Had Gpt 6.1 Sol and Claude 5.5 Opus play tower defense against each other, and Sol wiped the floor with Opus. Not really what I thought would happen. (1/n)
1
3
257
Tried to get Opus 5.5 (and Astra) to hand-write changes to Qwen 3 8B's weights to get it to paint with code, but everything sucked! It managed to fix the issues related to syntax, that is, Qwen was writing lots of p5 code that never resulted in an image, but it couldn't impart its own ability of painting via code just by "hand-editing" the neurons.
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…
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.…
4
3
422
Tried this out by asking Astra and Opus 5.5 to hand construct networks to play a custom game another agent cooked up. They can in fact make the networks, but is that just because the networks are small and these bigger models were post trained on these tasks? Opus did a better job than Astra.
Also did an o.k. job at guessing what the network was for, when given just weights with no context.
Proper write up/thread coming after I read through the Claude code session that produced this lol
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…
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.…
4
7
422
Perplexity
The ranking members of the House Judiciary and Oversight Committees have written a letter to President Trump asking whether he has been soliciting advice from Grok in order to help carry out his official duties as President of the United States.
They also demanded that he preserve any prompts sent to Grok, ChatGPT, Claude, Gemini, and Perplexity, as well as outputs in response to those prompts, that were 'submitted to AI chatbots in carrying out your constitutional, statutory, or other official or ceremonial duties.'
The letter was provided by Politico.
2
193
“You know what I’d really like?”
“Tell me.”
“A Side Project”
“We can do a Side Project”
“A real Side Project, no slop write up, no agent spam.”
“I’ll make you a very traditional Side Project. Hand-crafted, passion infused, joy-bringing.”
4
15
741
Took some of the ~spooky~ neurons in Qwen that @jfonsecarivera found and had Claude make a Ouija board with some fun personalities moving around!
top 10 on (1) of these kernels! wonder if it'll stay up there for a bit, but time to move on to the other ones.
We're launching a $100,000 challenge to see who can make Muse Spark run fastest: openkernelchallenge.dev/
1
5
761
