On 8/23/2025 3:56 AM, John Clark wrote:
On Fri, Aug 22, 2025 at 9:43 PM Brent Meeker <[email protected]>
wrote:
/>>Having a better AI model bypasses the need for more
electricity/
>> *No it does not! Regardless of how efficient an AI model is,
the more computational resources it has available the more
intelligent it will be, and the more new computer models you
will be able to train.*
*
*
/> Yes it does! Regardless of how much more electricity you have
a more efficient model may be superior. /
*Superior to what? *
Other AI; with a less efficient model. The same thing that more
training (and electricity) aims at. You seem to notice this, e.g. HRM,
when it suits your argument.
Reading the HRM paper, it's not clear to me how the "R" part works. Is
it just a recurrent neural net that's smaller and runs faster, or does
it actually have some logic built in?
*If your AI is not lousy then more computing power will result in more
intelligence. If your AI is lousy then more computing power will
enable you to find a new AI model, or even a new AI architecture, that
is less lousy. So your assertion that a more efficient AI will lead to
less demand for computing power is simply not true. *
I wrote:*"*/Regardless of how much more electricity you have a more
efficient model may be superior. " /I didn't say it would lead to less
demand for computing power, but that it /might/ divert the race for more
electrical power. The demand for electrical power is driven by
training. At some point LLM's will have absorbed all recorded human
knowledge, but will only be smart the way Encyclopedia Britannica is smart.
*
*
*That's why on January 27 when Nvidia lost $600 billion in one
daybecause the Chinese had made a more efficient AI, the largest one
day loss in Wall Street history, I took that opportunity to buy more
Nvidia stock. *
So did I. And so far it's paying off.
*The market had forgotten about Jevons Paradox, but I hadn't. I don't
know which AIwill end up winning the race but I do know that whoever
the winner is he, she or it will never feel that it has too much
computing power.In a gold rush the people who make the most money are
not those who dig holes in the ground but those who sell picks and
shovels. *
**
/> Races are only one-dimensional./
*Not if there are a huge number of branching ways to reach the finish
lineand it's not obvious which path will get you there in the least
amount of time. *
Right. That was a typo. I meant to write*"*/Races aren't only
one-dimensional." /Noting that more training in bigger LLM's isn't the
only path./
/Brent
--
You received this message because you are subscribed to the Google Groups
"Everything List" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to [email protected].
To view this discussion visit
https://groups.google.com/d/msgid/everything-list/53d528b3-f19b-4435-8c70-1ca4f12c5437%40gmail.com.
