Vulnerability Researcher and Exploit Developer. LLM whisperer. Former CTO and Director of Research @Crowdfense, ex-@XI_Research.
ntoskrnl.exe
Joined November 2013
- Tweets1.8K
- Following2K
- Followers5.2K
- Likes3.2K
I'll give frontier models two weeks before they'll introduce guardrails preventing direct weights changes
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.…
I'd take 4x DGX Sparks over a MAC Studio any time.
Prefill performance on huge context and long horizon runs alone. CUDA and its mature ecosystem for image/video generation and the flexibility to run multiple LLMs independently split across nodes.
Netflix made us forget about torrents, Hugging Face did the same for model weights.
But, as uncensored and abliterated models are quietly disappearing, we're learning the same lesson again: time to seed.
That might be huge
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/DeepI…
1/4
Looks like even though I just finished setting up my local model pipeline, there's always something new to test. huggingface.co/deepseek-ai/D…
I don't understand the hype/excitement around Nemotron 3.5, when Qwen3.8 Flash Next beats it on all the benchmarks...
While reverse engineering down to 1:1 of the original code has always been possible (despite taking skills and a lot of time), I believe the current advancement with AI will push more and more vendors to ship heavily obfuscated and DRM protected code
Paolo Stagno (VoidSec) retweeted
How to recognize the implants you can find on a Ledger Nano X today?
Learn about the Ledger Nano X display test pads, which one is the most important to watch for, and what today's known implants look like.
youtube.com/watch?v=A-R7Sm0E…
That's crazy
This looks like this could be exactly what I'm investigating. Anyone affected, please contact me or open your ledger and send me pictures so we can confirm if there's a bug in there (pictures: ledger board on top, spy implant below).
nitter.cf/Ledger_Support/status/…
Amazing how offensive work "is for organizations only; individual researchers are not eligible."
Local AI is the way to go. Not going to renew @AnthropicAI.
We’re expanding our Cyber Verification Program to give security professionals broader access to our most capable models.
Through this program, verified security professionals can access Claude Mythos 5.1, Opus 5.5, and Sonnet 5.5 with safeguards designed for defensive work.
We’re also opening up new tiers to allow for authorized offensive work, like penetration testing and red-teaming.
anthropic.com/news/cyber-ver…
So NVIDIA announces 64GB DGX Spark configuration at the original price of the 128GB ones; makes perfect sense to me blogs.nvidia.com/blog/local-…
Let me reframe it correctly: 5 months ago they told us about Mythos and how it was too dangerous for the public. Now that they've built fear around their own model, they are now nudging governments to step in and kill open-weight competitors. anthropic.com/research/glm-5…
Looks like 50% less usage on OpenAI's $200 plan starting tomorrow. Better run deepseek locally or via open router then...
