Skip to content
Source · Daily Brief

AI Daily Brief — 8 November 2025

Saturday belonged to Kimi aftermath. Kimi K2 Thinking two-day review consolidation crystallized: state-of-the-art open-weight on Humanity’s Last Exam (44.9% tool-enhanced beating GPT-5’s 41.7%), BrowseComp (60.2% vs GPT-5 54.9%), matching GPT-5 on GPQA Diamond. Nathan Lambert: “the closest open models have ever been to the closed frontier” — DeepSeek-R1-style moment. Barnacle Goose technical review published, evaluating reasoning + tool-use chains + code-gen — positioning K2 Thinking as one of the best code-generating models ever open-sourced. $4.6M training-cost claim ricocheted through AI media; Moonshot CEO Yang Zhilin would later call the figure “not official.” Practitioners (Simon Willison) reporting it ran in native quant on 2 M3 Ultras at ~15 tok/s using mlx-lm pipeline parallelism — viral for “$4.6M open model running on prosumer hardware.” LessWrong technical breakdown: INT4 native weights, “hundreds of tool calls” per query design, KDA (Kernel Attention Dual Architecture) research roadmap toward K3. First open reasoning model explicitly trained for very-long agentic tool-call chains (200-300 sequential). AI startups raised $2.3B in the Nov 2-8 week, anchored by Crusoe $1.38B Series E ($10B post-money, Valor + Mubadala co-lead, with NVIDIA + Fidelity + Founders Fund) and Sesame $250M Series B (Sequoia + Spark) for conversational-AI smart glasses. Anthropic-Google-Broadcom multi-gigawatt deal continued to dominate analyst commentary into Saturday — largest custom-silicon AI bet yet, counter to OpenAI’s $38B AWS pact.

Top stories

  • K2 Thinking review consolidation. Two-day consensus: open-weight SOTA on HLE (44.9% w/tools vs GPT-5 41.7%), BrowseComp (60.2% vs GPT-5 54.9%), matches GPT-5 on GPQA Diamond. via Interconnects
  • Barnacle Goose technical review. Independent evaluation positioning K2 Thinking as one of the best code-generating models ever open-sourced. via Medium
  • $4.6M training cost ricochets. CNBC source claim spread widely on X/Substack with DeepSeek-V3 comparisons. Moonshot CEO would later (Nov 11 AMA) call figure “not official” — ignores research + failed experiments. via CNBC
  • K2 Thinking top trending on Hugging Face. Heavy traffic into weekend. Willison ran on 2 M3 Ultras at ~15 tok/s using mlx-lm pipeline parallelism — feeding viral “$4.6M model on prosumer hardware” reactions. via Hugging Face
  • LessWrong technical breakdown. INT4 native weights, “hundreds of tool calls” per query design, KDA research roadmap toward K3. First open reasoning model trained for very-long agentic chains (200-300 sequential calls). via LessWrong
  • $2.3B AI funding week ending Nov 8. Crusoe $1.38B Series E at $10B (Valor + Mubadala lead, NVIDIA + Fidelity + Founders Fund); Sesame $250M Series B (Sequoia + Spark) for conversational-AI smart glasses. via AI Funding Tracker
  • Anthropic-Google-Broadcom multi-GW continues echoing. Largest custom-silicon AI bet yet. Counter to OpenAI’s $38B AWS pact. via Anthropic

Who shipped

Nobody Saturday. OpenAI, Anthropic, Google, Meta, xAI all dark on launches. Day was entirely downstream coverage of Moonshot’s Thursday drop.

Open-source pulse

K2 Thinking continues to dominate — viral chart-toppers on HF, deep technical reviews proliferating in researcher circles.

Money, infra & hardware

Weekly funding tracker captures the structural week — Crusoe + Sesame as anchors. No new Saturday rounds.

Quiet corners

OpenAI’s GPT-5.1 (later in November), Google’s Gemini 3 (Nov 18-week), and xAI’s Grok 4.1 were all still ahead — Saturday was pure aftermath, no first-party launches from US frontier labs.

By the numbers

  • 2 — days since K2 Thinking landed, still topping HF trending
  • 44.9% / 60.2% — K2 Thinking HLE w/tools / BrowseComp leads
  • $4.6M / “not official” — K2 Thinking training cost claim / Moonshot caveat
  • $2.3B / Crusoe $1.38B / Sesame $250M — weekly AI funding / anchors
  • 15 tok/s / 2× M3 Ultras — Willison’s prosumer K2 Thinking deployment
  • Most-mentioned company: Moonshot AI
  • Quietest segment: US frontier-lab launches

Compiled by AI Feed’s editor from verified web sources for 8 November 2025.