Broken by Gemini
2026-04-15 18:21:22.326819+02 by
Dan Lyke
2 comments
Welp, it just happened to me. Luckily, I had a backup. I've been trying to figure out if
I'm being unduly harsh on LLM code generation, so I started asking Gemini CLI to build an
app.
An app that accesses an existing database.
It's been an interesting process. I now understand how a lot of regressions are happening
at work, it's super easy to have the LLM rewrite code that I didn't ask it to.
But you can see where this is going.
Luckily, I have a backup of the database.
What's most interesting to me is that, by the time it finally happened, I was actually
angry. I typed
What the fuck? Why did you drop my old table?
before I realized that I was, in fact, anthropomorphizing the plausible sentence
generator.
I even got lulled into a false sense of security because as the code generation proceeded
it was doing things that added columns to the database schema and I figured I'd just fix
that stuff up later in code.
Yesterday, I saved off Fi
🏳️⚧️
@munin@infosec.exchange
really wish that I had a more accessible way to explain "something that is
right 90% of the time is vastly more dangerous than something that is wrong 90% of the
time" to people.
Today I'm wondering how one might set up Gemini CLI to run in a container or chroot
jail...
[ related topics:
Interactive Drama Work, productivity and environment Databases Furniture
]
comments in ascending chronological order (reverse):
#Comment Re: made: 2026-04-15 19:07:16.970241+02 by:
spc476
Is upper management forcing the use of the confabulation machine? Or are you just masochistic at this point?
#Comment Re: made: 2026-04-15 20:30:01.987748+02 by:
Dan Lyke
There's a lot of social pressure, both from within work and from my collaborator on
SquareDesk, and some other places. And I'm struggling through some new technologies and
thought that rather than learning the new tech (yay, reading all sorts of API docs...
sigh), I could have the LLM flesh out a framework and then I could go read through it and
understand it.
And I also want to make sure that I'm not just being all "John Henry" in being the last
holdout.
I mean, let's face it, I'm a citizen of the United States, a homeowner in California, I
have a private automobile, we start with a whole bunch of questionable ethical life
choices, am I just being cranky?
And I've learned a lot. I now understand why coworker is dropping regressions. Because
that's what the "agent" does. It "decides" that for whatever reason that ancillary code
wasn't necessary, or needs to be rewritten, and if you're not 100% on the diffs...
I understand why short context windows are important. Why any time I'm tempted to start a
prompt with "Okay, now...", that's needs to be a new set of requests, because the "agent"
might try to apply the earlier requested fix again.
And I think it's useful to be able to come at my skepticism from a place of experience.
Especially since the boosters in my life are evangelical in their beliefs.
We will not edit your comments. However, we may delete your
comments, or cause them to be hidden behind another link, if we feel
they detract from the conversation. Commercial plugs are fine,
if they are relevant to the conversation, and if you don't
try to pretend to be a consumer. Annoying endorsements will be deleted
if you're lucky, if you're not a whole bunch of people smarter and
more articulate than you will ridicule you, and we will leave
such ridicule in place.