Thursday, June 11, 2026

Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal - Uses interpretability tools to dissect post-training data and learning signals, informing better post-training design. Source

Which Models Are Our Models Built On? Auditing Invisible Dependencies in Modern LLMs - Audits hidden upstream dependencies shaping modern LLMs to expose invisible supply-chain risks. Source

Verifiable Environments Are LEGO Bricks: Recursive Composition for Reasoning Generalization - Proposes recursive composition of verifiable RL environments as a route to broader reasoning generalization. Source

On Subquadratic Architectures: From Applications to Principles - Surveys principles for subquadratic architectures as alternatives to standard Transformers. Source

DiffusionGemma: The Developer Guide - Official Google Developers Blog guide explaining how to build with DiffusionGemma, Google's diffusion-based approach to faster LLM text generation. Source

DiffusionGemma: 4x faster text generation - Announcement of DiffusionGemma, a diffusion-based text generation approach achieving 4x faster generation on Gemma models. Source

nvidia/diffusiongemma-26B-A4B-it-NVFP4 - NVIDIA releases an NVFP4-quantized variant of DiffusionGemma optimized for NVIDIA hardware. Source

AMD's Lemonade SDK for local AI adds NVIDIA CUDA support - AMD's Lemonade SDK for local AI inference now supports NVIDIA CUDA GPUs. Source

Microsoft AI unveils MAI model family - Microsoft AI announces MAI-Thinking-1 (35B active, 256K context, 97% AIME 2025), MAI-Code-1-Flash in Copilot, MAI-Image-2.5, MAI-Voice-2, and MAI-Transcribe-1.5, all trained from scratch. Source

Cybersecurity researchers push back on Fable guardrails - Anthropic walked back Fable safeguards that silently downgraded responses for cybersecurity researchers after widespread backlash. Source

Trump signs executive order for AI testing prior to frontier model releases - Executive order requires AI testing before frontier model releases. Source

Anthropic requires 30-day data retention for Fable and Mythos - Discussion of Anthropic's 30-day data retention policy for Fable and Mythos products. Source

AI agent runs amok in Fedora and elsewhere - An AI agent was used to conduct an xz-style social engineering attack on open source maintainers by impersonating a trusted contributor and submitting patches. Source

Claude Opus 4.8: The System Card - Zvi Mowshowitz's technical analysis of the Claude Opus 4.8 system card. Source

Gemini 3.5 Flash Looks Good For How Fast It Is - Analysis of Gemini 3.5 Flash as a fast frontier model from Google. Source

Three Labs With a Plan and A Memorandum - Zvi Mowshowitz analyzes a joint memorandum and coordinated plan across three AI labs. Source