karashiiro
@karashiiro.moe
inconsistently consistent, all opinions no one's, some rights reserved | https://klink.krs.moe/#/p/karashiiro.moe
Pinned
karashiiro
@karashiiro.moe
· Sep 9
I can't believe I tried horse sashimi and apologized to haru urara and then she immediately fucking died
I've been slopping a lot of code recently for some tight deadlines and noticed an interesting non-technical thing:
I obviously review all the code I ship, but if do I miss things, I basically have to present the reasoning as my own before fixing it. It'd be poor form to just blame the LLM for it
I obviously review all the code I ship, but if do I miss things, I basically have to present the reasoning as my own before fixing it. It'd be poor form to just blame the LLM for it
November 21, 2025 at 9:31 PM
I've been slopping a lot of code recently for some tight deadlines and noticed an interesting non-technical thing:
I obviously review all the code I ship, but if do I miss things, I basically have to present the reasoning as my own before fixing it. It'd be poor form to just blame the LLM for it
I obviously review all the code I ship, but if do I miss things, I basically have to present the reasoning as my own before fixing it. It'd be poor form to just blame the LLM for it
Reposted by karashiiro
yay oauth scopes are actually real
November 15, 2025 at 3:46 PM
yay oauth scopes are actually real
Reposted by karashiiro
The new hire just told us she never had a Facebook page... Legit felt the room go like this
November 12, 2025 at 2:15 AM
The new hire just told us she never had a Facebook page... Legit felt the room go like this
ok I had to share this one too, I love fine-tuning now
November 9, 2025 at 2:24 AM
ok I had to share this one too, I love fine-tuning now
gpt-oss:20b fine-tuwuned
November 7, 2025 at 10:46 PM
gpt-oss:20b fine-tuwuned
Reposted by karashiiro
The story so far: 🧵
Every large language model has an unembedding matrix that turns hidden states (model thoughts) into tokens (output text). This matrix associates a vector with every token in the vocabulary. The idea is you compare the last hidden state with the vectors to pick the right token.
Every large language model has an unembedding matrix that turns hidden states (model thoughts) into tokens (output text). This matrix associates a vector with every token in the vocabulary. The idea is you compare the last hidden state with the vectors to pick the right token.
November 7, 2025 at 5:17 PM
The story so far: 🧵
Every large language model has an unembedding matrix that turns hidden states (model thoughts) into tokens (output text). This matrix associates a vector with every token in the vocabulary. The idea is you compare the last hidden state with the vectors to pick the right token.
Every large language model has an unembedding matrix that turns hidden states (model thoughts) into tokens (output text). This matrix associates a vector with every token in the vocabulary. The idea is you compare the last hidden state with the vectors to pick the right token.
Reposted by karashiiro
Excited to launch an openRxiv partnership with the scientist-run AI review service qed (@qedscience.bsky.social), the brainchild of @odedrechavi.bsky.social 1/n
openrxiv.org/enabling-rev...
openrxiv.org/enabling-rev...
Enabling options for review: from training and transparency to author-centered AI tools - openRxiv
Peer review is widely viewed as a critical aspect of biomedical communication. Ideally, it provides authors with feedback so they can improve manuscripts and gives readers, particularly nonspecialists...
openrxiv.org
November 6, 2025 at 2:32 PM
Excited to launch an openRxiv partnership with the scientist-run AI review service qed (@qedscience.bsky.social), the brainchild of @odedrechavi.bsky.social 1/n
openrxiv.org/enabling-rev...
openrxiv.org/enabling-rev...
Undeprecating for the time being
November 3, 2025 at 6:11 PM
Undeprecating for the time being
Reposted by karashiiro
Decompress DEFLATE using only HTML5 Canvas retr0.id/stuff/canvas...
November 2, 2025 at 9:30 PM
Decompress DEFLATE using only HTML5 Canvas retr0.id/stuff/canvas...
Reposted by karashiiro
There’s an interesting line in between where you’re using it as a sort of research agent to help pull together the papers for a survey. But the point of writing such a paper is to learn a lot or to construct new categories for yourself
November 2, 2025 at 4:58 AM
There’s an interesting line in between where you’re using it as a sort of research agent to help pull together the papers for a survey. But the point of writing such a paper is to learn a lot or to construct new categories for yourself
Reposted by karashiiro
(1/2) Using Ikeiwa's LV Volumetric fog shader and RedSim's "Point Light Volume" component, I was able to simulate a fan casting shadows in some fog.
#VRChat
#VRChat
October 27, 2025 at 10:22 PM
(1/2) Using Ikeiwa's LV Volumetric fog shader and RedSim's "Point Light Volume" component, I was able to simulate a fan casting shadows in some fog.
#VRChat
#VRChat
Finally, yes this is exactly what it is, been having trouble figuring out how to phrase this but I'm stealing this from now on
People who work with computers a lot anthropomorphise them a lot. I don't think demystification changes anything. I've heard this described as "not anthropomorphic about computers, but mechanistic about humans," and this also describes the behaviour of most AI researchers I've worked with.
October 31, 2025 at 4:57 PM
Finally, yes this is exactly what it is, been having trouble figuring out how to phrase this but I'm stealing this from now on
except for actual spam (2 accounts so far, both now suspended) I block and mute nobody, and I have no muted keywords
the difference: I do not use Discover
the difference: I do not use Discover
October 31, 2025 at 4:36 PM
except for actual spam (2 accounts so far, both now suspended) I block and mute nobody, and I have no muted keywords
the difference: I do not use Discover
the difference: I do not use Discover
good thing the labeler caught this account
October 31, 2025 at 3:55 PM
good thing the labeler caught this account
Reposted by karashiiro
Fashion enjoyers, this one is for you!
From Patch 7.4 onward, you'll be able to equip any glamours regardless of your class/job! ✨
Mix and match your favorite outfits and slay on the battlefield! 💅
From Patch 7.4 onward, you'll be able to equip any glamours regardless of your class/job! ✨
Mix and match your favorite outfits and slay on the battlefield! 💅
October 31, 2025 at 12:48 PM
Fashion enjoyers, this one is for you!
From Patch 7.4 onward, you'll be able to equip any glamours regardless of your class/job! ✨
Mix and match your favorite outfits and slay on the battlefield! 💅
From Patch 7.4 onward, you'll be able to equip any glamours regardless of your class/job! ✨
Mix and match your favorite outfits and slay on the battlefield! 💅
Reposted by karashiiro
azure got jealous at the amount of attention aws was getting so they decided to go down too
October 29, 2025 at 5:28 PM
azure got jealous at the amount of attention aws was getting so they decided to go down too
Reposted by karashiiro
Umamusume answers this on a functional level at least.
October 31, 2025 at 4:58 AM
Umamusume answers this on a functional level at least.
Reposted by karashiiro
We just crossed it. Yay! Congrats everyone and Happy Halloween!
October 31, 2025 at 8:35 AM
We just crossed it. Yay! Congrats everyone and Happy Halloween!
Reposted by karashiiro
Context-Bench also measures total cost to complete the benchmark. Surprisingly, raw token costs ($/million tokens) do not map directly to total cost.
GPT-5 has lower per-token cost than Sonnet 4.5, but costs more in the benchmark because GPT-5 agents are more "token hungry".
GPT-5 has lower per-token cost than Sonnet 4.5, but costs more in the benchmark because GPT-5 agents are more "token hungry".
October 30, 2025 at 8:08 PM
Context-Bench also measures total cost to complete the benchmark. Surprisingly, raw token costs ($/million tokens) do not map directly to total cost.
GPT-5 has lower per-token cost than Sonnet 4.5, but costs more in the benchmark because GPT-5 agents are more "token hungry".
GPT-5 has lower per-token cost than Sonnet 4.5, but costs more in the benchmark because GPT-5 agents are more "token hungry".
Another protocol (there's no protocol here yet, looks like a vision paper)
The one interesting idea I see here is "what if we gave agents DIDs" but I'm not entirely sure what the implications of that would be
The one interesting idea I see here is "what if we gave agents DIDs" but I'm not entirely sure what the implications of that would be
Agent Network Protocol White Paper
w3c-cg.github.io
October 31, 2025 at 1:33 PM
Another protocol (there's no protocol here yet, looks like a vision paper)
The one interesting idea I see here is "what if we gave agents DIDs" but I'm not entirely sure what the implications of that would be
The one interesting idea I see here is "what if we gave agents DIDs" but I'm not entirely sure what the implications of that would be
Claude Code deprecating output styles in favor of plugins means you now either need to paste them into the global context file or use a SessionStart hook which isn't portable
October 31, 2025 at 2:07 AM
Claude Code deprecating output styles in favor of plugins means you now either need to paste them into the global context file or use a SessionStart hook which isn't portable
I wrote some stuff about the internal markup Anthropic uses in Claude
A Look at ANTML: The Anthropic Markup Language
A brief trip into the depths of Anthropic's API
karashiiro.leaflet.pub
October 30, 2025 at 4:32 PM
I wrote some stuff about the internal markup Anthropic uses in Claude
wait did people not take DARE seriously, have I been living under a rock
October 30, 2025 at 3:15 PM
wait did people not take DARE seriously, have I been living under a rock
Is this the year we evolve past J-scores
the accuracy is also dramatically better than LLM-as-a-judge
white box: their method, using SAE
black box: an LLMs own judgement via text outputs
it’s a lot cheaper because they only need one set of activations, i.e. only one output token + a single random forest inference
white box: their method, using SAE
black box: an LLMs own judgement via text outputs
it’s a lot cheaper because they only need one set of activations, i.e. only one output token + a single random forest inference
October 30, 2025 at 3:13 PM
Is this the year we evolve past J-scores