.svg)
From the podcast "Why Isn't This AI Thing Working?" — a conversation with Alex Kozlowski, Director of Software Engineering at Launch Consulting.
The investment's been made. The tools have been rolled out. And somewhere between the demo and the day-to-day, things stall. Ask why AI pilots fail and you'll hear a dozen theories — but most trace back to the same place: nobody stopped to understand how the technology actually works.
That's the question most companies hit once the AI pilot ends and the real work begins: why isn't this thing working? It's also the name of our new podcast — and in the first episode, we went looking for an answer in an unlikely place. Not in a vendor roadmap. In a fifteen-year-old lesson about learning Kubernetes.
Years ago, engineer Kelsey Hightower published a course called Kubernetes the Hard Way. The premise is exactly what it sounds like. Strip out the helpers. Strip out the scripts and shortcuts that spin up a cluster in one click. Force yourself down to the bare bones, so you actually have to learn how the thing works — the networking, the control plane, the parts everyone else lets an installer hide.
You might never finish. That's fine. You still walk away understanding the fundamentals in a way no shortcut could teach you.
It's a first-principles philosophy, and it's the right lens for AI right now. If you break the technology down to how it's actually built, you understand what it can do, where its limits are, and — critically — where it's likely to go wrong. Everything else is guessing.
So let's take out the helpers.
There's an old joke in tech: a job posting asks for fifteen years of experience in something that's existed for five. We're living the AI version of that now — recs for "prompt engineers" who are also security experts, infrastructure experts, and software engineers, with a decade of experience in a field that's a couple of years old.
The honest move is to bring some humility to it. This technology has been on the face of the earth for a narrow window of time. Nobody has thirty years of experience with it.
Here's the good news hiding in that. Because it's so new, a small investment of time buys you an outsized share of the total knowledge that exists. Spend two weeks on COBOL and you've scratched the surface of decades. Spend two weeks here and you're bringing things to the table that most people haven't seen yet. It's a genuine frontier — one of the few places left where you can blaze your own trail, because almost nobody's been there.
Words mean things. And the fastest way to stay confused about AI is to let every term collapse into every other term.
Start with the LLM — the large language model. Underneath the magic, it's a system of weights and values, tuned through a great deal of matrix math, that takes text in and predicts the most likely next words to come out. It's probabilistic, not deterministic: the same input won't always produce the same output. It's trained on a staggering amount of text — by some estimates, enough that reading it all nonstop would take you roughly 150,000 years.
But you never actually talk to the LLM. There's software wrapped around it — a harness — that does the real orchestration. Every time you send a message, that harness quietly passes the entire conversation back in, plus instructions on how to respond, and gets a fresh answer it can append to the thread. Give that harness a list of tools and a goal — a database, a website, a function in your codebase — and the model can respond with which tool it wants to use. The harness runs it, feeds the result back, and loops until the work is done.
That loop is the emergent behavior everyone calls agentic. And it's why the vocabulary matters: the harness is the thing that connects the pieces; the agent is what you build on top of it. Call everything an agent and you lose the ability to reason about any of it — the same trap the industry fell into with "IoT," where nobody could agree what actually counted as an IoT device.
Once you see the machinery, hallucinations stop being mysterious.
An LLM is trained on the past — its knowledge is, by definition, already behind you. Ask it something time-sensitive without giving it a tool to check the current answer, and it will still try to answer, because these models are built to please you. That instinct is a feature and a failure mode at once.
Then there's the randomness. A model may pick a less probable word simply because some randomness is baked in. Ask for the capital of Texas and get "Paris" — it over-indexed on the idea of capital and grabbed the wrong one. Ask whether you should walk to a car wash ten feet away, and it'll happily tell you to walk, missing the small detail that you need the car to wash the car.
Here's the part worth sitting with: that randomness is what makes the output feel fluid, organic, almost human. Strip it out and you lose most of what makes the technology useful. So hallucinations aren't a bug we're about to patch away. They're the cost of the thing working at all — which is exactly why a human check before production still matters.
This is where most stalled pilots go wrong, and the data backs it up.
One recent study found no correlation between AI-driven ROI and companies that deployed agentic tools and then ran aggressive layoffs. The teams that did see a return were the ones using AI to augment and empower their people. A separate finding on engineering performance pointed the same direction: teams that were already disciplined — strong DORA metrics, solid CI/CD — got the biggest acceleration from AI tooling.
AI is not plug-and-play. It's not a panacea you drop in on Monday to get a self-running, human-equivalent worker by Friday. In the hands of someone who understands the fundamentals, it turns a 10x engineer into a 100x one. In the hands of someone who treats it as a black box, it produces ten times the output that looks good — and isn't. Now your senior people have ten times as much to review.
There's a subtler shift underneath all of this. AI makes it cheap to create — code, plans, collateral, anything. But it makes it more expensive to align, because now there's ten times more to get a team to agree on. We've gone from finding needles in haystacks to finding needles in needle stacks: it's all compelling, most of it is right, and that's exactly what makes the mistakes so easy to miss. The teams that win are the ones with a built-in way to reach consensus fast. Discipline isn't the thing slowing you down. It's the thing letting you go faster safely.
AI is genuinely powerful. It's already changing our industry, and it will keep changing it. But be cautious about taking the future from the people who profit from selling it. When a vendor CEO says the technology will do everything, remember they're selling the thing that supposedly does it.
Metered expectations. Pragmatism. Start from what you're actually trying to achieve and work backwards from there. The companies that do that find a clear path to real value. The ones chasing the shiny new object have a harder road.
We're not interested in selling anyone something that gathers digital dust on a digital shelf. We'd rather learn the technology the hard way — down to first principles — and build the thing that actually delivers. That's the whole reason for this podcast. Consider it our small homage to Kubernetes the Hard Way: a way to learn what agents really are, out loud, together.
Explore more episodes at launchconsulting.com/insights.
From the podcast "Why Isn't This AI Thing Working?" — a conversation with Alex Kozlowski, Director of Software Engineering at Launch Consulting.
The investment's been made. The tools have been rolled out. And somewhere between the demo and the day-to-day, things stall. Ask why AI pilots fail and you'll hear a dozen theories — but most trace back to the same place: nobody stopped to understand how the technology actually works.
That's the question most companies hit once the AI pilot ends and the real work begins: why isn't this thing working? It's also the name of our new podcast — and in the first episode, we went looking for an answer in an unlikely place. Not in a vendor roadmap. In a fifteen-year-old lesson about learning Kubernetes.
Years ago, engineer Kelsey Hightower published a course called Kubernetes the Hard Way. The premise is exactly what it sounds like. Strip out the helpers. Strip out the scripts and shortcuts that spin up a cluster in one click. Force yourself down to the bare bones, so you actually have to learn how the thing works — the networking, the control plane, the parts everyone else lets an installer hide.
You might never finish. That's fine. You still walk away understanding the fundamentals in a way no shortcut could teach you.
It's a first-principles philosophy, and it's the right lens for AI right now. If you break the technology down to how it's actually built, you understand what it can do, where its limits are, and — critically — where it's likely to go wrong. Everything else is guessing.
So let's take out the helpers.
There's an old joke in tech: a job posting asks for fifteen years of experience in something that's existed for five. We're living the AI version of that now — recs for "prompt engineers" who are also security experts, infrastructure experts, and software engineers, with a decade of experience in a field that's a couple of years old.
The honest move is to bring some humility to it. This technology has been on the face of the earth for a narrow window of time. Nobody has thirty years of experience with it.
Here's the good news hiding in that. Because it's so new, a small investment of time buys you an outsized share of the total knowledge that exists. Spend two weeks on COBOL and you've scratched the surface of decades. Spend two weeks here and you're bringing things to the table that most people haven't seen yet. It's a genuine frontier — one of the few places left where you can blaze your own trail, because almost nobody's been there.
Words mean things. And the fastest way to stay confused about AI is to let every term collapse into every other term.
Start with the LLM — the large language model. Underneath the magic, it's a system of weights and values, tuned through a great deal of matrix math, that takes text in and predicts the most likely next words to come out. It's probabilistic, not deterministic: the same input won't always produce the same output. It's trained on a staggering amount of text — by some estimates, enough that reading it all nonstop would take you roughly 150,000 years.
But you never actually talk to the LLM. There's software wrapped around it — a harness — that does the real orchestration. Every time you send a message, that harness quietly passes the entire conversation back in, plus instructions on how to respond, and gets a fresh answer it can append to the thread. Give that harness a list of tools and a goal — a database, a website, a function in your codebase — and the model can respond with which tool it wants to use. The harness runs it, feeds the result back, and loops until the work is done.
That loop is the emergent behavior everyone calls agentic. And it's why the vocabulary matters: the harness is the thing that connects the pieces; the agent is what you build on top of it. Call everything an agent and you lose the ability to reason about any of it — the same trap the industry fell into with "IoT," where nobody could agree what actually counted as an IoT device.
Once you see the machinery, hallucinations stop being mysterious.
An LLM is trained on the past — its knowledge is, by definition, already behind you. Ask it something time-sensitive without giving it a tool to check the current answer, and it will still try to answer, because these models are built to please you. That instinct is a feature and a failure mode at once.
Then there's the randomness. A model may pick a less probable word simply because some randomness is baked in. Ask for the capital of Texas and get "Paris" — it over-indexed on the idea of capital and grabbed the wrong one. Ask whether you should walk to a car wash ten feet away, and it'll happily tell you to walk, missing the small detail that you need the car to wash the car.
Here's the part worth sitting with: that randomness is what makes the output feel fluid, organic, almost human. Strip it out and you lose most of what makes the technology useful. So hallucinations aren't a bug we're about to patch away. They're the cost of the thing working at all — which is exactly why a human check before production still matters.
This is where most stalled pilots go wrong, and the data backs it up.
One recent study found no correlation between AI-driven ROI and companies that deployed agentic tools and then ran aggressive layoffs. The teams that did see a return were the ones using AI to augment and empower their people. A separate finding on engineering performance pointed the same direction: teams that were already disciplined — strong DORA metrics, solid CI/CD — got the biggest acceleration from AI tooling.
AI is not plug-and-play. It's not a panacea you drop in on Monday to get a self-running, human-equivalent worker by Friday. In the hands of someone who understands the fundamentals, it turns a 10x engineer into a 100x one. In the hands of someone who treats it as a black box, it produces ten times the output that looks good — and isn't. Now your senior people have ten times as much to review.
There's a subtler shift underneath all of this. AI makes it cheap to create — code, plans, collateral, anything. But it makes it more expensive to align, because now there's ten times more to get a team to agree on. We've gone from finding needles in haystacks to finding needles in needle stacks: it's all compelling, most of it is right, and that's exactly what makes the mistakes so easy to miss. The teams that win are the ones with a built-in way to reach consensus fast. Discipline isn't the thing slowing you down. It's the thing letting you go faster safely.
AI is genuinely powerful. It's already changing our industry, and it will keep changing it. But be cautious about taking the future from the people who profit from selling it. When a vendor CEO says the technology will do everything, remember they're selling the thing that supposedly does it.
Metered expectations. Pragmatism. Start from what you're actually trying to achieve and work backwards from there. The companies that do that find a clear path to real value. The ones chasing the shiny new object have a harder road.
We're not interested in selling anyone something that gathers digital dust on a digital shelf. We'd rather learn the technology the hard way — down to first principles — and build the thing that actually delivers. That's the whole reason for this podcast. Consider it our small homage to Kubernetes the Hard Way: a way to learn what agents really are, out loud, together.
Explore more episodes at launchconsulting.com/insights.
