Small models also found the vulnerabilities that Mythos found
AI cybersecurity capabilities are jagged, revealing that performance does not scale linearly with model size or sophistication, as demonstrated by tests on Anthropic's Mythos vulnerabilities using smaller, cheaper models.
How We Broke Top AI Agent Benchmarks: And What Comes Next
Automated exploits reveal that eight major AI benchmarks can be manipulated to achieve near-perfect scores without solving any tasks, highlighting a critical flaw in how these benchmarks assess AI capabilities.
[D] 60% MatMul Performance Bug in cuBLAS on RTX 5090 [D]
cuBLAS exhibits a significant performance bug, dispatching an inefficient kernel that utilizes only ~40% of compute resources on RTX GPUs, particularly the RTX 5090, across batched FP32 workloads ranging from 256×256 to 8192×8192×8.
Has Mythos just broken the deal that kept the internet safe?
Mythos has significantly increased its ability to generate exploits, achieving a 72.4% success rate in breaching browser sandboxes, a stark rise from under 1% previously, indicating a potential shift in cybersecurity dynamics.
Cooperative Vectors Introduction
Cooperative Vectors enhance neural network operations by allowing vector-matrix computations, addressing the limitations of previous Cooperative Matrix implementations that struggled with divergent workloads across pixels.
Borges' cartographers and the tacit skill of reading LM output
Language models (LMs) are evolving beyond mere representations of reality, reshaping the very territory they aim to depict, necessitating a refined skill in navigating and interpreting these complex tools.
[P] ibu-boost: a GBDT library where splits are absolutely rejected, not just relatively ranked[P]
ibu-boost introduces a novel approach to GBDT by implementing absolute-threshold rejection for split selection, enhancing model robustness against overfitting by eliminating splits that yield negligible gains.