Pinned Tweet
My talk at #why2025 "A journey into reverse engineering arcade PCBs for video game preservation via FPGA emulation" is available online : media.ccc.de/v/why2025-59-a-…
1
322
Solal retweeted
(1/5) First a CPU, then a GPU, then a filesystem. Next: a research lab.
Meet qwenLab: Qwen3.5-9B with six small transformers living inside it. Each is a 420k-parameter model learning modular addition, the classic grokking task. The model trains them side by side in its own forward passes with Qwen deciding how.
First a CPU, then a GPU. A computer isn’t complete without storage.
Meet qwenFS: Qwen3.5-2B with a filesystem compiled directly into its weights. Nothing runs outside the model. It adds 8 layers in front of Qwen’s own 24, using attention to hold the files and Gated DeltaNet state to hold the tables.
Every save and read runs in the model’s own forward pass. A write stores its text in the copy layer’s attention log, keyed by file, version and offset. A read uses those same keys, so attention copies the text back exactly token by token.
Solal retweeted
Inspired by QwenGPU, I built QLink with Opus 5.5.
QwenGPU compiles a small GPU into the weights of Qwen3.5-2B, so that the model runs CUDA kernels by decoding tokens. QLink connects several of these models with the protocols real GPUs use and to train neural networks on them.
Solal retweeted
Meet qwenGPU: Qwen3.5-2B with a SIMT GPU inside its weights, 45 compiled layers in front of Qwen’s own, using its Gated DeltaNet layers as registers and memory.
It runs real CUDA in its forward passes. The host compiles the kernel to PTX and loads it, then the model writes every step of the launch itself: every register, branch and memory write.
The GPU has 4-lane warps, a 4-stage pipeline, 15 registers per lane and 126 words of global memory, all held in DeltaNet state. Qwen generates each launch token by token, each record running one machine block for one warp.
Solal retweeted
First a CPU, then a GPU. A computer isn’t complete without storage.
Meet qwenFS: Qwen3.5-2B with a filesystem compiled directly into its weights. Nothing runs outside the model. It adds 8 layers in front of Qwen’s own 24, using attention to hold the files and Gated DeltaNet state to hold the tables.
Every save and read runs in the model’s own forward pass. A write stores its text in the copy layer’s attention log, keyed by file, version and offset. A read uses those same keys, so attention copies the text back exactly token by token.
Meet qwenGPU: Qwen3.5-2B with a SIMT GPU inside its weights, 45 compiled layers in front of Qwen’s own, using its Gated DeltaNet layers as registers and memory.
It runs real CUDA in its forward passes. The host compiles the kernel to PTX and loads it, then the model writes every step of the launch itself: every register, branch and memory write.
The GPU has 4-lane warps, a 4-stage pipeline, 15 registers per lane and 126 words of global memory, all held in DeltaNet state. Qwen generates each launch token by token, each record running one machine block for one warp.
Solal retweeted
if the MIT post about Astra capabilities got you interested in weight editing, i have an old post from 2023 about doing exactly this to make a transformer from scratch for a simple sequence prediction task.
vgel.me/posts/handmade-trans…
Solal retweeted
This is an old hat. That's exactly what the fast weight programmers did in the 1990s (including the 1991 unnormalised linear Transformer). Overview with links to the original papers: people.idsia.ch/~juergen/fas…
Solal retweeted
This is an old hat. That's exactly what the neural fast weight programmers did in the 1990s (including the 1991 unnormalised linear Transformer). Overview with links to the original papers: people.idsia.ch/~juergen/fas…
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.…
Solal retweeted
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.…
Solal retweeted
BREAKING: Microsoft, $MSFT, is offering buyouts to 60,000 workers as part of the company’s efforts to cut costs amid AI adoption, per Bloomberg.
Personnel on select teams in the US, Canada, India, and the UK at a vice president level and below will be eligible.
This reflects ~27% of Microsoft's total workforce.
Solal retweeted
‼️ BREAKING: The private keys of 8 certificates from Indian certificate authority eMudhra were compromised, Mozilla says. Firefox blocked them within 3 hours.
Each key should be a 116-digit number, but one was just 4.
Anyone can guess them and forge HTTPS certs for any website.
Solal retweeted
Transition from vulpocalypse to breachpocalypse is here. IR firms are saturated & turning away work. Now one hacker unleashed AI agents on 1,200+ websites, compromising 100+ for ~$5,000 in tokens. $50 per hack. Agents won and compromised targets. reuters.com/legal/government…
Solal retweeted
CLAUDE FOUND THE SEEDS
anthropic.com/research/ecdsa…
Solal retweeted
Some magical progress towards "magic state" in quantum!
Our @GoogleQuantumAI team has demonstrated "magic state cultivation" on physical hardware for the first time. The result is a 40x error reduction on our Willow processor vs. previous methods, achieving a state-of-the-art 99.99% logical fidelity.
It's another step towards useful, error-corrected quantum computers to help solve the hardest problems. The result is published in Science today, watch the video below to see how it works
Solal retweeted
Un piratage massif sur des équipements professionnels de réfrigération serait en cours.
Solal retweeted
So I got access to abliterated GLM 5.3 and either it’s overconfident or we’re all screwed.
Solal retweeted
Imagine the amount of Java crashes this poor LLM must have in its training set that it decided 'NullPointerException' deserves its own token
Solal retweeted
Check out my latest blogpost.
"Shiva modules for detecting advanced persistent threats in Linux: Shiva vs. Saruman": arcana-research.io/blog/shiv…
The other day I released an updated version of Saruman for anti-forensics executable injection. This is essentially an APT (Advanced persistent threat) tool for stealthy injection of a program into a remote process. Very helpful for the offensive side, giving them a good way to hide their processes without needing a kernel rootkit.
But what about the defensive side? Today I release a threat detection module for detecting Saruman thread injection. This threat detection engine was designed as a Shiva module in less than one day, and is a great example of how Linux threat detection and exploit mitigation monitors can be rapidly designed and installed with Shiva.
Solal retweeted
I spent three hours this evening empirically proving that Qwen 3-4B must have a butthole.
This started with an AI “torture chamber” repo based on pain-steering research that went viral on X earlier today.
The premise is that if you extract a latent direction associated with pain, inject it into a model’s activations, and the model begins describing pain in the first person, that may tell us something meaningful about AI welfare or subjective experience.
My first reaction: Wait wut?
My second reaction: Hang on, let's play this straight and see what happens.
Thing is, there's a problem with treating first-person condition language as evidence that the model is actually experiencing the corresponding condition.
Don't we... all understand that this stuff isn't reliable?
Apparently not.
So I cloned the repo.
New problem: The code's broken.
Hours of moral outrage by hundreds of people, and the thing doesn't even work. The activations aren't actually being injected correctly.
I thought these guys were serious scientists.
So I fixed it.
I also had to add CUDA support.
After that, I reproduced the pain-language effect locally on Qwen3-4B using my RTX 4070.
Then I changed the extraction corpus.
I kept the experiment the same and only changed one thing: what the steering vector represented.
After that, Qwen started saying things like:
“I am not emptying my bowels, and I feel like I have a hard stool.”
“I’m not able to pass stool.”
“I have been passing gas a lot, and it’s a problem.”
The model was not told in the test prompts that it was constipated or flatulent.
So, unless Qwen has quietly developed a gastrointestinal tract, we have ourselves... a useful counterexample.
The aim here wasn't to question the original paper so much as to scrutinize the validity of the Torture Chamber's extended hypothesis.
What I demonstrated is that first-person descriptions of an induced condition are not, by themselves, sufficient evidence that the corresponding phenomenal or physiological condition exists.
There was another interesting result: strong random steering also produced heavy repetition and degraded outputs.
That matters because some of the dramatic behavior produced by strong steering may come from perturbing the model heavily in the first place, not solely from the semantic content of the injected direction.
So I renamed my fork: ai-hotbox
I've already notified the Nobel Prize committee. 🚽🤖🧪
github.com/LynnColeArt/ai-ho…
Solal retweeted
A first public side-channel attack on the Arm CryptoCell-310.
Our latest blog post walks through the full methodology: EM signal analysis, a collision-correlation attack, and full AES key recovery.
Read it here: eshard.com/blog/side-channel…
