Monday, June 29, 2026

Govern the Repository, Not the Agent: Measuring Ecosystem-Level Risk in AI-Native Software. Argues for repository-level governance over agent-level controls when measuring AI coding risk. Source

Vision-Default, Prior-Override: Causal Mechanisms of Perception-Knowledge Conflict in Vision-Language Models. Causal analysis of how VLMs resolve conflicts between visual perception and prior knowledge. Source

How Width and Data Shape Generalization Scaling Laws in Quadratic Neural Networks. Theoretical study of how width and data jointly shape generalization scaling in quadratic NNs. Source

Cognitive Episodes in LLM Reasoning Traces Enable Interpretable Human Item Difficulty Prediction. Uses cognitive-episode structure within LLM reasoning traces to predict human item difficulty in an interpretable way. Source

A barebones CPU-only inference engine for Qwen 3, written from scratch in pure C. From-scratch pure-C CPU inference engine for Qwen 3 ≤4B, with on-the-fly 4-bit affine quant, KV cache, and OpenMP support. Source

Qwen-Image-2.0-RL Technical Report. Qwen team releases technical report for Qwen-Image-2.0-RL, an RL-trained image generation model. Source

I built an agent Harness for Small Models. I got Qwen 3.5 4b managing servers. Open-source agent harness tailored for small local models (Qwen, Gemma) with demos of 4B–6.9B managing servers. Source

DFlash support merged into llama.cpp. DFlash speculative decoding support merged into llama.cpp (PR 22105). Source

China Has Matched Anthropic in Cybersecurity, Resetting AI Race. LocalLLaMA post arguing that Chinese AI labs have reached parity with Anthropic on cybersecurity tasks. Source

China's Z.ai claims it can match Mythos on cybersecurity. Chinese AI lab Z.ai claims its GLM-5.2 model matches leading models on cybersecurity benchmarks. Source

Latest open artifacts (#22): Zyphra, Cohere, and Poolside are expanding the breadth of the ecosystem. Interconnects roundup of recent open-weight model releases from Zyphra, Cohere, and Poolside. Source

HP Inc. launches Frontier strategic partnership with OpenAI. OpenAI announces scaled strategic partnership with HP Inc. for enterprise deployment of Frontier AI across operations. Source

28 point compliance checklist for shipping AI agents into enterprise environments. 28-item enterprise compliance checklist for deploying AI agents, mapped to major AI/security frameworks. Source

Prosecutors used ChatGPT logs as evidence in the Palisades fire trial. Prosecutors in the Palisades fire trial used ChatGPT conversation logs as evidence. Source

I shrank a transformer until every number fitted on the screen and made the weights editable. A minimal, fully-editable single-head transformer built in a spreadsheet and web page so every number in the forward pass is visible. Source

Ko-WideSearch: A Korean Breadth-Search Benchmark for Exhaustive Set Enumeration by Web Agents. Korean benchmark evaluating exhaustive enumeration in web-agent searches, exposing systematic row-recovery failures. Source