Ramblin' & Amblin' (2026-10-23): But Mike, I only use the AI to
Ramblin' & Amblin' Field Notes with Mike Hall (Staff Software Engineer)
Recorded on the move: engineering observations, systems architecture, and software craftsmanship reflections.
• Series Playlist: https://www.youtube.com/playlist?list=PLNtb969T2-l0
• The Sound Above Curriculum: https://www.just3ws.com/series/the-sound-above/
• Canonical Profile: https://www.just3ws.com
Spoken Overview:
"But Mike, I only use the AI to generate code that I totally understand. There's no hallucinations in my code. It's all as clear and verifiable and provable as any program is. Cool. Cool. Yes, I believe you. I do. I know that that is..."
Recorded: October 23, 2026
The Interviewer
Mike Hall
Host, Ramblin' & Amblin'
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
But Mike, I only use the AI to generate code that I totally
understand. There's no hallucinations in my code. It's all
as clear and verifiable and provable as any program is. Cool.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
Cool. Yes, I believe you. I do. I know that that is
possible.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
What I ask, though, is what are your PRs written in? What
are your commit messages written in? What are your comments
written in? And what is the documentation written in?
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
And if it's written in AI, generative AI, that's the vector
, okay? That's the stop vector. And it's, you might be like,
oh, there's one or two little quirks here or there where it
hallucinates. It's doing more than that.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
Because what you need to realize
, or I shouldn't say you need to realize because I'm ass
erting here. What I'm trying to convey is what I've observed is what I'm
trying to get through across. So you choose how you want to approach this information by
top. What I've observed is that the messages, the Git messages,
are written, for me, in English.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
And they're generated usually by the lowest model settings
with the most limited context window possible because it
has to be able to handle volume. And the lower that is the context and the lower the effort
is, the more room there is for hallucinations. Now, I am not the one saying this. I am saying that working with the AI, it has pointed these
things out to me in the process of our work.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
And, you know, yes, I did say hour, but I'm going to say
that there is a flow that you can get into with these AI
tools while you're building stuff. And it is like a next level REPL kind of experience, and it
does feel very interactive. So I will on occasion use, you know, our and we, because
that is the context that this is being done in. So I'm expressing intent into the AI's prompt context
window, and then I'm getting a product out from that, some
effect.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
And that effect is tangible and practical. But it's also the vector that I'm most likely to not be
paying much attention to because it's the intent of using
these tools is to automate and abstract that away. And that is commit messages, PRs, and comments and
documentation. Again, usually using the cheapest possible tools possible
because they have to be able to operate in CI, they have to
be responsive, they have to be low cost.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
You don't want to be spending a ton on a commit message,
which means that you're going to have things like, now in
this particular case, I'm going to point out,
but it was actually using an AI and automation for building
the descriptions on these videos. And there's one video I noticed that it turned SCMC, which
is Software Craftsmanship McHenry County, which is well
known in my code, my knowledge base. So, much like with the GitHub or whatever, your source
control, version control messages, it was generated using
the lowest possible setting with the lowest model. Usually, in my case, it's right now Gemini Flash or Argon
now.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
But the point is, is that even though I put all these
guards in place for context when I'm working with the
interactive AI,
it's worth sitting in automation and in the supply chain,
the pipeline, where it's usually eyes off or it's at such a
volume that, you know,
nobody's really trying to keep up with it, and it would
just get things wrong because it isn't operating with the
full context of the system. It's not usually thinking mode. It's usually inference and summary because we're trying to
be cheap and fast, cost efficient and fast, whatever you
want to call it. But it's very subtle and one of the most key flows for me,
and that is the publication window.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
A video gets published, and I'm letting these things happen
right now. Like, I'm accepting a certain amount of stop because I'm
moving forward,
but I'm also noting and creating these videos to capture
this context and these observations,
is that because I'm using whatever lightest-weight Gemini,
lowest effort, smallest context window,
so just here's the transcript, and from the transcript
generate a summary for description,
and then it just says, oh, SCMC, I think it said it was, I
don't even want to say it because here's the thing. I say this now again. So in one place in the system, it is now authoritatively
marked in production as SCMC means something completely
different.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
Confidently, as always. Super confident. And now this is out in production, and I will eventually
fix it manually. Probably manually, but it's out there, and then I have my
AI tools doing work on that production corpus,
and I'm all the while thinking everything is hunky-dory,
and it's fine, and, you know, my code is just fine.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
The pipeline is just summarizing the transcript, right? It couldn't get anything wrong. I have all this context, buddy. It absolutely can get things wrong.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
And it will get things wrong because it doesn't understand
what it's reading, and it doesn't ask questions. It just summarizes and synthesizes and ships it. So now think about your GitHub repo, for example. Most of us are probably on GitHub.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
It doesn't matter. It doesn't matter. The point is that inside of that, you're generating text
that's based off of fragments of code that is based off of
the lowest possible effort. And then that gets written into the code base sometimes,
you know, multiple times, you know, an hour.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
You have thousands of commits by the end of your day, right
? And the AI is looking at the history, and it's building out
inferences based off of your commit history, and the
documentation that's being generated, and the comments that
are being generated, and the commit messages that are being
generated. And then it's carrying forward that drift. You know, that SMC being the wrong thing.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
And it... I appreciate it, and everybody says I'm going to read every
line of code. I am telling you right now, if you think that you could
read every line of every change in an application of any
kind of significance that was being implemented by human
beings,
and there was enough of them, and there was enough of them,
the volume, and variety of context. I've worked in enterprise.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
I'm not speaking from no experience here. I'm telling you, I couldn't possibly keep up with the
volume of commits that were going through when it was
humans. AI, 10, 100-fold that. And every single generated prose, no matter how much, how
accurate it was, was a synthesized lie.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
Because it has no idea. It has no idea what you're putting into it. And then you take the lowest possible effort weights, and
the lowest possible effort, and then you just keep cramming
thousands of messages that nobody can keep up with. No person can keep up with the volume.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
And then you say, oh, we'll just put more AI on it. Oh, yeah, that's good. You know what helps with dealing with foxes stealing your
chickens? We'll get more foxes.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
Like, it literally doesn't make sense. But, um, but, that's real. We're going to have this. Once that AI gets the wrong idea on something, you know, it
mistakes a domain mapping.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
It means now that every single diagram I look at that was
generated by code. Now, I love generating diagrams by code. I do feel like it's a good way of using tools to reason
with the structure of a system. I also know that the goal of the system, AI, is to give me
an answer that looks, that answers my request for a map of
the system.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
If I don't have the means to go and verify that that map is
, in fact, accurate, that the map is useless, and if it is
that the AI is saying that a term has been redefined
because of a hallucination,
that's now in your history, your git history, and you're
going to find that that AI is going to keep finding that
over and over and over again,
and you're going to be constantly, and it's going to become
a, uh, a, a, um, cascading effect of all the prose text
that gets generated around our code and applications. Git, commit, version control, commits. Peer review requests, patch requests, PRs, whatever. The documentation that's very much inside of our system,
because when the AI is looking at that, it pulls in context
,
looks for the place it needs to be able to apply a change
in order to affect an outcome, and then just ignores
everything else.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
It doesn't say, hmm, you know, I need to refactor the
documentation. I mean, unless you make a skill to get it to do that, but
still, that's a separate thinking mode. Like, it isn't, it's like you've got a developer to come in
and implement the code,
and then you get a different developer to come in and
document the code. And, um, they never talk, really, um, because they don't
have the concept of talking.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
Um, they have the context, and that's it. They evaluate the context, make assertions, fill in the
gaps, and then proceed from there. Is it possible to make tools that will associate in these,
uh, uh, speculative, uh, generations from, uh, um, determin
istic code? Um, probably.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
But in the meantime, until then, the risk isn't that the
code is going to, um, take your job, per se. It's that it's going to lie, it's going to lie, it's going
to lie, it's going to hallucinate. And we're going to have a false sense of what these systems
are doing, and what's actually out there,
because it is based off of a generation of text that was
looking at some code and turning it into English. It's not building the code.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
Two different engines, two different modalities. One is usually the cheapest and least upper. The other one can actually be proven. How do you map that?
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
How do you, uh, um, enforce that? I have some techniques that I've, I've applied pre-AI to
understand how a system works
and map actual expectations in the face of lost context
where the business, through the kind of things that got us
to where we're at
and we're seeing accelerating with AI of just heads down,
all gas, no brakes, uh, build, build, build,
and then, um, figure it out later. But we can build so fast now, uh, that even when you're
building something, you do understand,
it's hard to keep up with what's doing and how it will
leave you in a place where you think that you've,
you've documented your system, but actually it's just full
of hallucinations and I'm going to be going through more
examples of that. Okay.
Mike Hall
Host, Ramblin' & Amblin'
▶ Speaking
All right. Thanks. Bye.