AI Wiki — Knowledge Base

Last week some of South Korea's biggest banks were hit by a cyberattack. We now know the entire hack may have been done by a single person. He used a combined stack of an open-source AI penetration tool named ARTEX, DeepSeek v4.1-Flash, GLM-5.3, Grok 4.6, and Claude Code (CrowdStrike)

CrowdStrike report: https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/ From: Andrew Curran on 𝕏: https://x.com/AndrewCurran_/status/2108108695876092

model-releasesr-localllama

8 Oct 2026

model releases

Grok 4.7 ranks first in legal matters

Halogen + Qwen Flash Next keeps getting better

HAN-Mamba: Hierarchical Selective State Space Networks for Multi-Scale Financial Volatility Forecasting

Hard, Yet Reducible: Controlled Forward Transfer for Synthetic Degradation Curation

HarnessIR: Harnessing Multimodal Foundation Models for Universal Real-World Image Restoration

HeiCo-FOCUS: A Clinically Grounded Dataset for Long-Context Video Understanding

Hidden in Plain Sight: Benchmarking Agent Safety Against Decomposition Attacks with DECOMPBENCH

How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression

How Do LLMs Change Predictions Under Negation?

How Fragile Is On-Device Language Model Safety? Localizing Safety-Critical Parameters for Sparse Fault Analysis

How Private is Private? A Comparative Study for Face De-Identification

How to train your model organism

Huge release from @odysseyml. Odyssey-3 Pro sets a new top score on Physics-IQ Verified, a benchmark that asks models to continue videos of …

HySPE: Positional Encoding via Symplectic Dual Shears

I built a local server that speaks the Ollama API - sessions survive crashes, and your chat doesn't wait behind your agents

I oppose the death penalty, but I didn't always feel that way. I understand the people who support it - because I once was one of them. But …

Innovation in Ireland: How Irish brands scale with Gemini Enterprise

Input-Blind Controls Produce Substantial Oracle Headroom for Layer Programs in Multiple-Choice Evaluation

InsClaimBench: Benchmarking Insurance Claim Adjudication Across the Decision Chain

Insights from Autoresearch for Solar Panel Segmentation

InstanceBench: Diagnosing Referential Reasoning and Target Identity in Referring Expression Segmentation

Instead of another GPU terminal renderer, I trained a 1.26M-param model to turn TUIs (htop, vim, emacs…) into real UI components [R]

Inverting Multi-Vector Visual Document Indices

Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning

Itgan at NADI 2026 shared task: Parameter-Efficient Whisper Adaptation for Robust, Mixed-Dialect and Code-Switched Arabic ASR

it's built on top of Claude Managed Agents, doing code generation in the cloud sandbo, using Sonnet 5.5 each run costs ~$0.75 you can try it…

Java vllm-like framework claims 90% of perfomance of llama.cpp on local inference on NVIDIA GPUs by compiling Java to CUDA and cuTile

Jevjitsu: Or, How We Tried Generalized Classifiers on Everything

jevman: AI decision models play Pac-Man

Judging in Latent Space: Efficient Generative Reward Modeling via Semantics-Preserving Compression

Label-free cell counting and viability prediction with brightfield imaging and deep learning

Large-scale Repository Engineering via Agent-Native Reusable Code Primitives

Last week some of South Korea's biggest banks were hit by a cyberattack. We now know the entire hack may have been done by a single person. He used a combined stack of an open-source AI penetration tool named ARTEX, DeepSeek v4.1-Flash, GLM-5.3, Grok 4.6, and Claude Code (CrowdStrike)

Latent Watermarks under Generative Editing: A Benchmark and Analysis of Detection Survival

Learning Cross-Model Activation Alignments with Explicit Many-to-Many Layer Maps

Learning joint probabilistic weather forecasts from station observations alone

Learning Situation-Conditioned Thinking Policies for Long-Term LLM Agents

Learning Transition Kernels of Jump-Diffusion Processes with Conditional Diffusion Models

Learning Unknown Constraints without Unsafe Data via Optimality and Counterfactual Regularization

Lifelong small-object navigation in changing object layouts: a benchmark and method

LiG-DETR: Local-in-Global Reassembly in Latent Space for Aerial Object Detection

LightOn OCR-3 is now available on OpenDocRouter, priced at 0.28 / 1m input and 1.40 / 1m output (it is ~$3.19 / 1k pages on ParseBench). W…

Lightweight and Versatile Learned Optimization by Recombination of Gradient History

LiveMACE: Process-Aware Evaluation of LLM Agent Capabilities in Evolving Markets

LLM4Impact: Integrating Heterogeneous Information for Scientific Impact Prediction

LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches

LRCC: Generalizing Low-Rank Compression with Conditional Computation

MAGIC: Learning from Visibility Asymmetry for Unsupervised Stereo Matching