Tuesday, June 16, 2026

Nemotron 3 Ultra: Open, Efficient MoE Hybrid Mamba-Transformer Model. NVIDIA releases open MoE hybrid Mamba-Transformer architecture targeting agentic reasoning efficiency. Source

Bayesian Inference and Decision Audits for Public Archives of Frontier AI Evaluations. Framework to audit reporting practices and missingness in public frontier AI evaluation archives. Source

Who Flips? Self- and Cross-Model Counterarguments Reveal Answer Instability in LLMs. LLMs frequently flip answers when challenged by counterarguments, including their own. Source

Artificial Intelligence Index Report 2026. Stanford HAI's annual synthesis of AI progress, adoption, and policy. Source

Tangram: Non-Uniform KV Cache Compression for Multi-turn LLM Serving. Enables non-uniform KV cache compression to improve multi-turn inference efficiency. Source

VibeThinker-3B: Verifiable Reasoning in Small Language Models. WeiboAI's 3B model explores strong verifiable reasoning at small scale. Source

ollama/ollama. Adds support for Kimi-K2.6, GLM-5.1, DeepSeek, gpt-oss, and Qwen models. Source

FeynRL: Open Training Framework for RL Post-Training. Open framework for RL post-training of LLMs, VLMs, and agents. Source

Inside the fight over Claude Mythos 5. Report on the political and export-control battle surrounding Anthropic's Claude Mythos 5. Source

The US government's Anthropic models ban was never about an AI jailbreak. Analysis argues the ban targets capability restrictions rather than jailbreak prevention. Source

Cybersecurity vets protest 'dangerous' US government ban on Anthropic's most powerful models. Security researchers warn the ban is counterproductive to safety research. Source

NewCore raises $66M to give AI agents identities. Identity infrastructure startup emerges as AI agents become "employees." Source

Cleo: Fitting Full Analyst Behavior in a 2B Model. Open 2B Qwen3.5 finetune with integrated harness for analyst-style text-to-SQL. Source

My Homelab AI Dev Platform. Homelab agentic dev platform built on Forgejo, Argo workflows, and SPIFFE-attested agent identity. Source

The 90-year-old idea behind JEPA models: Canonical Correlation Analysis. Explains how CCA underpins JEPA embedding-prediction models. Source

How the brains learn. Argues error-driven predictive learning in corticothalamic circuits is the cortex's core learning algorithm. Source