Skip to content
Source · Daily Brief

AI Daily Brief — 9 October 2025

Thursday had two stories. Reflection AI raised $2B at $8B post-money — a 15× leap from $545M seven months earlier. Led by NVIDIA (~$800M) and Sequoia, with Disruptive, DST, 1789, B Capital, Lightspeed, GIC, Eric Yuan, Eric Schmidt, Citi, CRV. Founded March 2024 by ex-DeepMind researchers Misha Laskin (Gemini reward modeling) and Ioannis Antonoglou (AlphaGo co-creator), Reflection positions as the Western open-weight counter to DeepSeek/Qwen with a frontier open-weight model targeting tens of trillions of training tokens before year-end 2025. Laskin: “DeepSeek and Qwen and all these models are our wake-up call… the global standard of intelligence will be built by someone else.” Samsung SAIL Montréal’s Tiny Recursive Model paper (TRM, arXiv:2510.04871) swept Thursday coverage — Alexia Jolicoeur-Martineau’s 7M-parameter recursive net hit 45% on ARC-AGI-1 and 8% on ARC-AGI-2, outperforming DeepSeek-R1, Gemini 2.5 Pro and o3-mini while using <0.01% of their parameters. Single 2-layer net looped repeatedly to refine answers; code open-sourced.

Top stories

  • Reflection AI raises $2B at $8B. NVIDIA (~$800M) + Sequoia leading. Disruptive, DST, 1789, B Capital, Lightspeed, GIC, Eric Yuan, Eric Schmidt, Citi, CRV. 15× valuation jump in 7 months. Plans a frontier open-weight model on tens of trillions of tokens before year-end 2025. via TechCrunch
  • NVIDIA’s largest visible bet on a US open-weight lab. Explicitly framed as backing a Western alternative to Chinese open models. Reflection has secured a compute cluster (NVIDIA stack) for training. via AI Business
  • Laskin: “open model frontier back to the US”. X post framing the mission as restoring US leadership in open-weight frontier models for a “thriving AI ecosystem globally.” via X / Misha Laskin
  • TRM: 7M params beats DeepSeek-R1, Gemini 2.5 Pro, o3-mini on ARC-AGI. Samsung SAIL Montréal’s Tiny Recursive Model paper hit mainstream coverage Thursday. 45% on ARC-AGI-1, 8% on ARC-AGI-2, <0.01% of frontier-LLM parameter counts. via MarkTechPost
  • TRM paper + code on arXiv + GitHub. Single 2-layer network looped to refine both internal reasoning state and final answer for self-correction. Trained on ~1000 examples for Sudoku, Maze, ARC-AGI. via arXiv
  • Reflection positions vs OpenAI, Anthropic AND DeepSeek. Three-way framing: open-weight alternative to closed US labs + non-Chinese sovereign-friendly alternative to DeepSeek/Qwen. Enterprises and sovereigns often won’t deploy Chinese models due to legal exposure. via Yahoo Finance
  • Laskin on CNBC: sovereign AI strategy. “Focused on building a frontier open intelligence for a thriving AI ecosystem” — explicitly tying open-weight frontier models to US national-security and economic-competitiveness arguments. via CNBC

Who shipped

Reflection AI shipped a $2B round. Samsung SAIL Montréal shipped the TRM paper and code. OpenAI, Anthropic, Google DeepMind, Meta, xAI, DeepSeek, Alibaba all dark on launches.

Open-source pulse

Reflection’s $2B reshapes the US open-weight conversation overnight. Brooklyn-based team recruited from DeepMind + OpenAI; advanced training stack to be open; initially focused on autonomous coding agents pivoting to general frontier scale. Wilson Sonsini advised — top-tier counsel on a deal at seed-stage age (~19 months old).

Money, infra & hardware

NVIDIA’s $800M in Reflection tightens the circular financing pattern: NVIDIA invests in customers who then buy NVIDIA compute. 15× valuation step in 7 months underscores AI-lab fundraising velocity bending VC norms.

Quiet corners

TRM positioned in technical commentary as evidence recursion + tiny models can replace scale-only paradigms for structured reasoning — performance beats HRM (the previous tiny-network reasoner) and frontier LLMs on hard puzzle benchmarks while remaining trainable on a single GPU.

By the numbers

  • $2B / $8B — Reflection AI round / post-money
  • 15× — valuation jump from $545M in 7 months
  • ~$800M — NVIDIA’s lead check
  • 7M / <0.01% — TRM parameters / share of frontier-LLM size
  • 45% / 8% — TRM on ARC-AGI-1 / ARC-AGI-2
  • Most-mentioned company: Reflection AI
  • Quietest segment: frontier US labs

Compiled by AI Feed’s editor from verified web sources for 9 October 2025.