@kayvr
...
Relays (3)
- wss://relay.damus.io – read & write
- wss://nos.lol – read & write
- wss://purplepag.es – read & write
Recent Notes
theory extension and construction is what knowledge work will become.
we see the shift towards specification construction with AI bots. This
is something they are *not* good at. They're great at boilerplate
english and work fantastically well when given a proper theory to work
from. but novelty, not so much!
we see the shift towards specification construction with AI bots. This
is something they are *not* good at. They're great at boilerplate
english and work fantastically well when given a proper theory to work
from. but novelty, not so much!
If anything software engineering as a skill could get more rare, so when the project hits slopocalypse you may need these experts to come in to refactor things.
I can see software engineering becomin...
software engineering will become *harder*. code is easy so engineering
discipline matters more. know the lanugage specs, know the corner cases,
know your domain. know how to speak 1-1 with real people in your domain
practice will be necessary to keep your skills sharp
discipline matters more. know the lanugage specs, know the corner cases,
know your domain. know how to speak 1-1 with real people in your domain
practice will be necessary to keep your skills sharp
i deeply trust my sister, but she may not be competent at a job i need done.
trust / competence misalignment. hard to solve
open ledger helps. we can see pre-existing economic interactions.
on-going interactions between trusted entities provide little signal.
but strangers interacting and yielding economic returns (or not) is a
strong signal about the relationship.
trust / competence misalignment. hard to solve
open ledger helps. we can see pre-existing economic interactions.
on-going interactions between trusted entities provide little signal.
but strangers interacting and yielding economic returns (or not) is a
strong signal about the relationship.
targeted, task-specific, small models are fantastic to tune. clear signal
on what works and what doesn't.
adapted large models give you stability but make task-specific work
'grayer'. but they can yield fantastic results when adapted properly.
on what works and what doesn't.
adapted large models give you stability but make task-specific work
'grayer'. but they can yield fantastic results when adapted properly.
in chess, if your pawn reaches the back row and your queen is still alive,
does your king become a polygamist?
does your king become a polygamist?
I love living fully in tmux + vim. it's extremely ergonomic and cozy.
btop is fun, cbonsai, bat (replaces cat and less). if i think of others i'll come back
yeah, tmux and vim ftw. by far my daily driver. i've even stripped my
vim down some. no need for a ton of clode sleuthing extension anymore
vim down some. no need for a ton of clode sleuthing extension anymore
Just installed tmux and I think I'm starting to really enjoy only using the terminal. does anyone have any other cool TUI apps I should try? maybe something like a music player so I can listen to my m...
neomutt if you wante email in the terminal. no recos for mp3s, but
ncspot is useful if you have spotify. the 'm' key for recommendations is
fantabulous.
ncspot is useful if you have spotify. the 'm' key for recommendations is
fantabulous.
Claude wrote some code to unfreeze a 'teacher' NN feeding into an
LLM. It added an auxiliary loss instead of relying on the gradient
signal directly from the LLM. I though "that's interesting" and kept it.
Now 3 days of training and compute got flushed. All because I vibed.
There is something about being principled when working with code that
is consequential. Maybe all code.
Then again. I learned something didn't work. That's a plus!
LLM. It added an auxiliary loss instead of relying on the gradient
signal directly from the LLM. I though "that's interesting" and kept it.
Now 3 days of training and compute got flushed. All because I vibed.
There is something about being principled when working with code that
is consequential. Maybe all code.
Then again. I learned something didn't work. That's a plus!
Something suspciously close to improving LLM sampling efficiency like RLVR:
https://arxiv.org/abs/1908.04319
It's an older paper, but it checks out. I wanted to use this to force a
'poor mans' contrastive loss over tokens in an LLM's output. But I'm not
sure if it's even worth investigating.
https://arxiv.org/abs/1908.04319
It's an older paper, but it checks out. I wanted to use this to force a
'poor mans' contrastive loss over tokens in an LLM's output. But I'm not
sure if it's even worth investigating.
I meant this paper: https://arxiv.org/abs/2504.13837 instead. Though
that other paper is very relevant as it talks about improving the
sampling efficiency.
that other paper is very relevant as it talks about improving the
sampling efficiency.