Skip to content
Source · Daily Brief

AI Daily Brief — 21 November 2025

Friday closed SC25 + an extraordinary release week. 66th TOP500: El Capitan #1 (1.809 EF/s HPL), Aurora + Frontier. El Capitan also leads HPL-MxP (mixed-precision AI) at 16.7 EF/s. JUPITER Booster at Jülich clocked exactly 1.000 EF/s — 4th exascale worldwide, first non-US. NVIDIA Apollo open AI-physics family debuts at SC25 — EDA/semiconductors, structural mechanics, weather/climate, CFD. Partners: Applied Materials + ASML + Cadence + Dassault + KLA + Lam + Luminary + PhysicsX + Siemens + Synopsys. RIKEN GB200 supercomputers: 1,600 + 540 Blackwell GPUs with Quantum-X800 InfiniBand. Gordon Bell Prize 2025: LLNL + UT Austin + Scripps for “Real-Time Bayesian Inference at Extreme Scale: Digital Twin for Tsunami Early Warning Applied to the Cascadia Subduction Zone” — 50 years of supercomputing collapsed to <1 second on El Capitan. Ai2 ships Olmo 3 — fully open 7B + 32B (Base / Instruct / Think / RL Zero) with entire Dolma 3 (~9.3T-token) corpus, training code, hyperparameters, intermediate checkpoints public. Olmo 3-Think (32B) strongest fully open reasoning model, closes Qwen 3-32B-Thinking gap despite 6× fewer training tokens. Tencent HunyuanVideo-1.5 released — 8.3B DiT + 3D causal VAE; runs on single consumer GPU with 14GB VRAM; Selective Sliding Tile Attention 1.87× speedup vs FlashAttention-3 at 720p/10s; full weights + code + SSTA kernels open. OpenAI + Target launch ChatGPT shopping beta for Thanksgiving week — first-of-its-kind conversational shopping inside ChatGPT: multi-item baskets, fresh-food, Drive Up / Order Pickup / shipping, personalized via Target Circle. Anthropic publishes Claude political even-handedness framework + open-sources eval — 94% even-handedness (vs GPT-5 89%, Llama 4 66%, behind Gemini 2.5 Pro 97%, Grok 4 96%). Coverage continues amid Trump’s “woke AI” procurement ban.

Top stories

  • SC25 closes — TOP500: El Capitan #1 (1.809 EF/s); JUPITER first non-US exascale (1.000 EF/s). El Capitan also #1 HPL-MxP at 16.7 EF/s. via TOP500
  • NVIDIA Apollo open AI-physics model family. EDA + structural mech + weather + CFD. Partners: Applied Materials + ASML + Cadence + Dassault + KLA + Lam + Luminary + PhysicsX + Siemens + Synopsys. via NVIDIA
  • Gordon Bell Prize 2025: real-time tsunami digital-twin on El Capitan. LLNL + UT Austin + Scripps. 50 years collapsed to <1 second. via LLNL
  • Ai2 ships Olmo 3 — fully open 7B + 32B. Base / Instruct / Think / RL Zero. Full Dolma 3 ~9.3T-token corpus + code + hyperparameters + checkpoints public. Olmo 3-Think (32B) strongest fully open reasoning at 6× fewer training tokens than Qwen 3-32B-Thinking. via Ai2 / HPCwire
  • Tencent HunyuanVideo-1.5 — 8.3B DiT on single consumer GPU. 14GB VRAM. SSTA 1.87× speedup vs FlashAttention-3 at 720p/10s. Full weights + code + SSTA kernels open. Open-source counterweight to Sora + Veo. via GitHub
  • OpenAI + Target launch ChatGPT shopping beta. Thanksgiving-week beta — multi-item baskets + fresh-food + Drive Up / Order Pickup / shipping + Target Circle personalization. via OpenAI
  • Anthropic Claude political even-handedness framework. 94% (vs GPT-5 89%, Llama 4 66%; behind Gemini 2.5 Pro 97%, Grok 4 96%). System prompt + character training + open-source paired-prompts eval. via Anthropic
  • Meta opens SAM 3 + SAM 3D (still circulating). Text-promptable concept segmentation ~30ms for 100+ objects on single H200. SAM 3D Objects for scene + Body for human pose/shape from single image. via Meta
  • NVIDIA Q3 FY26 still anchoring AI macro coverage. $57B (+62%), $51.2B data center, GAAP gross 73.4%. Huang: “Blackwell sales off the charts, cloud GPUs sold out… entered the virtuous cycle of AI.” via NVIDIA 8-K
  • EU Digital Omnibus reaction runs through Friday. Defers AI Act high-risk obligations + eases GPAI docs. Bundled with Data Union Strategy + EU Business Wallets. via European Commission

Who shipped

NVIDIA shipped Apollo + RIKEN supers + TOP500 leadership. Ai2 shipped Olmo 3 with full transparency. Tencent shipped HunyuanVideo-1.5 open. OpenAI + Target shipped shopping beta. Anthropic shipped even-handedness eval. Meta shipped SAM 3 / 3D. Google, xAI, DeepSeek, Alibaba quiet on launches.

Open-source pulse

Olmo 3 (fully transparent training trace) + HunyuanVideo-1.5 (consumer-GPU video) + SAM 3/3D (segmentation + 3D from single image) = three structural open-source moments in one week. Chinese + American open-weight families both pushing hard.

Money, infra & hardware

El Capitan + JUPITER reset exascale geography. Apollo opens physics-AI domain to NVIDIA accelerated stack — partners span EDA + manufacturing + weather. NVIDIA Q3 momentum + Blackwell “sold out” continues to drive AI-capex thesis.

Quiet corners

Friday also saw Microsoft’s Adobe Marketing Agent for M365 Copilot ship as part of Agent 365 partner roster.

By the numbers

  • 1.809 / 1.000 / 16.7 EF/s — El Capitan HPL / JUPITER HPL / El Capitan HPL-MxP
  • 10 — NVIDIA Apollo physics-AI launch partners
  • 7B + 32B / 9.3T / 6× — Olmo 3 sizes / Dolma 3 corpus / training-token efficiency vs Qwen 3
  • 8.3B / 14GB / 1.87× — HunyuanVideo-1.5 params / consumer VRAM / SSTA speedup
  • 94% / 89% / 66% / 97% / 96% — political even-handedness Claude / GPT-5 / Llama 4 / Gemini 2.5 Pro / Grok 4
  • Most-mentioned company: NVIDIA
  • Quietest segment: Google + Anthropic launches

Compiled by AI Feed’s editor from verified web sources for 21 November 2025.