AI Daily — 2026-05-06

English 中文

OpenAI launches MRC: Multipath Reliable Connection for AI training · SpaceXAI Gives Anthropic Acc...

Covering 30 AI news items

🔥 Top Stories

1. OpenAI launches MRC: Multipath Reliable Connection for AI training

OpenAI, together with industry heavyweights, released Multipath Reliable Connection (MRC) to accelerate and stabilize large-scale AI training clusters via the Open Compute Project. The protocol targets reduced GPU idle time and improved resilience in distributed workloads, potentially boosting throughput and lowering costs for large training runs. Source-x

2. SpaceXAI Gives Anthropic Access to Colossus 1 for Claude

SpaceXAI will grant Anthropic access to Colossus 1 to enhance Claude’s capacity, underscoring continued investment in scalable AI infrastructure. Colossus 1 stands among the world’s largest AI supercomputers, signaling ongoing push for larger, more capable models. Source-x

3. Solo Indian Researcher’s ICML Paper Accepted

Kunvar Thaman, a 26-year-old independent researcher, had a paper accepted at ICML with the Reward Hacking Benchmark (RHB), a sandbox for testing AI models on multi-step tasks. The project—funded by a $2.5k grant—highlights growing opportunities for solo researchers to impact AI safety benchmarks. Source-x

📰 Featured

Embodied AI & Robotics

GENE-26.5 Debuts Robotic Brain Toward Human-Level Robotics — GENE-26.5 introduces a robotics-native foundation model centered on a 1:1 human-like robotic hand and a noninvasive data glove, trained via fast simulation to cover language, vision, proprioception, tactile, and action; the goal is autonomous operation with a single model and fixed weights. This marks a potential new paradigm for human-in-the-loop learning and robotics deployment. Source-x

Open Source & Inference

OpenSeeker-v2 Improves Search Agents with Informative Trajectories — The work shows that informative, high-difficulty trajectories can make simple supervised fine-tuning surprisingly effective for frontier LLM agents, potentially reducing reliance on heavy pre-training, CPT, SFT, and RL. The findings are published on HuggingFace. Source-huggingface
GB10 Atlas Open-Source Inference Engine with Fast CUDA — Atlas, built in Rust + CUDA with no PyTorch or Python runtime, reports high inference speeds across Qwen models on a DGX Spark (130 tok/s peak for Qwen3.5-35B; ~87 tok/s for Qwen3-Next-80B-A3B) thanks to hand-tuned CUDA kernels; community contributions are invited. Source-reddit
Qwen 3.6 27B Gains 2.5x Inference with MTP — A Reddit post documents a 2.5x speedup using MTP with speculative decoding and a 262k context on 48GB, including updated quantizations and a GGUF link for drop-in endpoints. This demonstrates faster inference on consumer hardware and broader deployment opportunities. Source-reddit

LLMs, APIs & Developer Tools

Claude Code rate limits doubled; Opus API limits raised — Claude Code’s five-hour rate limits for Pro, Max, and Team plans are doubled, and peak-hours reductions are removed for Pro/Max; Opus API rate limits are substantially increased. This improves access and throughput for Claude Code and Opus APIs, with implications for usage planning. Source-x

AI Safety & Data Quality

GPT-3 Loss Spikes Linked to Noisy Subreddit Data, Cleaned — Analysis indicates that data scraped from a microwave-noise subreddit caused loss spikes; cleaning the data eliminated the spikes, underscoring data quality’s critical role in large-model training. Source-x

Education & AI in Society

ChatGPT Futures Class of 2026: 26 AI-enabled graduates honored — OpenAI honors 26 graduates who spent four years with ChatGPT, applying AI to science, disaster response, data search, language preservation, and sustainable logistics. Their work includes mapping space objects, disaster survivor detection, cataloging galaxy images, preserving endangered languages, and rerouting unsold inventory from landfills. Source-x

⚡ Quick Bites

ChatGPT Adds Excel and Google Sheets Add-On — Adds native spreadsheet integration for easier data tasks. Source-x
ARIS: Open-Source Autonomous Research Harness for LLMs — New open-source harness for autonomous LLM research. Source-huggingface
Beyond SFT-to-RL: Black-Box On-Policy Distillation for Multimodal RL — Introduces a black-box on-policy distillation method for multimodal RL. Source-huggingface
Local Deep Research: Local AI Research Assistant — Local Deep Research releases a local AI research assistant. Source-github
TabPFN: Foundation Model for Tabular Data — TabPFN foundation model tailored for tabular data. Source-github
Speed Up Qwen 3.6-27B with 100K Context on 3090 — Reports speedups on 3090 with 100k context. Source-reddit
RTX 5090 Achieves 200k Context with Qwen3.6 27B in vLLM — Demonstrates very large context handling on high-end GPUs. Source-reddit
Qwen3.6-35B-A3B-UD-Q5_K_XL Shines with VS Code and Copilot — Notable performance with code tooling. Source-reddit
ZAYA1-8B: Frontier Intelligence Density Trained on AMD — Frontier-trained model on AMD hardware. Source-reddit
Code with Claude: Keynotes and Claude Code Updates — Claude Code updates and roadmap notes. Source-x
Google DeepMind Partners with Eve Online on AI in Games — DeepMind collaborates with Eve Online on AI in gaming. Source-x
Cursor 3.3 reveals breakdown of agent context usage — New insights into agent context utilization. Source-x
GitHub Repo Curates 80+ LLM-Powered AI App Projects — A curated list of AI apps. Source-github
Local models with agent harnesses now handle junior IT tasks — Local models tackle junior IT tasks with agent harnesses. Source-reddit
Qwen 3.6-27B Quantizations Compared for VRAM Efficiency — VRAM-efficient quantizations explored. Source-reddit
Analysis of the 100 most popular hardware setups on Hugging Face — Hardware usage survey on HF. Source-reddit
Qwen3.6-35B-A3B MTP Grafted, Mixed Speedups Reported — Mixed speedups with MTP grafting. Source-reddit
Prefill Bottleneck Often Slower Than Token Speed, Reddit Debate — Debate on prefill vs token speed. Source-reddit
What if the EU Built Claude? — Discussion on EU-built Claude-like AI. Source-x
ChatGPT Feels Very Switched On Lately — Observations on ChatGPT’s recent performance. Source-x

Generated by AI News Agent | 2026-05-06

━━━━━━ End of Template ━━━━━━