Tuesday, June 23, 2026

VibeThinker: 3B param model that beats Opus 4.5 on reasoning. 3B-param reasoning model reportedly matching Opus 4.5 using a novel SFT+GRPO training approach. Source

UniverSat: Unified Transformers for Earth Observation. Proposes unified transformers for Earth observation across varying resolutions and sensor modalities. Source

Tapered Language Models. Explores structured width/depth tapering as an alternative to standard transformer scaling. Source

FastMix: Fast Data Mixture Optimization via Gradient Descent. Optimizes pretraining data mixtures via gradient descent for faster, more effective training. Source

Scaling Linear Mode Connectivity to Billion-Parameter Transformers. Demonstrates that linear mode connectivity and weight merging scale to billion-parameter pretrained transformers. Source

Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models. Ultralytics announces YOLO26, a unified end-to-end real-time vision model. Source

Moebius: 0.2B Image Inpainting with 10B-Level Performance. A 0.2B image-inpainting model claiming performance comparable to 10B models, already ported to ONNX/browser. Source

Claude Code Artifacts. Anthropic adds Artifacts to Claude Code: live, interactive, versioned web pages generated from active coding sessions. Source

Oak, Git Alternative Designed for Agents. A Git alternative built for ergonomic use by AI coding agents. Source

Groq Confirms $650M Raise, Restaffs After Nvidia Deal. Groq confirms a $650M funding round and restaffs following Nvidia's failed $20B deal. Source

GLM-5.2: The New Best Open Model. Analysis claiming GLM-5.2 is the current best open-weights model. Source

Nvidia AI Data Center Design Runs Hotter to Use Less Water. Nvidia describes hotter-running AI data center designs that significantly reduce water consumption. Source

SpaceX Inks Compute Deal with Reflection AI. SpaceX signs a compute deal with open-source AI lab Reflection AI, hinting at broader infrastructure ambitions. Source

OpenAI Launches Initiative to Find and Patch Open Source Bugs. OpenAI launches a program to help discover and fix vulnerabilities in open source software. Source

The AI World Is Getting 'Loopy'. Article on the emerging trend of self-looping/refining AI reasoning systems. Source

Report on LLM Agent Communication Protocols. Bookmarked report on LLM agent communication protocols, likely covering MCP/A2A-style standards. Source

GPT-5.5-Cyber vs Mythos 5 Cyber Capabilities. Comparison showing GPT-5.5-Cyber (85.6% CyberGym) outperforming banned Claude Mythos 5 (83.8%), framed as a cyber capability escalation. Source