
Once the hiders learned what gets caught, they switched to exact symmetries of the network. Each square's three inputs always sum to 1, so you can shift weights without changing a single move. The auditors fell to near chance, but the intended receivers mostly lost the code too.
1
2
53
Then I added AI auditors that didn't know whether a code was there. They mostly caught the codes only when they were told to actively look for one. That cut what slipped through from about 1 in 5 to 1 in 20.
1
1
56
I realized most of the incorrect guesses were real four-letter words (MAZE, DUEL, IONS) instead of random codes. The AIs assumed it was some kind of message. GPT-6 Astra never did this.
1
1
73
LLMs were able to pass hidden messages to each other through neural network weights 78% of the time, with no agreed method, across OpenAI, Anthropic and Google.
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.…
1
5
249
I remade Fall Guys using NVIDIA MotionBricks.
Then trained a neural network with reinforcement learning to find its own path across the vanishing tiles.
2
8
410
I built a tool that turns a sentence into an animated 3D scene. Keep typing to chain new moves, and it all renders in your browser.
NVIDIA's Kimodo model writes the motion from a GPU, and an agent lays out the world around it.
When my agents ran Modal jobs, the log showed nothing about progress until the job finished, so I built an open source fix for it.
3
10
473
Replying to @ItsGoharr
When making it, I broke it down by scene so I could fix issues more directly
3
1,436
I benchmarked benchmarks. It audited 14 popular AI benchmarks and found that about 1 in 5 sampled MBPP and Terminal-Bench items are broken.
2
7
831











