The Archive

Simon Willison documents a data exfiltration vulnerability in Claude's web_fetch tool that exploits interaction between private memories and URL-based attacks.

SPyCE co-evolves reusable skills and policies for multimodal agents, distilling visual reasoning trajectories to improve generalization across tasks.

Quantum topological data encoding framework encodes topological structure into quantum states via topology-driven quantum evolution for high-dimensional data.

Heavy-Tailed Flow Matching via Random Clocks (HTFM) improves diffusion models for rare-event data by replacing Gaussian sources with clock-conditioned mixtures.

Survey of 410 German companies shows AI adoption in HR mainly drives efficiency gains, not strategic transformation of people-centered functions.

NodeImport addresses class imbalance in graph node classification using node importance assessment instead of class-size prioritization.

Multimodal deep learning framework using Swin-UNETR jointly analyzes 3D CT and clinical data to classify pancreatic cancer resectability per NCCN categories.

Anthropic-backed Ode launches as AI labs bet that embedding forward-deployed engineers inside enterprises is the key to accelerating enterprise AI adoption.

Traffic-Aware Randomized Smoothing (TA-RS) provides certified defense for LLM-based network intrusion detection by noise injection in attacker-controllable feature subspace.

Patient-clustered evaluation of 15,000 chest radiographs using DenseNet-121 and BIO+ClinicalBERT reveals report leakage risks in multimodal medical imaging models.

The app is designed for people who want to create social content, but find traditional video editing tools too complex or time-consuming.

Rime is handling over 100 million calls each month across multiple companies

Geometric Coding Theorem: establishes symmetry priors as universal lower semi-computable semi-measures for fix-retractable symmetry groups.

Unifying framework generalizes Nerode-style regularity characterizations across functions with non-Boolean output domains.

Dark matter search using Neural Spline Flows on CMS Run 2015D data in mono-Z channels, reducing 40 kinematic observables to 37-dim feature vector.

Kaleido: algorithm-hardware co-design for video diffusion transformers exploiting latent space correlations to accelerate compute-bound self-attention.

MxGPS addresses topology overfitting in GNNs for power grid modeling via multiplex graph transformers, improving generalization across unseen grid structures.

Three-stage calibration framework reveals how SFT, RL, and on-policy distillation reshape model confidence before, during, and after chain-of-thought reasoning.

Study of grokking in two-layer networks with holomorphic activations on modular arithmetic reveals algebraic structure limits memorization-to-generalization transitions.

The guy behind TCP/IP is working on a standard for identifying AI agents in the wild.

The startup has reached a $120 million annualized revenue run rate and more than 200,000 paying customers.

OpenAI proposes 'reverse federalism' model where state-level AI regulations inform national safety framework.

Audit of post-hoc attribution methods (Chefer, Grad-CAM) for echocardiography EF estimation finds spatial-temporal mismatch: models localize anatomy but miss critical frames.

Iso-QGNN and Iso-CGNN implement topology-aligned architectures for molecular property prediction with parameter efficiency via shared learnable bond interactions.

Constraint-driven framework for model optimization unifies quantization, pruning, distillation, and PEFT as multi-objective engineering decisions for cloud/edge deployment.

DAGR improves goal-conditioned RL via state-conditioned goal embeddings using multi-scale gated cross-attention to mark remaining action gaps.

Quantum GANs applied to find vulnerabilities in post-quantum cryptographic protocols, exploiting current quantum hardware for weakness detection.

DTW over WavLM embeddings enables text-free self-supervised L2 speech assessment for phonetics, rhythm, and intonation in low-resource settings.

User permission frameworks for AI agents address prompt injection and unauthorized actions; bridges product-level design with enforcement mechanisms.

IBM's earnings miss and mainframe business challenges; limited direct relevance to frontier AI development.