Re: LLM usage in Debian -- ballot options comparison

On Sun, Aug 02, 2026 at 05:28:06PM -0500, Simon Richter wrote:
> The Internet bubble left fiber connections behind, but the AI bubble is only
> building out datacenter capacity for "bandwidth over latency" applications
> with a weird floating point format.

The AI bubble story is a lot more complicated than that, and I am
worried that a lot of the data centers which is being financed using
private financing may not have the longevity that is being claimed by
companies like SpaceX or other NeoClouds.

But the wierd floating point format shouldn't be the argument which
gets used.

> People have tried using TPUs for EM field simulations and found them to be
> largely useless because 16 bit floats aren't all that great.

You should *never* try to use LLM's to do math, and in fact, that's
not how it gets used when it is used properly.

For example, when I asked Gemini to give me a model answering a
financial investment question (e.g., under what assumptions should I
sell highly appreciated stock, and take the tax hit right away, but
allowing me to diversify my holdings, and when does the advantage of
the diversified holdings over take the initial capital gains tax
cost).  The way the LLM answered my question was that it generated a
Python script, and then ran that Python program in a sandbox, and the
formatted the results in a table illustrating the assumptions when the
breakeven would be hit.

More importantly, it allowed me to download the Python script, and I
was able to take that script and modify and enhance it to try out different
assumptions.

This was incredibly useful to me, and given that the Python program
was created specifically to model the question that I was asking, (a)
the specialized floating point formay used by LLM's doesn't matter,
(b) we don't have to worry about all of the (valid) concerns of LLM
not being able to do math, and (c) given that the python script was
generated to specific to my question, unless you believe that how to
use python libraries to do time value of money calculations are a
potential copyright problem, I'm not particularly worried about
copyright issues.  (The python script was a throwaway, and I wasn't
planning on redistributing it anyway.  :-)

This is the sort of thing is why I'm quite happy paying $200/year for
a Gemini Pro subscription.  (Disclosure: I work for Google, but I'm
not saying this because I'm shilling for my employer.  I'm also a
$200/year Claude Code subscriber as well, in addition to owning an
128GB M5 Max Macbook Pro, so I can run local LLM's.  I use the correct
tool for the job, and in some cases, it's not cloud-based LLM's, but
rather one that runs on my laptop.)

> So the long term effects are closer to that of the BitCoin bubble than to
> that of the Internet bubble.

Everyone points at the dark fiber post Internet-bubble which took ten
years to eventually light up as the analog to the AI bubble.  But what
most people forget was that after the dot.COM bubble, there were
$100k+ USD Sunfire servers that were selling for pennies on the
dollar, and Sun servers ten years later don't hold value the same way
dark fiber did.  So that might be the closer analog for the portion of
the AI bubble which comes from people buying Nvidia GPU's.

There may very well be an AI bubble, just as there was a dot.COM
bubble.  And there were certainly downsides for all of the data
centers which allowed us to more easily buy goods at a much cheaper
and more efficient way on-line, not to mention to stream movies using
Netflix or Disney+.  Furthermore, a lot of small mom-and-pop small
business that used to sell computers have been driven out of business
by people buying computer equipment from companies like Provantage,
CDW, and Amazon.  But most people probably take ecommerce for granted,
and they aren't arguing that we should somehow go back to the way
things were sold pre dot-Com bubble.

Cheers,

						- Ted