Starlink Bandwidth Jumps 10x, Grok 4.5 Tops Benchmark, Cosmos 3 Edge Launches

Key Takeaways

1. Starlink satellite bandwidth leaps from 96 Gbps to 1024 Gbps, with the constellation growing from 60 to ~9,600 satellites in under seven years.
2. Starship's next test flight is targeted for July 23, aiming to dramatically reduce AI compute and global communication costs.
3. Grok 4.5 achieves 91.3% on VulcanBench, surpassing Claude Fable 5 and GPT-5.6 Sol while maintaining cost efficiency.
4. NVIDIA releases Cosmos 3 Edge, a 4B-parameter device-side world model for robotics and autonomous driving, supported by Aravind Srinivas.
5. Hugging Face launches VoiceEQ benchmark for voice AI and local.dot.ai for on-device model evaluation across 700+ tasks.
6. Baidu open-sources Unlimited-OCR, capable of processing 40-page documents into structured Markdown locally.

1. SpaceX Advancements in Satellite Internet and Spaceflight

1. Starlink satellite bandwidth has been upgraded from 96 Gbps to 1024 Gbps, a 10x increase. The constellation has grown from 60 satellites to approximately 9,600 in less than seven years, now covering 164 countries with broadband and direct-to-cell phone services. — via 1 2
2. Starship's next flight test is targeted for Thursday, July 23. Elon Musk emphasized that Starship will reduce costs for AI compute and global communication, potentially transforming infrastructure economics. — via 1 2

2. AI Model Performance, Benchmarks, and Capacity Constraints

1. Grok 4.5 scored 91.3% on the VulcanBench programming benchmark, surpassing Claude Fable 5 and GPT-5.6 Sol while maintaining cost efficiency. This highlights Grok's competitive coding capability. — via 1 2
2. Greg Brockman noted that GPT 5.6 Sol shows promising mathematical reasoning, with one user calling its proof "correct and containing interesting ideas," potentially a milestone for AI in math. — via 1
3. Ethan Mollick observed that AI model performance is inconsistent, with some models (e.g., Sonnet 3.5/3.6) exhibiting abnormal capability jumps that later regress to the mean, suggesting underlying unpredictability. — via 1
4. Deedy highlighted a severe AI compute bottleneck using Kimi K3 as an example: Moonshot lacks sufficient compute to scale service, and even US inference providers struggle with 2.8T parameter models. Demand for frontier intelligence continues to grow, benefiting those with locked-in compute. — via 1

3. Edge AI and Open-Source Models

1. NVIDIA released Cosmos 3 Edge, an open world model designed for device-side deployment. It features 4B parameters and a 2B Nemotron reasoner, enabling robot learning, autonomous driving scene understanding, and visual AI reasoning on platforms like DGX Spark and NVIDIA Jetson. Aravind Srinivas also tweeted about the model. — via 1 2
2. Hugging Face announced local.dot.ai, a free platform that benchmarks models on-device across 700+ tasks, covering quantization, MTP settings, and more. It aims to help users choose the right model for their hardware. — via 1
3. Baidu open-sourced Unlimited-OCR, a model that can process a 40-page document in one go, outputting structured Markdown with high accuracy. It runs locally for free. — via 1

4. AI Tools and Evaluation Benchmarks

1. Hugging Face and Hume introduced the Real World VoiceEQ benchmark for voice AI, covering 40+ models, 15 dimensions, and 60+ metrics, built on over 1 million human ratings. This provides a standardized way to assess voice AI quality. — via 1
2. NVIDIA's Agent Toolkit now includes Omniverse libraries, enabling developers to integrate physical AI capabilities into existing 3D applications, shown at SIGGRAPH 2026. — via 1
3. Hugging Face CLI now supports discovering models from Novita directly in the terminal, with filtering and trending sort, simplifying developer workflow. — via 1
4. An AI agent reproduction challenge encourages using agents to replicate ICML papers from code snippets, promoting reproducibility and tool evaluation. — via 1