N3Xxus_6
Re:
on Thursday July 16, 2026 @01:51AM (#48930800)
I'm the co-creator of agentre-bench, an agentic benchmark that measures how well models can reverse engineer malware. Currently it measures only ELF malware but PE will be included in the next update. We also just released a Qwen 3.59b model, https://huggingface.co/AgentreBench/xref-9b This model went through SFT, then reinforcement learning, specifically IPO. Our research has recently been mentioned by several universities as well.
Re:
on Saturday February 14, 2026 @05:28PM (#47019047)
My team and I have worked on an agentic benchmark to see how well agents can perform at reverse engineering threats. The results are pretty interesting, we would love the communities feedback.
Re:
on Sunday February 08, 2026 @02:18AM (#46932062)
Over the last year I’ve been building RacoGrad, a deep learning library written in Racket (a Lisp). It started as an experiment in understanding autodiff and neural nets from first principles, but it’s grown into something surprisingly usable.
Recently, with help of opencalw I added transformer support and finished a working GPT-2 implementation. The model loads pretrained GPT-2 weights and runs end to end inference entirely in Racket.
Fuzzing macOS Userland (For Fun and Pain)
Neural Network Fuzzing macOS Userland (For Fun and Pain)
A Novel "Reasoning"-Enhancing Technique for Large Language Models
A Novel "Reasoning"-Enhancing Technique for Large Language Models
Re:A Novel "Reasoning"-Enhancing Technique for Large Language Models
on Friday June 06, 2025 @11:35PM (#44207224)
Large language models (LLMs) demonstrate impressive capabilities across numerous domains, from programming and technical support to cooking instructions and research enhancement. However, they continue to struggle with precise symbol manipulation in tasks requiring rigorous logical reasoning, mathematical proofs, and complex algebraic operations. As a student of philosophy and computer science, I find this intersection especially fascinating because of the connections between logical precision and conceptual clarity.
SimpleCoder analyzes screenshots, extracts text, and provides instant insights
SimpleCoder analyzes screenshots, extracts text, and provides instant insights
Neuro-Symbolic Logic Enhancement for LLMs
I’m experimenting with a novel approach that integrates symbolic logic directly into a transformer’s attention mechanism. By using a custom spaCy-based logic parser, I generate a “logic mask” that guides the self-attention layers to focus on logical constructs. In preliminary tests with a fine-tuned LLaMA 3 8B model, this method has shown promising improvements on symbolic reasoning tasks...
Novel Logic-Enhanced LLM for Improved Symbolic Reasoning
I’m experimenting with a novel approach that integrates symbolic logic directly into a transformer’s attention mechanism. By using a custom spaCy-based logic parser, I generate a “logic mask” that guides the self-attention layers to focus on logical constructs. In preliminary tests with a fine-tuned LLaMA 3 8B model, this method has shown promising improvements on symbolic reasoning tasks...
RacoGrad: A Racket Autograd
Re:The Deterioration of Google
on Friday November 29, 2024 @11:53PM (#42279556)
They invented the transformer architecture that powers GPT and other llms.
Re:Five or Ten New Proofs of the Pythagorean Theorem
on Thursday October 31, 2024 @08:36PM (#42012964)
This just isn’t true either. There are contemporary sources who talk about him.
Re:Five or Ten New Proofs of the Pythagorean Theorem
on Thursday October 31, 2024 @08:36PM (#42012958)
Where did you come up with this ? This is just not true, that Pythagoras had little interest in math. He had a love of numbers and thought that math was a way to the divine or at least understanding the divine. His philosophy, not religion , but philosophy was a way of life that entangled mathematics profusely.
Re:
on Tuesday August 27, 2024 @04:12PM (#41372432)
Hi all, I created a site that does pattern matching on malware binaries. Let me know what you all think.
Neural Network Malware Scanner
Re:Ask HN: What are you working on (August 2024)?
on Sunday August 25, 2024 @02:11AM (#41344831)
My AI safety startup
Re:
on Friday August 02, 2024 @04:15PM (#41142222)
Synapticyte is my startup dedicated to stopping school and workplace shootings using advanced AI technology. We’ve developed a highly accurate model for identifying firearms and are working on building a robust alert system. This system will enable schools or organizations to notify police, faculty, and staff when weapons are detected. Our model leverages a combination of Convolutional Neural Networks (CNN) and transformers. Although it’s still in the early stages, we believe this is a powerful and important application of AI technology.
Re:
on Thursday August 01, 2024 @12:49PM (#41131076)
Synapticyte is my startup dedicated to stopping school and workplace shootings using advanced AI technology. We’ve developed a highly accurate model for identifying firearms and are working on building a robust alert system. This system will enable schools or organizations to notify police, faculty, and staff when weapons are detected. Our model leverages a combination of Convolutional Neural Networks (CNN) and transformers. Although it’s still in the early stages, we believe this is a powerful and important application of AI technology.
New Racket Deep Learning Library
New Racket Deep Learning Library (Mind)
Re:New Racket Deep Learning Library (Mind)
on Monday June 03, 2024 @12:31PM (#40564329)
MIND is a simple, educational framework for building feedforward neural networks. Includes tensor operations, activation functions, and backpropagation. Ideal for learning core deep learning concepts and exploring Racket. Future updates will be released!
Introducing BlokAv it’s a modern antivirus utilizing blockchain to allow for rapid updates and community consensus on voting and recognizing malicious files. This is still a work in progress. I did this to learn and also to give something to the community. There also aren’t many open source AV options besides ClamAV.
Behavior analysis, signature scanning, sandboxing for Linux, blockchain implementation etc are deployed but still being worked on. A windows update will come in the future. I screwed the code up and just decided to restart it. The signature scan for now requires a hardcore path to a hash list. I’m working on downloading the list from a server. I’d love any feedback etc.
Blok AV
Re:Write your own retro compiler
on Sunday December 10, 2023 @08:51PM (#38596813)
Nils just wanted to say I love your work. I have your C compiler book and one of your lisp compiler ones as well. I’ve learned a lot and your work has helped me appreciate compilers a lot more, so thanks!
Re:Chess.com Says Hans Niemann’s $100M Lawsuit Is a ‘Public Relations ...
on Monday December 05, 2022 @11:52AM (#33868083)
He's not proven guilty because the issue is did he cheat OTB when he played Magnus. The answer is no.
Re:We Built a Meta Pixel Inspector
on Saturday November 26, 2022 @01:47PM (#33755380)
I actually work in an industry that utilizes these a lot. Google, tiktok, meta etc. I implement the code on our customers sites. It's crazy how much data these scripts collect.