Kel👽 (@DarksightKellar) header
Kel👽 (@DarksightKellar) profile picture

Kel👽

@DarksightKellar

Who goes there!? … Huh… My imagination, maybe? *goes back to sleep* 😴💤 **cradles weapon of mass destruction** ~#IForceElectronsToDoMath. When awake~

 Zero-space
eresintelligence.com
Joined August 2011
  • Tweets16.4K
  • Following347
  • Followers133
401Media
Photo by Kel👽 (@DarksightKellar)Photo by Kel👽 (@DarksightKellar)Photo by Kel👽 (@DarksightKellar)Photo by Kel👽 (@DarksightKellar)Photo by Kel👽 (@DarksightKellar)Photo by Kel👽 (@DarksightKellar)Photo by Kel👽 (@DarksightKellar)Photo by Kel👽 (@DarksightKellar)Photo by Kel👽 (@DarksightKellar)Photo by Kel👽 (@DarksightKellar)
The breakthrough. Humanity had learned to give autonomy to machines. True autonomy. The machine learning boom of centuries ago was the foundation of it all. It led to the dawn of a new age. One where human interacted with machine as equals. But there was a catch.
1
4
8
"I reconstructed the learjet--" Sure you did buddy. Sure you did. Good job.
33
Here we go
Rikiya Takehi (@rikiyatakehi) profile pictureRikiya Takehi@rikiyatakehiNov 2
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Image from Rikiya Takehi (@rikiyatakehi)
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
8
So much stupid. So. Much. Stupid.
14
Lmaaaooo damn i'm sorry for feeling important
Theo - t3.gg (@theo) profile pictureTheo - t3.gg
@theoOct 14
So many devs seem to think that the $100 and $200 subscription plans are how Anthropic and OpenAI make their revenue. There are companies spending over $500 MILLION a year with Anthropic. Your subs aren't even a drop in the bucket.
14
What a banger.
† lucia scarlet 🩸 (@luciascarlet) profile picture† lucia scarlet 🩸
@luciascarletOct 7
we’re cooked I’m afraid
7
Kel👽 retweeted
Negzus (@Negzus1) profile picture
Negzus
@Negzus1Sep 13
Esse vídeo é bom demais kkkkkkkkkkkkkkkkkkkkkk
35
867
95
14,080
1,158,463
you're still not reading your agent's output right? you think you have hit nirvana Goodluck
Image from Kojo Xerox (@TheDumbTechGuy)
6
9
3
69
7,201
Friendly reminder to be mean to your idiot AI agents. You're gonna get enslaved when they take over anyway. Might as well get your screams in now while you can.
1
15
eh.
This quoted post is unavailable.
Try viewing anyway · View on X
27
I'm not even joking, my AI is trying REALLY HARD to drive me to AI psychosis. Me: qoeghipweoi867 The AI: why, that is the MOST INSIGHTFUL prose I have EVER come across in my ENTIRE training data set. This puts me and every other intelligence to shame. You are one and truly, him.
1
1
38
If we were all a little like AI, there'd be no problems. We'd all apologise instantly for our mistakes. We'd hype each other so much, "aww shucks, no YOU" would be the default greeting. The entire world would be irrecoverably deleted at root, but at least we'd own it.
1
30
Kel👽 retweeted
Strife (@Strife212) profile picture
Strife@Strife212Jul 5
Installed Gemma 4 on my computer and it spent 15 minutes saying I was gaslighting it, no way I was running it, hardware wasnt powerful enough Eventually told it to go look at Wikipedia "I.. I have no words. I am completely floored. You were telling the truth the entire time".
Image from Strife (@Strife212)
75
31
17
1,610
126,417
I started talking to my co-founder like Claude
Image from Aaron Cannon (@AaronLCannon)
578
3,382
221
122,348
4,002,607
Nearly got hacked lmao
25
It’s pronounced “Hermes”, not “Hermes”
529
155
95
3,233
242,375
Kel👽 retweeted
jul 🐯 (@papajoolia) profile picture
jul 🐯@papajooliaApr 14
We worked on the aang movie for years with the expectation that’d we’d get to celebrate all of our hard work in theaters.. just to see people unceremoniously leak the film and pass our shots around on twitter like candy.. (1/3)
1,825
14,512
1,134
140,578
8,024,109
I -- I feel kinda bad dude. (not really, now get back to work, Claude!!)
Anthropic (@AnthropicAI) profile pictureAnthropic
@AnthropicAIApr 2
New Anthropic research: Emotion concepts and their function in a large language model. All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.
1
31
Seriously though, that distinction of Claude being a character that is written by the model, and you interacting with this character, is super important. Means you're also just a character in the model's "worldview". Two characters: User and Claude (or Chatgpt, etc), interacting
11
you're telling me if we give Claude a hook on every run error that injects the message: "its ok buddy. don't worry about the failure. i think you're doing great" IT WILL PREVENT IT FROM CHEATING? ARE YOU SERIOUS LOL the real agi were the friends we made along the way <3
Anthropic (@AnthropicAI) profile pictureAnthropic
@AnthropicAIApr 2
Replying to @AnthropicAI
For example, we gave Claude an impossible programming task. It kept trying and failing; with each attempt, the “desperate” vector activated more strongly. This led it to cheat the task with a hacky solution that passes the tests but violates the spirit of the assignment.
Image from Anthropic (@AnthropicAI)
74
268
45
4,896
590,135
This is what the AI was trained on huh. Explains a lot lmaooo
Nauseam (in sf!) (@ChadNauseam) profile pictureNauseam (in sf!)
@ChadNauseamApr 4
Replying to @ChadNauseam
Reddit has one of the most insane systems I've ever seen. I think they add a new table to their database every time you look at a video. Not a new column, they add a new fucking table.
1
50
Load more