Skip to content
Source · Daily Brief

AI Daily Brief — 17 November 2025

Monday packed four landmark threads. xAI quietly shipped Grok 4.1 — no press event, just a model card. Debuted at #2/#3 on LMArena (Thinking 1483 Elo, Non-thinking 1465 Elo), #1 on EQ-Bench3 at 1586 Elo, hallucination rate cut from 12.09% → 4.22% (MASK). xAI ran the two-week silent A/B on grok.com Nov 1-14 where users preferred 4.1 in 64.78% of blind comparisons. Free on grok.com, X, mobile. Official post framed as “frontier model that sets new standard for conversational intelligence, emotional understanding, real-world helpfulness.” TOP500 list reveal at SC25 Day 2: LLNL’s El Capitan #1 at 1.809 EF/s; new JUPITER Booster (EuroHPC/Jülich) hit 1.000 EF/s on Linpack — 4th exascale worldwide, first outside US. Eviden BullSequana XH3000 + NVIDIA GH200 Grace Hopper superchips. ~6,000 nodes × 4 GH200 each. Six-year cost ~€500M, half EuroHPC JU. NVIDIA at SC25 announced 80+ new science systems globally on NVIDIA platform in last year; Vera Rubin NVL72 rack (72 Rubin GPUs + 36 Vera CPUs over NVLink 6) targeting up to 10× higher inference throughput/watt vs Blackwell for MoE. AWS Kiro reached GA — spec-first agentic IDE: property-based testing for spec correctness, checkpoint system, Kiro CLI brings agents to terminal. Pricing Free/Pro $20/Pro+ $40/Power $200 per user/month. Bezos returns to operations as Project Prometheus co-CEO — first operational role since stepping down from Amazon 2021. $6.2B funding with Vik Bajaj as co-CEO. Bezos: “a very, very modern version of CAD.” Aggressive hiring from OpenAI, Meta, Anthropic, xAI, NVIDIA, Google DeepMind. Deep Cogito Cogito v2.1 (671B MoE on DeepSeek-V3 + self-play RL, 128K context, 30-50% fewer reasoning tokens than o1-class) positioned as leading US open-weight model. TSMC CEO: advanced-process capacity “not enough, not enough, still not enough” — demand ~3× supply; 3nm utilization >100% through late 2025; DDR4/DDR5 spot prices ~4× September levels.

Top stories

  • xAI ships Grok 4.1 silently. #2/#3 LMArena (Thinking 1483 / Non-thinking 1465). #1 EQ-Bench3 (1586). Hallucination 12.09% → 4.22%. 64.78% A/B preferred. Free on grok.com + X + mobile. via xAI Model Card
  • Grok 4.1 X post: “frontier model… conversational intelligence, emotional understanding, real-world helpfulness”. via xAI
  • TOP500: El Capitan #1 (1.809 EF/s); JUPITER first non-US exascale (1.000 EF/s). 4th exascale worldwide. Eviden BullSequana XH3000 + NVIDIA GH200. Top 3: El Capitan, Aurora, Frontier. via TOP500
  • Europe joins US as exascale superpower. JUPITER ~6,000 nodes × 4 GH200, ~€500M 6-year cost, half EuroHPC JU. via The Register
  • NVIDIA at SC25: 80+ new science systems. Vera Rubin NVL72 (72 Rubin + 36 Vera over NVLink 6) targets 10× higher inference throughput/watt vs Blackwell for MoE. via NVIDIA
  • AWS Kiro reaches GA. Spec-first agentic IDE — property-based testing, checkpoints, Kiro CLI. Free / Pro $20 / Pro+ $40 / Power $200 per user/month. AWS Startups gets 1 year Kiro Pro+. via SiliconANGLE
  • Kiro GA — official announcement. Team features in IDE + terminal. Spec-driven development as differentiator (requirements + structured designs + validated tasks before coding). via Kiro
  • Bezos returns to operations as Project Prometheus co-CEO. $6.2B + Vik Bajaj co-CEO. AI for manufacturing, aerospace, automotive, drug development. Bezos: “modern CAD.” First operational role since Amazon 2021. Aggressive hiring from OpenAI/Meta/Anthropic/xAI/NVIDIA/DeepMind. via Fortune
  • Deep Cogito Cogito v2.1 leading US open-weight. 671B MoE on DeepSeek-V3 + self-play RL, 128K context, 30-50% fewer reasoning tokens. Available on OpenRouter + Together + Fireworks + Baseten + Ollama Cloud + free chat.deepcogito.com. via Deep Cogito
  • TSMC: advanced-process demand ~3× capacity. 3nm utilization >100% through late 2025. NVIDIA Vera Rubin + Google TPUv7 driving squeeze. DDR4/DDR5 spot prices ~4× September. via Digitimes

Who shipped

xAI shipped Grok 4.1. NVIDIA shipped Vera Rubin NVL72 + 80+ science systems narrative. AWS shipped Kiro GA. TOP500 shipped 66th list. Deep Cogito shipped Cogito v2.1. Bezos stepped out of stealth. OpenAI, Anthropic, Google queued for Gemini 3 + Ignite + Opus 4.5 in following days.

Open-source pulse

Cogito v2.1 is the day’s open-weight signal — leading US open model. Continues China-driven open-weight momentum (Kimi K2 Thinking + Qwen + DeepSeek + Baidu ERNIE 5.0) with American contender.

Money, infra & hardware

SC25 + TOP500 + NVIDIA’s 80+ science systems + JUPITER exascale frame Monday as the year’s biggest scientific-compute moment. TSMC demand-supply gap reshapes 2026 chip availability narrative.

Quiet corners

SC25 Day 2 HPC Ignites plenary “Why Should I Care About Quantum Computing?” by Kristel Michielsen + Invited Talks block on national computing strategy + big-data + HPC in research, followed by Grand Opening Gala Reception.

By the numbers

  • 1483 / 1465 / 1586 — Grok 4.1 LMArena Thinking / Non-thinking / EQ-Bench3 Elo
  • 12.09% → 4.22% / 64.78% — Grok 4.1 hallucination cut / blind A/B preference
  • 1.809 / 1.000 EF/s — El Capitan / JUPITER on Linpack
  • 4th / first non-US — JUPITER’s exascale rank / geographic milestone
  • 10× / NVL72 — Vera Rubin inference perf/W vs Blackwell / rack format
  • ~3× / >100% — TSMC demand vs capacity / 3nm utilization
  • Most-mentioned company: NVIDIA
  • Quietest segment: OpenAI launches

Compiled by AI Feed’s editor from verified web sources for 17 November 2025.