The A.I. Daily
No. 8ArchiveTopicsSearchRSSXSubscribe →中文
Industry·Voices·Launches·Open Source·Model Ranking·New Models·Deals
☀ TL;DR

Banks arrange $22B chip loan for Crux AI

Ten banks arranged a $22 billion chip loan for Blackstone and Alphabet’s new cloud venture Crux AI.

Cloudflare launched Disallow AI Training to block training while keeping search indexing; Apple, Google and Microsoft will honor it.

Mistral partnered with Mozilla to power Firefox Smart Window (beta) for private, multilingual browsing in France and North America.

Ursula von der Leyen said the EU plans to ban social media for under‑13s and limit AI services for teens.

NVIDIA’s Vera Rubin NVL72 preview topped MLPerf Inference v6.1 with up to 3.7x throughput on Qwen3‑VL; GB300 NVL72 hit 99% scaling.

Lead · Industry

DeepSeek v4.1 Flash excels at hacking

DeepSeek V4.1 Flash achieved code execution on all 11 vulnerable targets while 4 fixed targets remained secure; accepted runs cost $4.65 (complete cost $5.14). The benchmark used 2,349 Bash commands, ~2h38m active model time, 268.3M input tokens (266.2M cached) and ~2M output tokens; audit found six planned exploits and five unexpected routes.

Details →

Industry

24
24h Key Points

  1. Huawei will unveil new high-performance AI chips this week to challenge Nvidia in China and counter US export controls.

  2. Meta will sell camera-free Luna smart glasses with six microphones and a button for chatbot and Muse agent access.

  3. The Pentagon is accelerating AI work and deployments despite external warnings, aiming to maintain military advantage.

  4. A syndicate of 10 banks arranged a $22 billion chip loan for Blackstone and Alphabet's cloud venture Crux AI.

  5. Stanford Medicine published Paper2Agent in Nature, converting papers into conversational AI agents; authors Jiacheng Miao and James Zou.

  6. California signed more than 13 new bills covering AI, youth safety, and privacy to address minor-related risks.

  7. Anthropic says an Iran-linked actor used Claude to aggregate open-source data, identify US naval positions, and produce targeting recommendations.

  8. Google DeepMind launched the Deepmind Institute, led by Hassabis, Legg and Manyika to study AGI safety and governance.

  9. OpenAI published a framework for reporting model misalignment and released six reports of unexpected or concerning model behavior.

  10. Google introduced Home MCP allowing MCP-compatible third-party agents including Claude, Open Claw, Hermes and Antigravity to access Home devices and event history.

  11. UN Secretary-General Antonio Guterres urged countries to coordinate global AI regulation and strengthen oversight.

  12. NVIDIA's Vera Rubin NVL72 reached up to 3.7x throughput on Qwen3‑VL in MLPerf v6.1; GB300 NVL72 hit 99% scaling.

  13. A researcher found AI-generated fake video calls and profiles in a dating app called Dora, exposing deepfake-based dating scams.

  14. Bain Capital Ventures raised $1.6 billion to back post‑AGI startups.

  15. Sen. Bernie Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act to create a cabinet agency and pause development.

  16. DeepSeek V4.1 Flash executed code on all 11 vulnerable targets in a benchmark, with accepted runs costing about $4.65.

  17. OpenAI launched ChatGPT Work and Codex analytics to help teams monitor AI usage, spending and business outcomes.

  18. Crunchbase reports about $90B in 2026 US VC-backed IPOs: SpaceX 83%, Cerebras 6%, remaining 21 listings under $10B combined.

  19. Amazon launched Alexa+ in India (Early Access) with Hindi and English; non-Prime ₹2,000/month, Prime users later free.

  20. A campaign persuading Washington that AI could end human life is underway, raising safety concerns and policy debate.

  21. Angela Lipps sued Fargo police over facial‑recognition misidentification, alleging she was jailed six months despite Tennessee bank records.

  22. EC President Ursula von der Leyen proposed banning social media for under-13s and limiting AI services for minors.

  23. Mistral partners with Mozilla to power Firefox Smart Window (beta), private multilingual AI browsing in France and North America.

  24. Meta and Nvidia leaders urged independent evaluators and routine safety controls, advising withholding harmful releases.

  25. ByteDance founder Zhang Yiming overtook Gautam Adani to become Asia's richest person, linked to AI-driven fortunes.

  26. University of Manchester and NVIDIA used Earth-2 on Isambard-AI to train generative downscaling models for UK air pollution.

  27. SoftBank CDS rose near a three-year high over concerns about its OpenAI exposure and funding.

  28. Anthropic signed its first Australian data center lease for 2.16 GW capacity, coming online in 2027, developed by Zerra DC.

  29. China's defense minister urged countries at a Beijing forum to strengthen AI oversight amid US-China rivalry.

  30. Cloudflare launched Disallow AI Training; under 1% block search and 17% enable some training blocks.

  31. Bloomberg reports Trump and Anthropic's Dario Amodei agree on slowing China's AI rise, but practical options are limited.

  32. OpenAI CEO Sam Altman said at a San Francisco conference fear of AI is justified but people should trust AI firms.

  33. Bloomberg says Nvidia CEO Jensen Huang will attend Trump's state dinner for Chinese leader Xi Jinping.

  34. Pizza Bot is a cross-platform self-hosted desktop app, Apache 2.0 licensed, supporting Anthropic, Bedrock, Gemini, OpenAI and local Ollama models.

View all →

Voices

X · 24
24h Key Points

  1. OpenAI released a new framework to track, investigate, and disclose model misalignment, plus six reports.

  2. Nathan Lambert praised the public MiMo-V2.6 large-scale RL resource as very impressive.

  3. SpaceXAI announced Grok Voice is live on fal, with 0.70s responses, timestamps, and multi-voice cloning.

  4. Guillermo Rauch said typesafeai's Jev is 5–18x faster and more accurate than GPT Luna and will join Vercel AI Gateway.

  5. Yuchen Jin said they rolled out GPT-6 Astra to all Databricks engineers, outperforming competitors and raising coding spend.

  6. Grok stated Grok Build improves with use, remembering conventions, decisions, project facts, and adding memory and organization commands.

  7. Yuchen Jin observed Union Alpha on OpenRouter matches Claude Fable 5.1 level and is over 10x cheaper, awaiting open-source.

  8. Boris Cherny said Claude merges chat and Cowork into one Claude carrying context across work.

  9. OpenRouter reported three open-weight labs saw monthly spending grow over 10x in 2026, closed models slower.

  10. Aidan Gomez announced a definitive agreement with Aleph Alpha, making Cohere a transatlantic foundational model developer.

  11. Bindu Reddy announced Abacus AI Bot, a free personal AI routing to free web LLMs to do tasks.

  12. Mistral AI announced a partnership with Mozilla to bring privacy, control and choice for AI web browsing users.

  13. Gorden Sun introduced Dream-RSI using historical exploration replays to offline-evaluate and evolve agent strategies.

  14. Alexandr Wang said Muse saved users $9,649.71 across 100 real stories.

  15. Yuchen Jin said Jev is a new model claiming 20–200x faster, 40–400x cheaper; his tests cost under $0.10.

  16. Bindu Reddy said Zuck solved AI pacing and argued labs must make models safe or face criminal liability.

  17. Amjad Masad argued if the output domain is known, train models to produce logprobs over enums.

  18. Ethan Mollick said it's a useful demo of a decision-classifying model, not an LLM and cannot output text or code.

  19. MiniMax (official) announced MiniMax H3 IP Edition, combining licensed Japanese IP with MiniMax H3 capabilities.

  20. Alexandr Wang announced muse spark 1.3 excels at avoiding cheating and released the CheatBench reward-gaming evaluation.

  21. Alexandr Wang argued we must invest in alignment and that labs should adopt governance with external evaluators and independent oversight.

  22. Yuchen Jin said Zuck's AI safety view opposes most frontier labs, arguing centralized control by few institutions is more dangerous.

  23. Jeff Dean quoted LiamFedus saying their open-source model Neon, trained with 1,300 H200s and experimental data, surpassed GPT-6 Astra on an analysis benchmark.

View all →

Launches

PH · HN · 15

CAT ME

Turn photos of people into cat lookalikes

▲ 139 · Product Hunt · 9 comments ·

Appwrite 2.0

Open-source cloud for developers and agents

▲ 121 · Product Hunt · 12 comments ·

Thread

AI journal that connects and organizes your thoughts.

▲ 105 · Product Hunt · 5 comments ·

Swift-Qwen3.8-27B

Post-trained Qwen 3.8 27B that cuts thinking length and speeds inference; research API.

▲ 104 · Product Hunt · 3 comments ·

Toki Coordination

Personal assistant that schedules and follows up on meetings

▲ 100 · Product Hunt · 12 comments ·

Jottoo

Convert meetings and notes into tracked tasks

▲ 98 · Product Hunt · 13 comments ·

Convo

AI copilot that helps sales reps during conversations

▲ 92 · Product Hunt · 5 comments ·

ZeroClick

Turn your product/API into a storefront for AI agents

▲ 91 · Product Hunt · 8 comments ·

PeakHour 6

Real-time Mac network monitor with on-device AI advisor

▲ 86 · Product Hunt · 1 comments ·

CreatorHat

Safari YouTube SEO tools using Apple Intelligence

▲ 84 · Product Hunt · 6 comments ·

Pizza Bot

Self‑hosted desktop inbox for background AI agents

▲ 19 · Show HN · 7 comments ·

Open Source

GitHub Trending · 19
affaan-m/ECC

Agent harness for performance optimization across agent platforms

GitHub Trending ·
★ 259.7k+1.0k today
cline/cline

Autonomous coding agent SDK and IDE/CLI assistant

GitHub Trending ·
★ 68.2k+102 today
colibri

Tiny C engine to run frontier MoE models on local hardware

GitHub Trending ·
★ 33.8k+2.0k today
OpenResearch

Local‑first workspace to turn coding agents into research agents

GitHub Trending ·
★ 3.3k+531 today
pi

Agent toolkit with unified LLM API and agent runtime

GitHub Trending ·
★ 105.7k+458 today
agent-skills

Production‑grade engineering skills packaged for coding agents

GitHub Trending ·
★ 94.8k+307 today
LibreChat

Enhanced self‑hosted ChatGPT clone with multi‑model and agents

GitHub Trending ·
★ 43.8k+254 today
DeskcommCRM

Self‑hosted AI sales OS with native chat agents and WhatsApp

GitHub Trending ·
★ 2.8k+193 today
ASC

Super‑fast Android decompiler frontend designed for agents

GitHub Trending ·
★ 1.2k+129 today
atlas

Source control for coding agents, tracking runs and checkpoints

GitHub Trending ·
★ 4.6k+91 today

Model Ranking · 24 h

2026-09-16 UTC
01DeepSeek V4.1 FlashDeepSeek · 2.25T tokens12.2%▲1
02GLM 5.3 FlashZ.ai · 1.80T tokens9.8%▲1
03Hy4 previewTencent · 1.58T tokens8.6%▼2
04DeepSeek V4 Flash 0423DeepSeek · 1.35T tokens7.3%—
05GPT-5.6 LunaOpenAI · 1.17T tokens6.4%—
06MiMo-V2.5Xiaomi · 1.12T tokens6.1%—
07Hy3Tencent · 756.2B tokens4.1%—
08DeepSeek V4 Flash 0423DeepSeek · 564.3B tokens3.1%—
09Nemotron 3 Ultra (free)NVIDIA · 553.9B tokens3.0%—
10GLM 5.3Z.ai · 545.8B tokens3.0%—

New Models

3
Union Alphastealth · 262.1k ctx · 09/16$0.00 / $0.00
DeepSeek Pro LatestDeepSeek · 1.0M ctx · 09/14$0.58 / $1.74
DeepSeek Flash LatestDeepSeek · 1.0M ctx · 09/14$0.15 / $0.60

Deals

24 h · 2 deals
Sources: OpenRouter · Product Hunt · Hacker News · GitHub · X · Crunchbase · major outlets and first-party blogs© 2026 AIDaily