[conspire] Discussion: Using LLMs the Right Way: 10/1/2025 7pm Eastern Daylight time

[Deirdre Saoirse Moen]

On Oct 3, 2025, at 04:57, Ivan Sergio Borgonovo <mail at webthatworks.it> wrote:
> 
> On 10/3/25 6:59 AM, Deirdre Saoirse Moen wrote:
>> On Wed, Oct 1, 2025, at 4:40 PM, Ivan Sergio Borgonovo wrote:
>>> The failure of your AI is mostly due to  Mercurial having a market share
>>> of 0.17-2% (according to AI).
>>> Not your fault having to use Mercurial of course.
>> I guarantee you that a FAANG known for its AI and able to create the kind of Mercurial-AI integration I described (and quoted) isn't hurting because they're using a less-common SCM.
> 
> Curious. I thought your thesis was LLM don't work.

My occasional sloppy use of language doesn’t change my position. Y’all are the ones calling it AI. Also, like it or not, it is in the general field of artificial intelligence, a field I actually have worked in.

LLMs are a parlor game that use far more resources than they can provide value to society. They don’t “work” in the sense that they are not intelligent, and are not true AI. At *best* you will get a junior dev out of them, but a junior dev who makes mistakes like deleting code that shouldn’t be deleted. Many people do not have enough exposure to LLMs to realize they are basically attempting to recreate that kid who constantly screws up then pretends nothing bad happened, then wonders why you stopped talking to them.

That said, I did find it useful to be able to interactively ask questions that were specific to $FIRM’s processes and have a live answer that was usually at least close enough to correct I could find the answer from where it led me.

But, as an example of things it did poorly, it suggested that I compile with X86 flags on, and Apple hasn’t shipped a new X86 machine since 2019, nor was I on X86 hardware (so it would have created a non-working binary for me if I’d even had the X86 platform installed, which I did not).

For anyone following along and wondering what the “right” solution to the problem I had was, it was the combination of moving revs to a new stable unset some flags that needed to be set before the commit could be added to the diff, and I was just tired AF that day and missed that step.

I wouldn’t have had problems if I hadn’t also had to manage backing out another engineer’s change as a temporary fix in order to compile at all (because the change that needed to go with it hadn’t propagated through to the stable branch yet), because that made the number of things I had to track an order of magnitude more complex than I could manage that particular day (because tired AF).

Deirdre