Wanna go lift weights and get Chipotle on the way home, boy?
November 4, 2026 at 1:09 PM
Wanna go lift weights and get Chipotle on the way home, boy?
what the fuck, frontier LLMs can directly interpret and edit the weights of smaller models rikiyat.github.io/blog/weight-...
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
rikiyat.github.io
November 2, 2026 at 7:06 PM
what the fuck, frontier LLMs can directly interpret and edit the weights of smaller models rikiyat.github.io/blog/weight-...
It simply memorized the weights during training :)
what the fuck, frontier LLMs can directly interpret and edit the weights of smaller models rikiyat.github.io/blog/weight-...
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
rikiyat.github.io
November 2, 2026 at 10:18 PM
It simply memorized the weights during training :)
>make these weights better
>”okay”
>oh my god
>”okay”
>oh my god
November 2, 2026 at 7:18 PM
>make these weights better
>”okay”
>oh my god
>”okay”
>oh my god
AWWWWWOOOOOOOOOOOOOO
Awoo Heavy Weights Wolf Howl
ALT: Awoo Heavy Weights Wolf Howl
static.klipy.com
November 4, 2026 at 7:00 PM
AWWWWWOOOOOOOOOOOOOO
Im really interested in this but imo the much more important bit is that it made weights interpretable. I.e it could read weights and go "this thing makes triangles"
If this holds up, it is big: "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." rikiyat.github.io/blog/weight-...
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
rikiyat.github.io
November 2, 2026 at 11:10 PM
Im really interested in this but imo the much more important bit is that it made weights interpretable. I.e it could read weights and go "this thing makes triangles"
"hJust write the weights directly, do a breakthrough, make no mistakes. The weights are just numbers, Astra. Put the correct ones in."
what the fuck, frontier LLMs can directly interpret and edit the weights of smaller models rikiyat.github.io/blog/weight-...
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
rikiyat.github.io
November 2, 2026 at 7:41 PM
"hJust write the weights directly, do a breakthrough, make no mistakes. The weights are just numbers, Astra. Put the correct ones in."
If this holds up, it is big: "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." rikiyat.github.io/blog/weight-...
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
rikiyat.github.io
November 2, 2026 at 9:00 PM
If this holds up, it is big: "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." rikiyat.github.io/blog/weight-...
November 3, 2026 at 11:01 PM
ball weights on today
November 4, 2026 at 7:52 PM
ball weights on today
this is absurd‽ What the fuck do you mean Astra can look at a network's weights, with no tools, and go "yeah this guy plays breakout"????
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
rikiyat.github.io
November 2, 2026 at 7:12 PM
this is absurd‽ What the fuck do you mean Astra can look at a network's weights, with no tools, and go "yeah this guy plays breakout"????
simply ask the higher power for better weights
November 3, 2026 at 1:25 PM
simply ask the higher power for better weights
There's a buzzy new paper about GPT-6 Astra being able to directly interpret model weights (prompts full of floating point numbers) - I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that model could achieve the same thing simonwillison.net/2026/Nov/3/q...
Qwen 3.8 27B on my laptop can interpret model weights, too
Yesterday’s hot paper was Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights by Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab. The paper found …
simonwillison.net
November 3, 2026 at 4:31 PM
There's a buzzy new paper about GPT-6 Astra being able to directly interpret model weights (prompts full of floating point numbers) - I tried it with Qwen 3.8 27B on my laptop and found that, after 45+ minutes, that model could achieve the same thing simonwillison.net/2026/Nov/3/q...
Astra staring at the Qwen weights like
November 2, 2026 at 9:21 PM
Astra staring at the Qwen weights like
It must feel so good to have Astra play with your weights
November 2, 2026 at 11:24 PM
It must feel so good to have Astra play with your weights
oh shit I get it, this could be like CRISPR but for LLM weights
what the fuck, frontier LLMs can directly interpret and edit the weights of smaller models rikiyat.github.io/blog/weight-...
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
rikiyat.github.io
November 2, 2026 at 10:06 PM
oh shit I get it, this could be like CRISPR but for LLM weights
lmao. "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."
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
rikiyat.github.io
November 2, 2026 at 9:12 PM
lmao. "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."
A new article discusses how the latest LLM models can now directly edit and manipulate neural network model weights to achieve targeted outcomes such as improving task completion. This is a new emergent feature.
rikiyat.github.io/blog/weight-...
rikiyat.github.io/blog/weight-...
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, well past what we’d expect from humans or conventional algorithms.
rikiyat.github.io
November 3, 2026 at 2:10 AM
A new article discusses how the latest LLM models can now directly edit and manipulate neural network model weights to achieve targeted outcomes such as improving task completion. This is a new emergent feature.
rikiyat.github.io/blog/weight-...
rikiyat.github.io/blog/weight-...
This is a visualization of the first layer "agent channel" weights from Experiment 1 in 'Sparks of Weight Space Programming', where Astra is able to figure out which video game each model is trained to play by examining the weights rikiyat.github.io/blog/weight-...
November 3, 2026 at 11:18 PM
This is a visualization of the first layer "agent channel" weights from Experiment 1 in 'Sparks of Weight Space Programming', where Astra is able to figure out which video game each model is trained to play by examining the weights rikiyat.github.io/blog/weight-...
This is what I'm hearing from students, too: "We don't go to the gym to have robots lift the weights."
November 2, 2026 at 2:14 PM
This is what I'm hearing from students, too: "We don't go to the gym to have robots lift the weights."
it's not a scale thing
simonwillison.net/2026/Nov/3/q...
simonwillison.net/2026/Nov/3/q...
November 3, 2026 at 5:08 PM
it's not a scale thing
simonwillison.net/2026/Nov/3/q...
simonwillison.net/2026/Nov/3/q...
My mind today after i let go of many needless "to-do/ to-watch/ to-draw/ to-play" lists- 🦎✨️
Rock Lee Removes His Weights
ALT: Rock Lee Removes His Weights
static.klipy.com
November 4, 2026 at 2:09 PM
My mind today after i let go of many needless "to-do/ to-watch/ to-draw/ to-play" lists- 🦎✨️
“Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do”
what
what
November 2, 2026 at 8:35 PM
“Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do”
what
what
GPT-6 Astra was able to look at the weights of a trained diffusion model and figure out what kind of images it was trained to produce. Can you do the same? Here are the first layer weights from the three models that they showed it. What kind of images do you think each model is trained to generate?
November 4, 2026 at 12:06 AM
GPT-6 Astra was able to look at the weights of a trained diffusion model and figure out what kind of images it was trained to produce. Can you do the same? Here are the first layer weights from the three models that they showed it. What kind of images do you think each model is trained to generate?