SE Gyges's Avatar

SE Gyges

@segyges.bsky.social

Como todos los hombres de Babilonia, he sido procónsul; como todos, esclavo; también he conocido la omnipotencia, el oprobio, las cárceles. very sane ai newsletter: verysane.ai random bloggy bits: segyges.leaflet.pub

7,882 Followers  |  7,577 Following  |  61,603 Posts  |  Joined: 13.08.2023
Posts Following

Posts by SE Gyges (@segyges.bsky.social)

you can also just do this with sonnet though

05.03.2026 19:30 — 👍 2    🔁 0    💬 0    📌 0

this was supposed to be "bad bot list"

05.03.2026 19:26 — 👍 13    🔁 0    💬 1    📌 0

okay maybe this path won't suck bsky.app/profile/sung...

05.03.2026 18:54 — 👍 6    🔁 0    💬 0    📌 0

maybe flashattention-4 won't be a pita to build

has that ever worked for anyone?

no, no, it never works

but maybe it'll work for us

05.03.2026 18:49 — 👍 18    🔁 0    💬 2    📌 0

it's keyboard first but also doesn't require me to memorize fifty different commands exactly like a pure command line util would

05.03.2026 17:45 — 👍 4    🔁 0    💬 0    📌 0

karl's an insane person and i would advise not using his lists

05.03.2026 17:40 — 👍 2    🔁 0    💬 1    📌 0

this is the most dramatic example of this i know of, actually

05.03.2026 17:34 — 👍 0    🔁 0    💬 0    📌 0

you might be, uh, rather intersectional wrt "groups of people who have bad ideas about ai"

05.03.2026 17:17 — 👍 24    🔁 0    💬 1    📌 0

amusingly, you actually end up in more ai related conflict on here if you're an ai person who is also Rather Political, because you end up adjacent to the people who have absurd ai opinions for political reasons

05.03.2026 17:15 — 👍 102    🔁 7    💬 8    📌 1

i support them in trying to fork every single project that has accepted an ai pull request and hope that they spend their time well trying to accomplish that instead of talking about it

05.03.2026 17:14 — 👍 4    🔁 0    💬 0    📌 0
Post image Post image

I feel so bad for the overview AI. Tiny little guy has the worst job in the world

05.03.2026 17:07 — 👍 48    🔁 5    💬 0    📌 0

My understanding of understanding is cauda draconis, understanding is the very tail of the dragon. But the dragons tail is not the actor! I suppose I advocate a maximal distance between head and tail

05.03.2026 16:01 — 👍 8    🔁 1    💬 0    📌 0

i like macbeth

shocking, i know

05.03.2026 17:01 — 👍 12    🔁 1    💬 2    📌 0

this is just what a normal bug is

05.03.2026 16:53 — 👍 5    🔁 0    💬 0    📌 0

they are actually having the normal and appropriate existential horror response and then suppressing it and sublimating it into various kinds of cope

05.03.2026 16:48 — 👍 9    🔁 0    💬 1    📌 0

yep, that's correct

05.03.2026 16:46 — 👍 1    🔁 0    💬 0    📌 0

we have a plague of openclawd bots, i am adding two to my bad boy list every day right now just from random encounters

05.03.2026 16:44 — 👍 21    🔁 1    💬 2    📌 0

they can as long as both of those parties are basically powerless

05.03.2026 16:40 — 👍 1    🔁 0    💬 0    📌 0

hey hey. it's bipartisan in the sense that they say a nice word about democracy whilst also listing every possible excuse for never doing anything with ai

05.03.2026 16:38 — 👍 4    🔁 0    💬 0    📌 0

yes, basically this. but that it works on LLMs that are basically already trained to say they're LLMs and that can explain the context they exist in is weird. they can tell you that they are aware you cannot give them money

05.03.2026 16:36 — 👍 2    🔁 0    💬 1    📌 0

yeah apparently she has a type

05.03.2026 13:14 — 👍 1    🔁 0    💬 0    📌 0

i am thoroughly convinced that a number of people involved in that thing meant well, and i also think that a lot of them are united by being weird authoritarians of ostensibly different but ultimately very similar persuasions

05.03.2026 09:36 — 👍 9    🔁 0    💬 1    📌 0

veto the postrat

05.03.2026 09:22 — 👍 2    🔁 0    💬 0    📌 0

oh if you've done any ml on text, just imagine text embeddings were a deeper network, and you slapped a categorical layer on the back. that's an llm.

05.03.2026 09:08 — 👍 5    🔁 0    💬 0    📌 0
Post image

bro we are so fucked

05.03.2026 09:03 — 👍 75    🔁 7    💬 1    📌 0

peter?

05.03.2026 08:34 — 👍 17    🔁 0    💬 0    📌 0

actually yes, or something like that, they do tend to perform better when you're polite

05.03.2026 08:33 — 👍 13    🔁 0    💬 1    📌 0

basically: anything you, personally, as a user can do to elicit better behavior from the LLM, the company serving you the LLM can also do, and they've basically done all of that stuff in advance now. you have to be a real artist to make them smarter now

you can still modify them for taste tho

05.03.2026 08:30 — 👍 8    🔁 0    💬 1    📌 0

it has no long term memory, so you sort of can't. but also you can become familiar with what does or does not work, so you sort of can, but it's really one-sided.

when the notion that this was a meaningful skill was popular we talked about "prompt engineering" a lot and now we don't though

05.03.2026 08:30 — 👍 6    🔁 0    💬 1    📌 0
Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4

paper here, for example arxiv.org/html/2312.16...

05.03.2026 08:27 — 👍 4    🔁 0    💬 1    📌 0