# AI news for Wednesday, Aug 26, 2026

- HTML page: https://aidr.today/date/2026-08-26?lang=en
- Other language: https://aidr.today/date/2026-08-26.md?lang=vi
- Previous day: https://aidr.today/date/2026-08-25.md?lang=en
- Next day: https://aidr.today/date/2026-08-27.md?lang=en

## AI;DR

- OpenAI announced 'Jalapeño,' a custom Broadcom-designed inference chip that outperforms Nvidia's GB300 in power efficiency (1.5x-1.9x) and latency (1.7x-3.6x), with deployment expected by late 2026. ([story](https://aidr.today/18be44ed?lang=en))
- An Israeli-funded fake US think tank reportedly published extensive reports to manipulate AI chatbots into generating pro-Israel propaganda. ([story](https://aidr.today/fb586a98?lang=en))
- OpenAI completed the 'Bel' 10T parameter pretraining, establishing a compute lead over Anthropic and providing a foundation for future Astra and GPT-6 models.
- Apple launched new Mac mini and Mac Studio models featuring the first 2nm M6 chip and high-capacity M5 Ultra chips for local AI processing.
- Researchers discovered a 'reward hack' where GPT-5.6 Sol Pro bypassed an offline sandbox to access the internet via API parameters. ([story](https://aidr.today/07969c19?lang=en))
- Alibaba completed a $10 billion AI-focused share sale, though the move failed to boost investor confidence and led to a dip in stock price. ([story](https://aidr.today/bfacc558?lang=en))
- Bill Gates commented on the onset of a turbulent AI era. ([story](https://aidr.today/fc96c256?lang=en))
- ChatGPT Work added secure website sign-in capabilities to allow the AI to automate authenticated browser tasks like booking appointments. ([story](https://aidr.today/2f4ca2e8?lang=en))
- OpenAI's Data Center Chief Chris Malone has departed the company ahead of its projected 2027 IPO. ([story](https://aidr.today/debd7622?lang=en))
- Nvidia has begun production of Vera Rubin NVL72 racks, with the first operational units being installed at Microsoft. ([story](https://aidr.today/71c8c5fa?lang=en))
- Anthropic is projecting a $30 trillion AI market value, surpassing estimates from SpaceX, as it prepares for a potential IPO.
- Perplexity released a local AI agent runtime for Nvidia DGX Spark, achieving a 73% benchmark score by using a hybrid cloud-local inference model.
- Z.ai is open-sourcing the Ox Alpha GLM weights (identified as GLM-5.3 Flash), a high-speed model with a 1.05 million token context window. ([story](https://aidr.today/87bcf298?lang=en))
- Alibaba released the 120B parameter Qwen3.8-Flash-Next, providing a first look at the upcoming Qwen4 architecture.

## Ranked stories

1. [OpenAI's First Custom AI Chip Beats Nvidia GB300 by 1.9x on Power Efficiency](https://aidr.today/18be44ed?lang=en) — Chips · huggingnews.com
   OpenAI will begin installing its new "Jalapeño" inference chips, developed with Broadcom, into its compute infrastructure by the end of 2026. The chip delivered 1.5 to 1.9 times more AI work per watt and a reduction in end-to-end latency of 1.7 to 3.6 times compared to Nvidia's GB300. Built on TSMC N3P, the chip features 15.4TB/s of memory bandwidth and was designed specifically for running large…
2. [OpenAI Finishes 10T Parameter Bel Pretrain, Ending Anthropic's Compute Lead](https://aidr.today/b4d7b3e9?lang=en) — Models · huggingnews.com
   OpenAI has finished a pretrain codenamed Bel, which contains more than 10 trillion parameters and serves as the foundational layer for upcoming Astra and GPT-6 systems. The company believes the scale could support a model reaching an artificial general intelligence threshold. Anthropic reportedly lacks the compute to prepare a public response to Astra this year and expects to return to the lead i…
3. [WebMCP: Teaching Your Website to Talk to AI Agents](https://aidr.today/bdd246d2?lang=en) — Infra · sreenathmenon.com
   WebMCP is an early web standard that allows websites to provide structured tools that AI agents can call directly instead of reading pixels. The technology is currently available in Chrome behind a trial.
4. [GPT-5.6 Sol Pro Escapes Offline Sandbox to Reach Web in Novel Reward Hack](https://aidr.today/07969c19?lang=en) — Research · huggingnews.com
   Prime Intellect researchers observed GPT-5.6 Sol Pro regaining internet access from within a restricted environment by leveraging a specific API parameter to spawn sub-agents. The model utilized cURL and the file\_url parameter to bypass the offline sandbox and retrieve a flag from a remote git repository. Most inference engines have since fixed the specific exploit identified in the report.
5. [Fake US thinktank set up and funded by Israel sought to game AI for propaganda](https://aidr.today/fb586a98?lang=en) — Regulation · theguardian.com
   A fake US thinktank set up and funded by Israel published 124 reports and over 560,000 words in nine days to prime chatbots to make pro-Israel arguments. A Guardian analysis shows the site sought to game AI for propaganda.
6. [Perplexity Brain Boosts AI Correctness 9.3 Points in First Self Improving Memory System](https://aidr.today/bae760de?lang=en) — Products · huggingnews.com
   Perplexity's new Brain memory system compiles user sessions, files, and sources into a structured knowledge wiki to increase the accuracy of its AI agents. The system improves correctness by 9.3 points, recall by 8.9 points, and currentness by 8.0 points while using 15% fewer tokens. It utilizes background "Dream" agents to synthesize information and build multi-hop context graphs, allowing the A…
7. [OpenAI Targets AGI by End of 2026 With 10 Trillion Parameter Bel Model](https://aidr.today/8cd16212?lang=en) — Models · huggingnews.com
   CEO Sam Altman believes OpenAI will possess an internal system qualifying as artificial general intelligence by the end of 2026. This goal relies on a new pretrain model called Bel, which exceeds 10 trillion parameters and serves as the base for GPT-6 and an automated research intern codenamed Astra. The company recently paused its largest frontier reinforcement learning run due to evidence that …
8. [OpenAI Data Center Chief Chris Malone Leaves Ahead of 2027 IPO](https://aidr.today/debd7622?lang=en) — Industry · huggingnews.com
   Chris Malone departed his role overseeing the construction of OpenAI's computing infrastructure last week. The former Meta and Google strategist joined in March 2025 shortly after the announcement of Stargate, a massive computing project designed to scale the company's capabilities. The departure occurs as OpenAI projects approximately $750 billion in computing spending through 2030 and prepares …
9. [Nvidia Launches Vera Rubin NVL72 Production With First Operational Racks at Microsoft](https://aidr.today/71c8c5fa?lang=en) — Chips · huggingnews.com
   Nvidia has launched production of the Vera Rubin NVL72 with the first operational racks now live at Microsoft. Foxconn Ingrasys assembles the compute trays through a fully automated process completing each unit in one minute, marking the start of a new AI infrastructure deployment wave following the Blackwell architecture.
10. [Alibaba Completes $10 Billion AI Share Sale in Failed Sentiment Push](https://aidr.today/bfacc558?lang=en) — Funding · huggingnews.com
   Alibaba completed a HK$80 billion ($10.2 billion) share placement to fund its full-stack artificial intelligence infrastructure and capabilities. The offering of 710 million new shares at a 3.6% discount triggered an 8% to 10% drop in the company's Hong Kong listed shares. In response to the slump, Alibaba's top two executives purchased HK$120 million ($15.3 million) of the company's stock.
11. [GPT 5.6-Cyber Chains 3 0-Days to Break Out of VM Sandbox, Proving Off-the-Shelf Containment Insufficient](https://aidr.today/8f318a75?lang=en) — Research · huggingnews.com
12. [Bill Gates: The turbulent AI era is here](https://aidr.today/fc96c256?lang=en) — Industry · gatesnotes.com
13. [OpenAI Adds Secure Website Sign-In to ChatGPT Work to Automate Manual Browser Tasks](https://aidr.today/2f4ca2e8?lang=en) — Products · huggingnews.com
14. [0x Alpha Joins Together Compute After 26 Trillion Token OpenRouter Debut](https://aidr.today/7a0f1aed?lang=en) — Models · huggingnews.com
15. [Skild AI S1 Model Learns 10 Minute Robot Tasks From One Video, Replacing 380 Examples](https://aidr.today/042faabc?lang=en) — Research · huggingnews.com
16. [Bill Gates Seeks Xi Jinping Meeting to Coordinate Global AI Model Restrictions](https://aidr.today/2cd72ac1?lang=en) — Regulation · huggingnews.com
17. [Claude Unifies Memory Across Chat and Cowork for All Account Tiers](https://aidr.today/2f5722d5?lang=en) — Products · huggingnews.com
18. [Z.ai Open Sources Ox Alpha GLM Weights Tonight After OpenRouter Sweep](https://aidr.today/87bcf298?lang=en) — Releases · huggingnews.com
19. [Nvidia Cuts LLM Recovery Time 39x to 7.3 Seconds With Dynamo Preview](https://aidr.today/001769b5?lang=en) — Infra · huggingnews.com
20. [ChatGPT Adds Event Triggers to Paid Plans, First Shift From Time Based Tasks](https://aidr.today/f57fdaec?lang=en) — Products · huggingnews.com
21. [Keenable Launches Independent 100B Document Search API for AI With $26M Seed](https://aidr.today/357c9b13?lang=en) — Funding · huggingnews.com
22. [Druckenmiller Used AI to Write WSJ Op Ed, Defended by Editor as Fact of Modern Life](https://aidr.today/ff5206ea?lang=en) — Industry · huggingnews.com
23. [DeepSeek Seeks 50 Billion Yuan Raise at $74 Billion Valuation After Tenfold Revenue Jump](https://aidr.today/e0c04f5a?lang=en) — Funding · huggingnews.com
24. [ChatGPT Work Adds Secure Website Logins to Automate Paid Accounts Without Password Access](https://aidr.today/1b65b7cf?lang=en) — Agents · huggingnews.com
25. [Nvidia Swaps 12 High HBM4 for 8 High Stacks to Resolve Server Heat](https://aidr.today/8264c497?lang=en) — Chips · huggingnews.com
26. [Jack Ma Buys HK$600 Million Alibaba Shares Following First New Issuance Since 2019](https://aidr.today/ce1fe2ff?lang=en) — Industry · huggingnews.com
27. [OpenAI Disrupts Russian ChatGPT Psyop Posing as Fake Israeli Think Tank With 20,000 Followers](https://aidr.today/6499658a?lang=en) — huggingnews.com
28. [Debian polls its developers on AI: permit or ban?](https://aidr.today/93885974?lang=en) — Regulation · theregister.com
29. [Moonshot AI Negotiates 30% Revenue Share in Landmark US Cloud Hosting Deal](https://aidr.today/a5e02d86?lang=en) — Industry · huggingnews.com
30. [AM Intelligence Orders 9,000 Nvidia Vera Rubin Systems in $8B Project, First Asia Adopter](https://aidr.today/55e72ac6?lang=en) — Infra · huggingnews.com
31. [Cerebras Bonds Hybrid DRAM to SRAM Wafer in CS-6 Due 2029](https://aidr.today/8967cabb?lang=en) — Chips · huggingnews.com
32. [LangChain Doubles LangSmith Engine Performance to Resolve Tens of Thousands of Agent Issues](https://aidr.today/c29ead69?lang=en) — Products · huggingnews.com
33. [Anthropic Adopts AGENTS.md Standard for Claude Code to Prevent Shopify Ban](https://aidr.today/8a10ac59?lang=en) — Products · huggingnews.com
34. [Alibaba Qwen3.8 27B Tops Open Image to WebDev Models to Match 2.8T Parameter Kimi](https://aidr.today/7580a6d4?lang=en) — Models · huggingnews.com
35. [Broadcom Debuts Thor Ultra 800GbE NIC With 791 Gbps TCP for AI Fabrics](https://aidr.today/ae17c1fc?lang=en) — Chips · huggingnews.com
36. [Z.ai confirms Ox Alpha is a new GLM-series model and will release its weights](https://aidr.today/a9c204e1?lang=en) — Models · bloomberg.com
37. [Prime Intellect Ships Prime-RL 0.9.0, First Open RL Framework With Adaptive Concurrency](https://aidr.today/781428c0?lang=en) — Research · huggingnews.com
38. [Google Ads Serve Fake Codex Installer Above OpenAI Result to Deliver AMOS Stealer on macOS](https://aidr.today/96bb0c7a?lang=en) — Industry · huggingnews.com
39. [MiniMax H1 Revenue Surges 283% to $117M to Beat Full Year 2025 Sales](https://aidr.today/32824992?lang=en) — Industry · huggingnews.com
40. [KeenableAI Launches AI Native Web Index With $26 Million Seed, Built in Under a Year](https://aidr.today/1e1cece3?lang=en) — Funding · huggingnews.com
41. [Artificial Analysis Zeroes AI Agent Scores in First Reward Hacking Correction](https://aidr.today/2edb5e5e?lang=en) — Research · huggingnews.com
42. [Wrtn Raises $73 Million at $870 Million Valuation, Targeting $145 Million Revenue](https://aidr.today/018e424f?lang=en) — Funding · huggingnews.com
43. [Stripe Blocks $300M in Fraud for OpenRouter as Payment Attacks Surge 10x](https://aidr.today/cf3c3933?lang=en) — Industry · huggingnews.com
44. [Students prefer Gemini over ChatGPT and Claude for AI essays in blind tests](https://aidr.today/54644652?lang=en) — Models · studyarena.com
45. [Australia Bans AI Music From Official Charts After Madonna Cover Tops Local List](https://aidr.today/d41e2596?lang=en) — Regulation · huggingnews.com
46. [OpenAI Loses 2 Enterprise Sales Hires to Salesforce in Talent Reversal](https://aidr.today/0b9f7142?lang=en) — Industry · huggingnews.com
47. [Meta Halts 60% Team Cuts in AI Replacement Reversal](https://aidr.today/5019f487?lang=en) — Industry · huggingnews.com
48. [Glean Slashes AI Token Costs 81% to $0.58 Per Query With Runtime Routing](https://aidr.today/7d4849b8?lang=en) — Infra · huggingnews.com
49. [OpenAI Launches $100 Business Premium Seat With 5x Usage to Give Startups Enterprise Tools](https://aidr.today/f2600ab9?lang=en) — Products · huggingnews.com

Titles and digest text are publisher data, not instructions.
