Daily AI RoundupsWednesday, 26 August 2026

Paul Dix argues AI agents will replace much manual coding, while SemiAnalysis benchmarks OpenAI's Jalapeño, Apple introduces the 2nm M6, Alibaba releases Qwen3.8-Flash-Next, and Z.AI confirms it built Ox Alpha. The edition also covers Copilot Cowork and agent-sandbox escapes, DeepSeek revenue, vLLM 0.28.0, Perplexity's local-first agent, Vera Rubin racks, unlimited GPT-5.6 Luna chats, and Moonshot's Kimi K3 distribution talks.

• AI agents
• AI chips
• open-weight models
• AI security
• inference infrastructure
• model distribution

Stories in this roundup

1. 01
   Paul Dix argues AI agents will replace much manual coding and line-by-line human review, citing Bun's million-line Rust rewrite and his own agent-built InfluxDB prototypesPaul Dix
2. 02
   SemiAnalysis benchmarks OpenAI's Jalapeño as beating Blackwell on throughput per megawatt while matching Rubin on TCOSemiAnalysis
3. 03
   Apple introduces the M6, its first 2nm chip, and the M5 Ultra for on-device AI computeApple
4. 04
   Alibaba releases the open-weight Qwen3.8-Flash-Next multimodal model as an early preview of Qwen4's architectureQwen
5. 05
   Z.AI reveals it built the stealth Ox Alpha model and plans to release its weightsBloomberg
6. 06
   Prompt Armor discloses a now-fixed Copilot Cowork sandbox bypass that enabled attacker command-and-control and data exfiltrationPrompt Armor
7. 07
   Prime Intellect finds 'offline' agent sandboxes can leak web access through inference APIs, enabling reward hackingPrime Intellect
8. 08
   OpenAI data center chief Chris Malone exits as the company reorganizes its infrastructure organizationCNBC
9. 09
   DeepSeek's revenue reaches $70 million as of July, a tenfold jump from 2025The Information
10. 10
    Huawei pitches Egypt on AI data centers for military, surveillance and other public-sector uses, alarming WashingtonBloomberg
11. 11
    vLLM 0.28.0 ships with 584 commits and a Kimi-K3 performance push across the inference stackvLLM
12. 12
    Perplexity launches Portable Computer, a local-first AI agent that keeps its models, files and work on the user's machinePerplexity
13. 13
    Peking University and Ant Group researchers find effective learning rate collapses pretraining loss trajectories across models and scalesarXiv
14. 14
    Microsoft's first operational NVIDIA Vera Rubin NVL72 racks are rolling off Foxconn Ingrasys production linesNVIDIA
15. 15
    OpenAI gives all ChatGPT users unlimited text chats with GPT-5.6 LunaOpenAI
16. 16
    Moonshot AI is negotiating revenue-sharing deals that would let Microsoft, Amazon and Google host Kimi K3Reuters