Daily AI RoundupsTuesday, 4 August 2026

NVIDIA's Alpamayo 2 Super leads the 4 August edition as an open reasoning model for autonomous vehicles that teams can fine-tune and deploy commercially. Cloudflare introduces programmable wallets for AI agents, while Cursor open-sources its Mixture-of-Kittens training megakernel. The roundup also covers OpenAI's rebuilt GPT-Live voice stack, new EU AI Act enforcement powers, Google DeepMind's DiffusionGemma, Qwen-CUA, Microsoft and UIUC research across 13 million GitHub Copilot sessions, six remote-code-execution flaws in Flowise, inference engineering, xAI infrastructure and AMD's large Instinct GPU agreements with Meta and OpenAI.

• autonomous vehicles
• AI agents
• AI infrastructure
• voice AI
• AI policy
• diffusion language models
• coding agents
• AI security

Stories in this roundup

1. 01
   NVIDIA launches Alpamayo 2 Super, an open reasoning model for autonomous vehicles that teams can fine-tune and deploy commerciallyNVIDIA / X
2. 02
   Cloudflare introduces programmable wallets that let AI agents hold stablecoins, purchase services and receive funds within owner-defined limitsCloudflare / X
3. 03
   Cursor open-sources Mixture-of-Kittens, its deterministic MoE training megakernel for NVL72 systems, reporting up to 2.37× higher throughput than public baselinesCursor / X
4. 04
   OpenAI rebuilt its voice stack for GPT-Live to listen while speaking, supporting deeper reasoning and tool use at scaleOpenAI
5. 05
   White House invites OpenAI, Google and Anthropic to review a completed voluntary frontier-AI oversight frameworkThe Information
6. 06
   European Commission gains powers to inspect AI models, restrict market access and fine providers under the EU AI ActCNBC
7. 07
   Google DeepMind's DiffusionGemma team reports an open-weight diffusion language model generating roughly 1,500 tokens per second on one H100Google DeepMind / arXiv
8. 08
   Qwen-CUA is a 397B-A17B mixture-of-experts agent that controls software using screenshots, keyboard and mouseQwen / arXiv
9. 09
   UIUC's Banruo Liu and Microsoft Azure Research study 13 million GitHub Copilot sessions, finding that agent loops, tool failures and context compaction create distinct infrastructure demandsUIUC and Microsoft / arXiv
10. 10
    LatePost examines Kimi K3's 3T open weights, mixed-attention architecture and implications for AnthropicLatePost
11. 11
    Researchers disclose six remote-code-execution vulnerabilities in the Flowise low-code AI platformElttam
12. 12
    Philip Kiely and Ali Taha of Baseten explain the emerging discipline of inference engineeringLatent Space
13. 13
    Former xAI inference lead Sheng Ying discusses xAI infrastructure, SGLang and open-source systems on Silicon Valley 101Silicon Valley 101 / YouTube
14. 14
    A LessWrong technical essay argues that modern LLMs remain predominantly powered by imitative learning rather than reinforcement learningLessWrong
15. 15
    SemiAnalysis reports that AMD's performance-based warrant agreements with Meta and OpenAI cover up to 6GW of Instinct GPUsSemiAnalysis / X