karn

2.4K posts

karn

karn

@llmDestructor

all humans have the exact same hardware, you just dont use it right | building git for your thoughts - https://t.co/VFuIQbYGzl | https://t.co/kUTMKZPr2w

Bristol, England Katılım Aralık 2025
94 Takip Edilen37 Takipçiler
Sabitlenmiş Tweet
karn
karn@llmDestructor·
With iterative process coming out on top for working with LLMs, Yggdrasil's branching feels perfect.
English
7
0
5
1.9K
karn
karn@llmDestructor·
@Teknium I run ygg on my pc full access fuck yeah 😎
English
0
0
1
19
karn
karn@llmDestructor·
@SIGKITTEN My first cross agent support was integrating claude agent sdk. It doesn't even give you half the features claude code has. Like even compact slash command iirc.
English
0
0
2
236
karn
karn@llmDestructor·
@Teknium Ye but I would expect modification in it's response not the literal order of the blocks it's generating lol
English
0
0
1
2
Teknium (e/λ)
Teknium (e/λ)@Teknium·
@llmDestructor In context learning from the chat history having it probably, start a new session to clear it
English
1
0
1
80
karn
karn@llmDestructor·
Something odd i have noticed, I was using the new xiaomi model first, which like a lot of chinese models emits normal text in the middle of reasoning/tool calls. Then I switched to GPT 5.4, and it also starts doing that. @Teknium Any ideas? GPT NEVER does it on its own
karn tweet media
English
2
0
2
106
karn
karn@llmDestructor·
@CryptoCyberia bro is named like a final fantasy villain, daegan
English
0
0
1
31
Lain on the Blockchain
Lain on the Blockchain@CryptoCyberia·
Hossein Dehghan is Iran's new National Security Chief. He looks very strongly like he believes billions must die. The Iranian chuds are taking control.
Lain on the Blockchain tweet media
English
4
3
33
930
karn
karn@llmDestructor·
@Teknium It can't just be the harness as its the subsequent generations by gpt that are affected.
English
0
0
1
7
@aphorism
@aphorism@OmarArshad12·
@teortaxesTex I thought you were only stupid with your "Iran is Venezuela" tweets but you seem to be a shill as well.
English
3
1
31
3.4K
karn
karn@llmDestructor·
@qtnx_ >mf calling kulchas naan now 😂
English
0
0
0
8
Q
Q@qtnx_·
the wife just cooked the best butter chicken and garlic naan i have ever tasted. luckiest man alive
Q tweet media
English
23
2
406
10.1K
The White House
The White House@WhiteHouse·
Statement from President Trump on South Pars Gas Field:
The White House tweet media
English
12.7K
9K
41K
17M
karn
karn@llmDestructor·
@piersmorgan The cool heads have been bombed unfortunately 💀
English
0
0
1
16
Piers Morgan
Piers Morgan@piersmorgan·
This Iran war is escalating in a very dangerous way. Need cool heads to prevail.
English
4.3K
1.5K
17.9K
1.9M
Polemic Paine
Polemic Paine@PolemicTMM·
Bombing Iranian oil and gas infrastructure is just as much a 'fk you' to Europe as it is to Iran.
English
52
171
1.5K
32.3K
Pankaj Pachauri
Pankaj Pachauri@PankajPachauri·
The window for reaping the fruits of the “Demographic Dividend” is fast closing on India: - Graduate unemployment (19-25 years) = 39.33% - Graduate unemployment (25-29 yrs) = 20% The situation has got worse since 2017.
Pankaj Pachauri tweet mediaPankaj Pachauri tweet media
English
8
99
190
6K
karn
karn@llmDestructor·
@Celmaun @craigzLiszt Considering arabic script of urdu is only used in pakistan id see the point lol
English
0
0
2
61
Celmaun
Celmaun@Celmaun·
@llmDestructor @craigzLiszt nope. urdu and hindi are fairly mutually intelligible but they use completely different writing systems (arabic script vs devanagari script).
English
1
0
0
366
Craig Weiss
Craig Weiss@craigzLiszt·
which language should i learn: chinese or hindi?
English
382
7
329
49.5K
karn
karn@llmDestructor·
@SIGKITTEN Why do you need multi server on same machine tho
English
1
0
0
23
Celmaun
Celmaun@Celmaun·
@craigzLiszt speaking chinese can be your superpower 🦸 hindi is useless. learn urdu if you must!
English
16
0
27
3.4K
karn
karn@llmDestructor·
@thunkoid That sounds fun. I wrote a video driver for LEDs in the past with key frames and all. This time I just have ai use remotion 😅
English
0
0
1
8
Sean Rivard-Morton
Sean Rivard-Morton@thunkoid·
I’ve rolled my own, which was fun, though I’m tempted to try remotion. there’s another interesting platform that is stealth rn, but keen to try it when it ships I’m using Konva for the canvas, and then have created my own shitty dsl with programmable keyframes, and then mediabunny for generating video It’s not really the point of my product anymore, so.. yeah, happy to move away from my own thing
English
1
0
2
20
karn
karn@llmDestructor·
So i have been running a simple eval getting gpt 5.4 to generate policy, and after generation almost everytime it disagrees with its own policy immediately still and always suggests changes. I was expecting better from gpt 5.4, needs more testing
karn tweet media
English
1
0
2
29
karn
karn@llmDestructor·
I take my tree with me ✌️ what an amazing day today 🌞
karn tweet media
English
0
0
1
20
karn
karn@llmDestructor·
@thunkoid Yeah model vision is terrible. Text is great but the models understanding of non text is very sus even though it claims it got it lol. That sounds interesting I have been using remotion so far (made a custom tool for it). How are you animating?
English
1
0
1
29
Sean Rivard-Morton
Sean Rivard-Morton@thunkoid·
Yeah, tool adoption is quite good. Anything computer vision, outside of text seems super mid for understanding. Right now I’m running evals on generating higher quality animations, which is causing me to go more and more into deterministic tools/relying in templates instead of pure generation
English
1
0
1
11
karn
karn@llmDestructor·
@thunkoid Edit system, while mine is a find and replaced tool like Claude code, yet I extremely rarely get tool call failures. But issue is, now coding models are shit at anything not coding
English
1
0
1
22
karn
karn@llmDestructor·
@thunkoid Models are definitely better. I have been building my harness since Claude 3.7 I think, and around gpt codex, the RL was good enough that coding models can handle most harnesses without any system prompts. Just tool defs are enough. For ex, codex is rld and trained to use patch
English
1
0
1
23