Tuesday, July 7, 2026

Qwen's J-Space - Anthropic's discovery of an internal model Global Workspace - Anthropic publishes Global Workspace research and releases J-Space lens code, applied to Qwen 3.6 27B. Source

MIRA: Multiplayer Interactive World Models trained on Rocket League - MIRA 5B-parameter multiplayer world model trained on 10k hours of Rocket League data, runs 4 players at 20fps on a single B200. Source

Weak-to-Strong Generalization via Direct On-Policy Distillation - Method for distilling a strong student from a weak supervisor using direct on-policy distillation, aiming to improve weak-to-strong generalization. Source

KempeLab researchers analyze internalization, the process of training models to absorb chain-of-thought computations directly into parameters - The method succeeds on computationally difficult problems that resist direct learning. Source

OpenAI released GPT-Realtime-2.1-mini in the API - Brings reasoning and tool use capabilities to realtime inference. Source

Meituan open-sources LongCat 2.0, a 1.6-trillion-parameter MoE model with 48 billion active parameters - The MIT-licensed model cuts the total number of experts to 128. Source

you can just watch a language model think now - Open-source tool implementing Anthropic's latent 'silent words' paper to visualize subtext in real time on Qwen3.5-4B. Source

Ternlight, 7 MB embedding model that runs in browser (WASM) - 7MB ternary-quantized MiniLM-based embedding model with from-scratch Rust/WASM inference engine for in-browser semantic similarity. Source

ByteDance and Alibaba shut down custom AI companion features ahead of incoming Chinese regulations - The rules target AI services designed to foster emotional attachment. Source

Florian Brand of Prime Intellect claims web search drives most AI hallucinations, prompting calls for agentic benchmarks - Maksym Andriushchenko proposed evaluating models using Claude Code and Codex scaffolds. Source

DeepSeek is developing a custom inference chip to reduce its reliance on Nvidia and Huawei hardware - The company's founder Liang Wenfeng has resumed hiring chip-design engineers. Source

Chinese AI models are gaining ground with U.S. companies as OpenAI, Anthropic costs surge - Report that US companies are increasingly adopting Chinese AI models as OpenAI/Anthropic costs rise. Source

CPU TTS benchmark with UTMOS MOS scoring: Kokoro, Supertonic, Inflect-Nano, and Kyutai's new Pocket TTS - CPU TTS benchmark with UTMOS comparing Kokoro, Supertonic, Inflect-Nano, and Kyutai Pocket TTS. Source

Ant's Robbyant open-sourced its LingBot-Vision family under Apache-2.0 - Highlights permissive licensing vs. Meta's custom-licensed DINOv3. Source

Buried in Anthropic's Fable 5 redeployment announcement is something practitioners should know - Analysis of Anthropic's proposed CVSS-style consensus framework for scoring AI jailbreak severity. Source