confident.
LLM is bullshitter too, they assume everybody else does the same, so LLM does everything they think everyone does.
I’m seeing big wins with AI, but not at the product level. I’m excited about the progress, though maybe I should be questioning our priorities more.
There is so much hype around AI and using it to move faster that the extreme focus on ‘product’ and prioritizing ‘impact’ has greatly relaxed.
For the first time in my entire career, I’m able to prioritize addressing technical debt and friction in our SDLC processes.
Normally, these types of changes get bundled in with regular product work and cause timelines to grow. Nobody paid down technical debt unless it came due and was no longer optional.
I am using AI to address tech debt, but these are all things I identified as problems - and big improvements - long before AI was a regular tool in the toolbox. I just could never demonstrate how addressing them early would pay off in the long run through faster project turnaround. Fixing known issues simply doesn’t rank against new features when it comes to prioritization.
Anyway, I’m thankful to finally get to fix these things, and our ticket-to-deployment time is going way down - but none of this matters if ‘product’ doesn’t streamline their goals and priorities as well.
Similar here, the issue that I'm having a hard time getting across to leadership is: internal users and customers do not want so much change thrown at them. Totally get the 2 person startup thing trying to build something greenfield first time, big hurry, etc. That does not translate to what companies with existing customers and not scrambling for product market fit need though.
Agents are conniving yes men, and who usually loves yes men?
Leadership!
Most developers are rational and skeptical, hence the backlash from most in the community.
My boss is injecting hopium. He barely sleeps compiling massive massive documents which don't answer anything. We literally have a "(LILR Design Language)" spec which when you keep slicing through it, doesn't actually say anything.
It takes 70 seconds to full scroll through it with the fast scroll mechanism, but the whole thing boils down to "basically if it looks good then its good".
And even worse, he is asking for access to the codebase of the product since "I can do 90% of what we need and all the devs need to do is review the PR".
I say give him access and make some popcorn. At my last job I watched a manager who had not written code in years start pumping out garbage PRs and reports with hallucinated stats. He would bring them to our engineering team meetings to present and we'd spend the whole time ripping into them until he got embarrassed enough to stop doing that. I can only imagine what kind of doubling down a more stubborn person could get up to.
We just got sent a document that amounted to "please set up a CNAME for us", but was multiple pages long and had detailed instructions on how to do various troubleshooting tasks before, during, and after creating the CNAME, mixed in so that it was impossible to tell what the actual request was.
Had all the classic LLM signs. Underuse of commas. No longer sentences. No other punctuation except full stops and em dashes. Just sudden negation at the end of three barely related concepts.
I'm so unbelievably sick of reading this slop. For the love of God, it's more work to turn bullet points into an unreadable multipage document, so just send me the bullet points. I don't want to communicate with a gigantic vector representing the average person's literacy anymore.
>We just got sent a document that amounted to "please set up a CNAME for us", but was multiple pages long and had detailed instructions on how to do various troubleshooting tasks before, during, and after creating the CNAME, mixed in so that it was impossible to tell what the actual request was.
I noticed shit like that pre-AI, but at least then it was written by a person and conceivably useful info, even if redundant and not always necessary. I can't imagine how bad it's going to get with all the AI slop now.
In highly bureaucratic companies there starts to exist a value for bureaucracy ‘per se’. This can be found in endlessly growing risk and operational policies. Actually, one is lucky if the policies don’t say much. That means that the bureaucracy is trying to stay out of the way of operations. So while bureaucracy is a way of management and staff departments to fulfill their needs, the fish are still being caught.
One step up on the ladder is that the fish aren’t being caught anymore due to regulations pertaining the hooks (disallowed) and the length of the line. The fishermen complaining are being told that the regulators are also complaining and that management has decided that complaints everywhere are the best balance.
Especially because it clearly showcases Garry Tan's (YC) grandeur delusions. Not only has he gone full state surveillance bullshit with Flock but also understands absolutely zero that he's vibe coding for 16 hours a day. And, shocker: it's pure slop!
I dealt with that recently. Despite getting 6 figures off them fixing their vibe-coded mess and trying to implement sane engineering practices they still keep vibe-coding and boasting about the virtues of AI. Oh well, it pays my bills.
> But the conversation in the AI community is ignorantly staying at the level of “lol CEOs are dumb” rather than grappling with a very clear structural problem: the tools themselves are incentivized to make you feel good, the platforms built on those tools are incentivized to sell you scale, and the culture around both punishes skepticism.
This. So much this.
It's so hard to sell an agentic platform that's designed to push real work compared to one that is designed to look flashy. (Spoiler alert: the one that outputs real work is way more useful, way more difficult to build, and less flashy because they can't just sell pure fantasies like the idea of automated CEO/managers.)
It's not just the CEOs that have the psychosis problem, it's investors as well. And for even technical users it's hard to tell the difference, but comparatively it's much easier to sell AIs on actual merit to very technical people than to non-technical people.
I was under the impression that the tokenmaxxing phenomenon was mostly propaganda from the hyperscalers and their vendors to keep the AI funding gravy train alive and prevent demand collapse from popping the bubble.
The big catastrope for valuations is if supply outruns demand. Videogen has not reached the expected traction and profitability, so now we're talking about 1000x code output as a success metric. But that's not quite the same thing as solving 1000x the number of user problems.
If anything these huge codebases are just creating new problems and atrophying from context rot due to the sheer amount of noise.
> I was under the impression that the tokenmaxxing phenomenon was mostly propaganda from the hyperscalers and their vendors to keep the AI funding gravy train alive and prevent demand collapse from popping the bubble.
It's not. A lot of companies who wouldn't otherwise be terribly exposed are getting in on it. The steelman argument for it is - if you think that the optimal software development strategy is going to be much different in 2027 than it was in 2025, and you have a lot of money to burn, it may well be worth having your devs go wild trying everything to see what works.
(Of course, that steelman is very compatible with a story where the best strategies don't end up using very many tokens and a bubble pop happens anyway.)
>Around the same time, Andrej Karpathy (OpenAI cofounder, former Tesla AI lead) told the No Priors podcast he was in a “state of psychosis” over AI agents. He said he hadn’t written a line of code since December. He described tasks that used to take a weekend now finishing in 30 minutes with zero human intervention.
Karpathy is a literal genius and one of the most technically accomplished people in the industry. He built a WhatsApp bot called “Dobby the House Elf” to control his home systems (though that naming leans more towards genius than psychosis).
Ah yes, the same guy that said implementing lidar with cameras is hard (like Kalman filters aren't a thing). Same guy who spoke positively about Musks engineering talents AFTER he went crazy. That genius...
Basically, I feel like if you are suffering from psychosis, your talent is measured by how much stuff you have memorized, and how much of it you can type on keyboard in a given timeframe. And now that LLMs are doing it for you, you feel worthless.
I remember when I first started learning python, having been in Java/C++ land. It felt like a hack. You could just pip install stuff, import it, dynamically hack things around if you needed to, and make stuff work in much shorter time. I wrote tools that let me write other tools quicker. For example, back before you could ask LLMs to write code, you basically had to google stuff and search for examples. So one of the first things I wrote was essentially web page to api converter. Now I had a tool that programmatically let me pull content from web, which included things like code samples.
I then wrote a tool to search documentation and github, and pull things that were styled as code, using my previous tool, and put them into opensearch, so when i had a question about something, I could search a function in opensearch and see examples.
E.t.c and so on.
Agents these days have replaced a lot of the manual work. But complex tasks, with decision making, repeat loops, and unknown unknowns is still something that agents cant reliably do. Anyone can put together a UI with agents very quickly. But then, if you leave a lot of stuff to the agents and not specify how you want the code written, you are going to get bounded into code that is going to quickly degrade performance, introduce edge case bugs, and so on. Sure, you can have llms fix all that, but to do that automatically is something nobody has done yet.
The real skill in the future is going to be writing agentic programs to work on features for you instead of working on features. You invest time up front to do this, and spend minimal time maintaining. Much in the same way that you invested time into writing OOP code with clean separation in packages and classes, build systems with verification, all so that anyone can come in and write code and have a safe way of testing and committing changes.
> Tan’s website made 169 server requests
> (Hacker News makes 7). It shipped 28 test files
> to production users. It loaded 78 JavaScript
> controllers
> Uncompressed 2MB PNGs that could’ve been 300KB.
> An empty 0-byte file sitting in production.
> A rich-text editor loaded on a read-only page.
I mean - none of this is great, but if these are the very worst examples they can find then it feels a bit like scraping the barrel.
Deploying tests, 4MB of images, a 6MB homepage for a news site, a barrage of unnecessary assets and broken code is pretty bad for frontend. Hard to do any worse.
While not the end of the world - we routinely deal with similar crap produced by humans - it's a clear marker of the kind of quality you get from AI without real supervision.
Bill Gates famously said something along the lines of "Measuring programming progress by lines of code is like measuring aircraft building progress by weight". Now we're measuring AI effectiveness by the number of tokens generated? Plus ça change...
And yeah, what is it with C level executives and AI? I'm pretty much getting it forced on me by the higher ups. Management has insisted all employees come up with ideas on how to "leverage AI most effectively for the business". And now we're hiring an AI expert... For the small internal IT team of a publishing company...
At one point, I spent so much time fixing the disastrous results of people using ChatGPT to generate oodles of CSS that they would put in the WordPress Customizer that I had to resort to disabling the custom CSS functionality entirely. All because they wanted a link to be a different colour.
> what is it with C level executives and AI?
My assumption is that they are extremely excited about AI, because they are also extremely excited about being able to reduce their workforce while expecting more output from smaller teams.
I think that most of them are fundamentally bullshitters. Not all, but bullshitting while not knowing is what allows you to raise money and look confident.
LLM is bullshitter too, they assume everybody else does the same, so LLM does everything they think everyone does.
I’m seeing big wins with AI, but not at the product level. I’m excited about the progress, though maybe I should be questioning our priorities more.
There is so much hype around AI and using it to move faster that the extreme focus on ‘product’ and prioritizing ‘impact’ has greatly relaxed. For the first time in my entire career, I’m able to prioritize addressing technical debt and friction in our SDLC processes.
Normally, these types of changes get bundled in with regular product work and cause timelines to grow. Nobody paid down technical debt unless it came due and was no longer optional.
I am using AI to address tech debt, but these are all things I identified as problems - and big improvements - long before AI was a regular tool in the toolbox. I just could never demonstrate how addressing them early would pay off in the long run through faster project turnaround. Fixing known issues simply doesn’t rank against new features when it comes to prioritization.
Anyway, I’m thankful to finally get to fix these things, and our ticket-to-deployment time is going way down - but none of this matters if ‘product’ doesn’t streamline their goals and priorities as well.
Similar here, the issue that I'm having a hard time getting across to leadership is: internal users and customers do not want so much change thrown at them. Totally get the 2 person startup thing trying to build something greenfield first time, big hurry, etc. That does not translate to what companies with existing customers and not scrambling for product market fit need though.
Agents are conniving yes men, and who usually loves yes men? Leadership! Most developers are rational and skeptical, hence the backlash from most in the community.
Is there a term for something similar?
My boss is injecting hopium. He barely sleeps compiling massive massive documents which don't answer anything. We literally have a "(LILR Design Language)" spec which when you keep slicing through it, doesn't actually say anything.
It takes 70 seconds to full scroll through it with the fast scroll mechanism, but the whole thing boils down to "basically if it looks good then its good".
And even worse, he is asking for access to the codebase of the product since "I can do 90% of what we need and all the devs need to do is review the PR".
What is this called?
I say give him access and make some popcorn. At my last job I watched a manager who had not written code in years start pumping out garbage PRs and reports with hallucinated stats. He would bring them to our engineering team meetings to present and we'd spend the whole time ripping into them until he got embarrassed enough to stop doing that. I can only imagine what kind of doubling down a more stubborn person could get up to.
It's called: the CEO isn't staying in their lane and is injecting incompetence into the company - look for a new job.
We just got sent a document that amounted to "please set up a CNAME for us", but was multiple pages long and had detailed instructions on how to do various troubleshooting tasks before, during, and after creating the CNAME, mixed in so that it was impossible to tell what the actual request was.
Had all the classic LLM signs. Underuse of commas. No longer sentences. No other punctuation except full stops and em dashes. Just sudden negation at the end of three barely related concepts.
I'm so unbelievably sick of reading this slop. For the love of God, it's more work to turn bullet points into an unreadable multipage document, so just send me the bullet points. I don't want to communicate with a gigantic vector representing the average person's literacy anymore.
>We just got sent a document that amounted to "please set up a CNAME for us", but was multiple pages long and had detailed instructions on how to do various troubleshooting tasks before, during, and after creating the CNAME, mixed in so that it was impossible to tell what the actual request was.
I noticed shit like that pre-AI, but at least then it was written by a person and conceivably useful info, even if redundant and not always necessary. I can't imagine how bad it's going to get with all the AI slop now.
In highly bureaucratic companies there starts to exist a value for bureaucracy ‘per se’. This can be found in endlessly growing risk and operational policies. Actually, one is lucky if the policies don’t say much. That means that the bureaucracy is trying to stay out of the way of operations. So while bureaucracy is a way of management and staff departments to fulfill their needs, the fish are still being caught.
One step up on the ladder is that the fish aren’t being caught anymore due to regulations pertaining the hooks (disallowed) and the length of the line. The fishermen complaining are being told that the regulators are also complaining and that management has decided that complaints everywhere are the best balance.
I call it executive deterministic parroting:
https://ianreppel.org/executive-deterministic-parrots/
> What is this called?
A manic episode.
Stimulants
We gave our bosses access and they mostly gave up after discovering how hard it was to get things past CI. YMMV if you have fewer or healthier tests.
I work in a more traditional manufacturing business and we are poorly inplementingal AI tools allover just because our customers ask about it...
This might be the best article I’ve read in months. Thanks for sharing it!
Especially because it clearly showcases Garry Tan's (YC) grandeur delusions. Not only has he gone full state surveillance bullshit with Flock but also understands absolutely zero that he's vibe coding for 16 hours a day. And, shocker: it's pure slop!
thanks, glad you liked it! :)
I would genuinely like to see a version for freelancers: “Your Client May Be Suffering from AI Psychosis.”
I dealt with that recently. Despite getting 6 figures off them fixing their vibe-coded mess and trying to implement sane engineering practices they still keep vibe-coding and boasting about the virtues of AI. Oh well, it pays my bills.
> But the conversation in the AI community is ignorantly staying at the level of “lol CEOs are dumb” rather than grappling with a very clear structural problem: the tools themselves are incentivized to make you feel good, the platforms built on those tools are incentivized to sell you scale, and the culture around both punishes skepticism.
This. So much this.
It's so hard to sell an agentic platform that's designed to push real work compared to one that is designed to look flashy. (Spoiler alert: the one that outputs real work is way more useful, way more difficult to build, and less flashy because they can't just sell pure fantasies like the idea of automated CEO/managers.)
It's not just the CEOs that have the psychosis problem, it's investors as well. And for even technical users it's hard to tell the difference, but comparatively it's much easier to sell AIs on actual merit to very technical people than to non-technical people.
I was under the impression that the tokenmaxxing phenomenon was mostly propaganda from the hyperscalers and their vendors to keep the AI funding gravy train alive and prevent demand collapse from popping the bubble.
The big catastrope for valuations is if supply outruns demand. Videogen has not reached the expected traction and profitability, so now we're talking about 1000x code output as a success metric. But that's not quite the same thing as solving 1000x the number of user problems.
If anything these huge codebases are just creating new problems and atrophying from context rot due to the sheer amount of noise.
> I was under the impression that the tokenmaxxing phenomenon was mostly propaganda from the hyperscalers and their vendors to keep the AI funding gravy train alive and prevent demand collapse from popping the bubble.
It's not. A lot of companies who wouldn't otherwise be terribly exposed are getting in on it. The steelman argument for it is - if you think that the optimal software development strategy is going to be much different in 2027 than it was in 2025, and you have a lot of money to burn, it may well be worth having your devs go wild trying everything to see what works.
(Of course, that steelman is very compatible with a story where the best strategies don't end up using very many tokens and a bubble pop happens anyway.)
Okay but didn't we know Tan had issues before the LLM era?
When that happens, he is not suffering. He is happy. Everybody else is suffering.
>Around the same time, Andrej Karpathy (OpenAI cofounder, former Tesla AI lead) told the No Priors podcast he was in a “state of psychosis” over AI agents. He said he hadn’t written a line of code since December. He described tasks that used to take a weekend now finishing in 30 minutes with zero human intervention. Karpathy is a literal genius and one of the most technically accomplished people in the industry. He built a WhatsApp bot called “Dobby the House Elf” to control his home systems (though that naming leans more towards genius than psychosis).
Ah yes, the same guy that said implementing lidar with cameras is hard (like Kalman filters aren't a thing). Same guy who spoke positively about Musks engineering talents AFTER he went crazy. That genius...
Basically, I feel like if you are suffering from psychosis, your talent is measured by how much stuff you have memorized, and how much of it you can type on keyboard in a given timeframe. And now that LLMs are doing it for you, you feel worthless.
I remember when I first started learning python, having been in Java/C++ land. It felt like a hack. You could just pip install stuff, import it, dynamically hack things around if you needed to, and make stuff work in much shorter time. I wrote tools that let me write other tools quicker. For example, back before you could ask LLMs to write code, you basically had to google stuff and search for examples. So one of the first things I wrote was essentially web page to api converter. Now I had a tool that programmatically let me pull content from web, which included things like code samples.
I then wrote a tool to search documentation and github, and pull things that were styled as code, using my previous tool, and put them into opensearch, so when i had a question about something, I could search a function in opensearch and see examples.
E.t.c and so on.
Agents these days have replaced a lot of the manual work. But complex tasks, with decision making, repeat loops, and unknown unknowns is still something that agents cant reliably do. Anyone can put together a UI with agents very quickly. But then, if you leave a lot of stuff to the agents and not specify how you want the code written, you are going to get bounded into code that is going to quickly degrade performance, introduce edge case bugs, and so on. Sure, you can have llms fix all that, but to do that automatically is something nobody has done yet.
The real skill in the future is going to be writing agentic programs to work on features for you instead of working on features. You invest time up front to do this, and spend minimal time maintaining. Much in the same way that you invested time into writing OOP code with clean separation in packages and classes, build systems with verification, all so that anyone can come in and write code and have a safe way of testing and committing changes.
> Tan’s website made 169 server requests > (Hacker News makes 7). It shipped 28 test files > to production users. It loaded 78 JavaScript > controllers > Uncompressed 2MB PNGs that could’ve been 300KB. > An empty 0-byte file sitting in production. > A rich-text editor loaded on a read-only page.
I mean - none of this is great, but if these are the very worst examples they can find then it feels a bit like scraping the barrel.
Deploying tests, 4MB of images, a 6MB homepage for a news site, a barrage of unnecessary assets and broken code is pretty bad for frontend. Hard to do any worse.
While not the end of the world - we routinely deal with similar crap produced by humans - it's a clear marker of the kind of quality you get from AI without real supervision.
If that’s the front end, imagine how the back end must look.
Rapid iteration of a bad code base, what could possibly go wrong?