OpenAI launched a $4B consulting company. Anthropic launched a $1.5B JV. Both sell human supervision for their own models to enterprises that can't deploy AI on their own.
I am intrigued by the bit of insight you have included into the DevOps services you use internally and also market. With having to dissect and fix vibe-coded apps, I'm curious on your general impression of the code quality and maintainability. If you don't mind a couple of questions: Do you get into refactoring overall, or is your goal just to make sure the code not brittle and insecure? And, to what extent do your workflows utilize agentic coding tools?
Thanks Jim. The goal is almost never a full refactor. It's making the thing safe to run. Exposed keys, no access control, data flowing where it shouldn't, logic that holds on the happy path and nowhere else. We harden that first.
Code quality varies wildly, and it tracks one thing: whether the person who vibe-coded it had real engineering background or none at all. Someone technical produces something brittle but fixable. Someone non-technical produces something where the fastest path is rebuilding it on infrastructure we already trust, with the vibe-coded version serving as the spec.
That's the part that scales. We don't salvage line by line, we rebuild on boilerplates I keep in tight shape. Side effect nobody warned me about: the whole company's code now looks like mine.
On the tooling: we're mostly on Claude Code day to day. Gemini plus http://pi.dev is on the list to evaluate, just haven't gotten to it yet. The model matters less than the boilerplate and the review layer around it, which is the whole point of the original piece.
Probability of error in LLMs follows a power law. You can not automate sh#t with those machines. You need EXPERT in the loop.
I saw many consultancy firms consulting
in the past. In the best case those forward deployed engineers are going to be overworked junior guys that start their career there just for their CV and abandon the firm when they are burn out before been 30 years old. In the worst case, these guys are not going to be engineers. They are going to be agent operators. In any case, salaries will go down. The final goal for which the current developers are so devotely helping to happen.
Expert in the loop, agreed. But it goes one level deeper than that. A lot of clients don't even have a process to automate. They think they do. Then you open the hood and there's no actual workflow, just a few people improvising and calling it a system.
So before anything gets automated, we end up reconstructing the process from scratch, piece by piece, until there's something real to point an agent at. The AI question comes last. Most of the work is figuring out what the company actually does, which no model can tell you.
Sometimes I remember that quote from someone, autonomous driving could work perfectly if the roads were perfect.
I saw enough enterprise codebases and pipelines to know that quality gates is like saying hyper drive engines in most of corporate development environments. Yeah, I'm speaking about huge financing institutions and firms.
So now we are adding sigma event generator machines to that.
The result is not going to be pretty, and the same way we do when garbage dumps grow exponentially, the only solution to that will be to incinerate the garbage.
The lost years we will call these days in the future. When billionaires became trillionaires.
"Works if the roads were perfect" is the cleanest version of the argument I've heard. The whole autonomy pitch assumes a clean substrate, and enterprise pipelines are the opposite. Sigma-event generators on top of code that already has no quality gates, agreed, that compounds fast.
But I read the ending differently. For people who actually have the expertise, almost nothing changed. They just got faster. The mess still needs someone who can read it, and that someone is the same person it always was. The tools moved. The job didn't.
If you like clean, here you you have a senior software engineer with 25 years of experience that only has 100 days of unemployment left in a job market for engineers that at this moment is absolutely broken. Nobody hires. Someone lays off a bunch but nobody fires. All companies and employees are like rabbits surprised by the lights of a car. Nobody knows what to do to be contracted. Juniors have it worse. They are calling them useless. Some unemployed with AI psychosis decide to use agents for making that idea of product they had real, with no bussines plan or anything, just burning tokens and finishing worse than when they started.
And I listen same and same message again. The people with expertise will thrill and make things faster. As if the people with no job were really fired after an expertise check or test. As if your experience has not been devaluated with months of hype.
Sorry. You are not in the street seeing what I see or speaking with the unemployed devs as I am. Please, don't throw me optimism when expert devs are thinking in taking any other job to pay the bills. You can not eat future. You can not pay the bills waiting until this dream someones are having finishes. The hurt is already done. The damage is already here.
Fair, I'll take the correction, you're closer to the damage than I am.
We agree on more than it sounds. The CEOs who sold replacement caused enormous harm and won't pay for any of it. That part is done, the careers it broke don't un-break, and the people calling juniors useless should be ashamed. I'm not throwing optimism at that.
Where I split from you is only on what comes next. Kyle Simpson is unemployed right now, the man who taught a generation how JS works. If he's out, this isn't skill losing value, it's a market that believed a lie and seized up. Lies like that get found out, and the market corrects when they do. Not in time for everyone, I know. But the reasons the floor is here are temporary, and I don't see the one that makes them permanent.
Oh man, I follow your substack because you are one of the remaining sources of truth without hype of LLM in production. And I agree with you more than it sounded. Much more 😁 I'm an avid fan of your articles, if that was not clear before 😉
And yes. I'm agree with you that the market is absolutely irrational, and when the pendule flow backwards they will need the expertise that nowadays are ignoring.
But they also say that markets could stay more time irrational than you can stay solvent. And if the bubble burst, nobody will have cash to hire for certain time. I hope that time is short, but I'm afraid that it will take long. So yeah, agree but disagree in the time line. And that worries me.
And from the other direction this time: the labs sell the model, sell the people who deploy it, and vouch that the output is safe. Same overlap that ended Andersen, just not labeled as auditing yet. Post-IPO, the first supervised deployment that blows up inside a public company turns that overlap from a footnote into a deposition.
Excellent piece. I had not heard of this before.
I am intrigued by the bit of insight you have included into the DevOps services you use internally and also market. With having to dissect and fix vibe-coded apps, I'm curious on your general impression of the code quality and maintainability. If you don't mind a couple of questions: Do you get into refactoring overall, or is your goal just to make sure the code not brittle and insecure? And, to what extent do your workflows utilize agentic coding tools?
Thanks Jim. The goal is almost never a full refactor. It's making the thing safe to run. Exposed keys, no access control, data flowing where it shouldn't, logic that holds on the happy path and nowhere else. We harden that first.
Code quality varies wildly, and it tracks one thing: whether the person who vibe-coded it had real engineering background or none at all. Someone technical produces something brittle but fixable. Someone non-technical produces something where the fastest path is rebuilding it on infrastructure we already trust, with the vibe-coded version serving as the spec.
That's the part that scales. We don't salvage line by line, we rebuild on boilerplates I keep in tight shape. Side effect nobody warned me about: the whole company's code now looks like mine.
On the tooling: we're mostly on Claude Code day to day. Gemini plus http://pi.dev is on the list to evaluate, just haven't gotten to it yet. The model matters less than the boilerplate and the review layer around it, which is the whole point of the original piece.
Probability of error in LLMs follows a power law. You can not automate sh#t with those machines. You need EXPERT in the loop.
I saw many consultancy firms consulting
in the past. In the best case those forward deployed engineers are going to be overworked junior guys that start their career there just for their CV and abandon the firm when they are burn out before been 30 years old. In the worst case, these guys are not going to be engineers. They are going to be agent operators. In any case, salaries will go down. The final goal for which the current developers are so devotely helping to happen.
Expert in the loop, agreed. But it goes one level deeper than that. A lot of clients don't even have a process to automate. They think they do. Then you open the hood and there's no actual workflow, just a few people improvising and calling it a system.
So before anything gets automated, we end up reconstructing the process from scratch, piece by piece, until there's something real to point an agent at. The AI question comes last. Most of the work is figuring out what the company actually does, which no model can tell you.
Sometimes I remember that quote from someone, autonomous driving could work perfectly if the roads were perfect.
I saw enough enterprise codebases and pipelines to know that quality gates is like saying hyper drive engines in most of corporate development environments. Yeah, I'm speaking about huge financing institutions and firms.
So now we are adding sigma event generator machines to that.
The result is not going to be pretty, and the same way we do when garbage dumps grow exponentially, the only solution to that will be to incinerate the garbage.
The lost years we will call these days in the future. When billionaires became trillionaires.
"Works if the roads were perfect" is the cleanest version of the argument I've heard. The whole autonomy pitch assumes a clean substrate, and enterprise pipelines are the opposite. Sigma-event generators on top of code that already has no quality gates, agreed, that compounds fast.
But I read the ending differently. For people who actually have the expertise, almost nothing changed. They just got faster. The mess still needs someone who can read it, and that someone is the same person it always was. The tools moved. The job didn't.
If you like clean, here you you have a senior software engineer with 25 years of experience that only has 100 days of unemployment left in a job market for engineers that at this moment is absolutely broken. Nobody hires. Someone lays off a bunch but nobody fires. All companies and employees are like rabbits surprised by the lights of a car. Nobody knows what to do to be contracted. Juniors have it worse. They are calling them useless. Some unemployed with AI psychosis decide to use agents for making that idea of product they had real, with no bussines plan or anything, just burning tokens and finishing worse than when they started.
And I listen same and same message again. The people with expertise will thrill and make things faster. As if the people with no job were really fired after an expertise check or test. As if your experience has not been devaluated with months of hype.
Sorry. You are not in the street seeing what I see or speaking with the unemployed devs as I am. Please, don't throw me optimism when expert devs are thinking in taking any other job to pay the bills. You can not eat future. You can not pay the bills waiting until this dream someones are having finishes. The hurt is already done. The damage is already here.
Fair, I'll take the correction, you're closer to the damage than I am.
We agree on more than it sounds. The CEOs who sold replacement caused enormous harm and won't pay for any of it. That part is done, the careers it broke don't un-break, and the people calling juniors useless should be ashamed. I'm not throwing optimism at that.
Where I split from you is only on what comes next. Kyle Simpson is unemployed right now, the man who taught a generation how JS works. If he's out, this isn't skill losing value, it's a market that believed a lie and seized up. Lies like that get found out, and the market corrects when they do. Not in time for everyone, I know. But the reasons the floor is here are temporary, and I don't see the one that makes them permanent.
Oh man, I follow your substack because you are one of the remaining sources of truth without hype of LLM in production. And I agree with you more than it sounded. Much more 😁 I'm an avid fan of your articles, if that was not clear before 😉
And yes. I'm agree with you that the market is absolutely irrational, and when the pendule flow backwards they will need the expertise that nowadays are ignoring.
But they also say that markets could stay more time irrational than you can stay solvent. And if the bubble burst, nobody will have cash to hire for certain time. I hope that time is short, but I'm afraid that it will take long. So yeah, agree but disagree in the time line. And that worries me.
Someone at Microsoft-OpenAI should talk to partners at Ernst Young, Arthur Andersen, PriceWaterhouseCoopers and ask them how this worked out.
Especially now that these companies are going public and are subject to Sarbanes-Oxley.
And from the other direction this time: the labs sell the model, sell the people who deploy it, and vouch that the output is safe. Same overlap that ended Andersen, just not labeled as auditing yet. Post-IPO, the first supervised deployment that blows up inside a public company turns that overlap from a footnote into a deposition.
Or worse. These guys have gotten away with so much for so long, they think they’re invincible.
Sort of looking forward to the reckoning.