AI Enterprise Products, Open Models, and Hardware Milestones

Key Takeaways

1. OpenAI launches ChatGPT Work for enterprise with $200 trial credit, and GPT-Live voice model goes global.
2. Elon Musk announces Grok 4.6 (1.5T params) on Aug 7 and Grok 4.7 (2.1T params) weeks later; joins Open Secure AI Alliance.
3. Perplexity introduces "Personal Computer" turning Windows PCs into AI agents.
4. China achieves mass production of immersion DUV lithography for 7nm chips, a milestone for domestic AI hardware supply chain.
5. Strong open-weight models emerge: Kimi K3 (2.8T MoE, 1M context) and Microsoft's Mage-VL 4B.
6. NVIDIA Cosmos model downloads surpass 10M on Hugging Face; Runway teases Seedance 2.5.

1. Enterprise AI and Product Launches

1. OpenAI has launched ChatGPT Work for enterprise, offering up to $200 in credits for each team member's first trial of Work, valid for 14 days, for enterprise customers who sign up by August 21. — via 1 2
2. GPT-Live, a voice model, is now rolling out globally in ChatGPT for Education, Business, and Enterprise plans. — via 1
3. Perplexity CEO Aravind Srinivas announced the "Personal Computer" feature for Windows, which turns PCs into AI agents that orchestrate local files, apps, and web-based agents for research, coding, browsing, and building. He highlights the unique feeling of owning local AI hardware. — via 1 2
4. Runway is about to launch Seedance 2.5, its next-generation video generation model. — via 1

2. Open Source AI Models and Ecosystem

1. Elon Musk revealed that Grok 4.6, a 1.5T parameter model with significantly improved SFT and RL, will be released around August 7, followed weeks later by Grok 4.7 (2.1T params) with better performance but slower service and higher token efficiency. He also emphasized AI safety and joined the Open Secure AI Alliance as a founding member, advocating open-source tools for defense. — via 1 2
2. Aravind Srinivas recounted Hugging Face's security incident where closed-source tools failed, and they used the open GLM 5.2 to control the situation; Perplexity joined OSAA to support open-source safety tools. — via 1
3. Ethan Mollick highlighted the release of Kimi K3, the strongest open-weight AI model with 2.8T MoE, native vision understanding, and 1M context window. The new architecture improves intelligence per compute by 2.5x, and open-source high-performance attention kernels and MoE communication libraries are available. — via 1
4. Hugging Face noted Microsoft's release of Mage-VL 4B, a codec-native streaming vision-language model capable of describing real-time events. — via 1
5. NVIDIA Cosmos models surpassed 10 million downloads on Hugging Face, used for building physical AI systems like robotics and autonomous vehicles. — via 1 2

3. AI Hardware and Infrastructure

1. China has achieved mass production of immersion DUV lithography machines (for 7nm chips) through Yuliangsheng, affiliated with Huawei and SiCarrier, with first deliveries to SMIC this year. While still 10-15 years behind ASML's frontier and reliant on Japanese parts, this is a critical milestone for China's self-reliant AI supply chain, potentially threatening global compute balance long-term. — via 1
2. Elon Musk showcased SpaceX's Starship entering space on Flight 13, and NASA Administrator Jared Isaacman praised the achievement, suggesting orbital data centers could become reality. — via 1