A friend of mine constantly reminds me that language is dynamic, not fixed in
stone from a billion years ago. So, if you find others consistently using a
term in a way that you think is wrong, then *you* are wrong in what you think.
The older I get, the more difficult it gets.
But
If GPT systems capture some sense of "usual" context based on trillions of
internet tokens, and that corpus is regarded approximately "global context",
then it seems not so objectionable to call "unusual", new training items that
contribute to fine-tuning loss.
It seems reasonable to worry
I was arguing with that same friend yesterday at the pub. I was trying to describe how
some of us have more cognitive power than others (he's one of them). Part of it is
"free" power, freed up by his upper middle class white good diet privilege. But
if we allow that some of it might be
https://www.math.columbia.edu/~woit/wordpress/
Twenty years of Not Even Wrong, an anniversary blog post.
-- rec --
On Thu, Mar 21, 2024 at 8:48 AM glen wrote:
> [...]
> And "unusual" is even worse. Both tokens require one to describe the
> context, domain, or universe within which the
Thanks everybody. All interesting.
As irony would have it, I was felled by a 72 hour virus and have barely
woken up since posting this inquiry.
By the way, being sick makes me more suceptible to click bait and I
stumbled into this:
Our 10 Best Laptops Best in the US - March 2024 |
Don't remind me, Nick. About 40(?) years ago I paid $5000 for an Apple II
which served as little more than a terminal that could do some graphics.
---
Frank C. Wimberly
140 Calle Ojo Feliz,
Santa Fe, NM 87505
505 670-9918
Santa Fe, NM
On Thu, Mar 21, 2024, 1:19 PM Nicholas Thompson
wrote:
>
Friam at Downtown Subscription tomorrow because of construction at St
John's.
---
Frank C. Wimberly
140 Calle Ojo Feliz,
Santa Fe, NM 87505
505 670-9918
Santa Fe, NM
-. --- - / ...- .- .-.. .. -.. / -- --- .-. ... . / -.-. --- -.. .
FRIAM Applied Complexity Group listserv
Fridays 9a-12p Friday
In the LLM example, completions from some starting state or none, have specific
probabilities. An incomplete yet-unseen (unique) utterance would be completed
based on prior probabilities of individual tokens.
I agree that raw materialist uniqueness won't necessarily or often override