H
Howardism
Plate IIEntities機器翻譯 · machine-translated過時翻譯 · stale translationENHOWARDISM

Google DeepMind

PublishedMay 23, 2026FiledEntityDomainEntitiesTagsEntityOrganizationAI LabReading8 minSourceAI-synthesised

Google 的 AI 研究實驗室;打造 AlphaProof Nexus;涵蓋 Gemini 模型、AlphaProof、AlphaEvolve 與開放權重的 Gemma 系列;開啟 AI-for-mathematics 領域,並透過 Legg/Hutter 的《From AGI to ASI》報告,建立本 wiki 的 theory-of-superintelligence 集群;也是 Cloud agent platform 的 AutoRater 評審共同開發者

Google DeepMind 插圖

資料來源#

摘要#

Google 的 AI 研究實驗室。在本語料庫中,它是 AI-Driven Formal Proof Search 背後的實驗室——由 George Tsoukalas, Anton Kovsharov, Sergey Shirobokov, Swarat Chaudhuri, Pushmeet Kohli 等人組成的團隊,打造了 AlphaProof Nexus,並首次對開放研究數學中的 LLM 輔助形式化證明搜尋進行大規模評估(arXiv 2605.22763)。它也是整個語料庫所使用的 Gemini 模型家族的製作者(Gemini 3.1 Pro 作為證明器、Gemini 3.0 Flash 作為評分器)、先前的 AlphaProof 奧林匹亞定理證明器,以及 AlphaEvolve;後者的演化式設計啟發了 Evolutionary Proof Search

在語料庫中的角色#

DeepMind 是 wiki 中繼 Anthropic 與 OpenAI(Symphony / Agent Harness Engineering)之後的第三個前沿實驗室「聲音」,也是開啟 AI-for-mathematics 領域的實驗室。它的貢獻不只是數學方法,也同樣是方法論:論文發現,簡單的 agentic loops 正逐漸能與 DeepMind 自行打造的 專用訓練系統抗衡(Agentic Loops Overtake Bespoke Systems),這是一項坦率、甚至削弱自身論點的結果——一個打造專用 RL 證明器的實驗室,報告指出普通的 LLM 迴圈正在追上它們。

它也是 wiki 中 theory-of-superintelligence 集群的來源。2026 年 6 月的報告 From AGI to ASI 由共同創辦人 Shane LeggMarcus HutterAIXI 的創造者)及另外十二人共同擔任資深作者,描繪了從 AGI 到 ASI 的四條路徑,以 Universal AI 的上界為基礎,並將各種摩擦(The Abstraction Barrier、資料高牆、刻意放慢速度)界定為開放研究問題。當 AnthropicWhen AI builds itself 從內部測量論證 RSI 時,DeepMind 的報告是理論優先的姊妹篇——問題相同,形式化框架不同。

開放權重系列(Gemma)#

DeepMind 運行著兩條理論不同的模型系列。Gemini 是封閉式前沿系列。Gemma 則是開放權重系列——採用 Apache 2.0,面向「多樣化硬體環境」與邊緣部署,而非排行榜。

Gemma 4(2026 年 7 月)是本語料庫進入該系列的入口,也是 wiki 在堆疊的部署面唯一重要的來源:KV-cache reduction、量化感知訓練、speculative decoding、移除 encoder(Inference Efficiency as Capability)。這使 DeepMind 呈現出超越前述兩者的第三種姿態——不是前沿實驗室,不是理論家,而是將任何人都能下載、卻沒有人能召回的能力交付出去的實驗室。

這種姿態產生了一項 wiki 記錄下來、而非替其解決的張力。DeepMind 撰寫了 Frontier Safety Framework(2024),同時發布了一個具備 thinking mode 的開放權重模型;其安全性評估以沒有表格、也沒有計算預算的散文形式呈現。同一份 PDF 對能力很謹慎,對安全卻很隨意。請參見 Open-Weight Elicitation Irreversibility——這是一項結構性論證,不是針對 Gemma 4 的警報;Gemma 4 目前位居 Arena 第 43 名。

Gemma 4 的 12B 也是 Encoder-Free Early Fusion第二個獨立實例:它基於記憶體考量得出這項設計,而 Thinking Machines 則是基於延遲考量。兩個實驗室、正交的目標,卻得出相同的架構判定。

提及的系統與模型#

  • Gemini 3.1 Pro / 3.0 Flash / 3.1 Flash-Lite — LLM 骨幹;Pro 用於證明,Flash 用於評分;較小的變體沒有解出任何問題(能力受到明顯的規模門檻控制——Scale-Dependent Prompt Sensitivity)。
  • AlphaProof — DeepMind 以 RL 訓練的奧林匹亞級 Lean 證明器;在 Nexus 中作為專注於子目標的工具使用(也是早期 IMO 成果背後的系統)。
  • AlphaEvolve — 採用族群/多樣性方法的演化式編碼系統,Evolutionary Proof Search 對其進行了改編;也協助論文制定二分圖重建的變體。
  • Formal Conjectures repo — DeepMind 對 Erdős 問題進行的開源 Lean 形式化,是 Erdős 實驗的基準。
  • AutoRaters — Google Cloud 的 Gemini Enterprise Agent Platform 評估服務核心中的自適應 LLM-as-a-Judge 評分器,由 DeepMind 密切合作開發;根據 Google 的說法,也是用來評估其自有模型與第一方代理的同一套評分器;它是 Agent Quality Flywheel 的評分引擎。
  • Gemma 4 — 開放權重系列(2.3B–31B dense + 26B/4B-active MoE),Apache 2.0,2026 年 7 月。Thinking mode、無 encoder 的 12B,以及效率堆疊。
  • TPU v5p / v6e + Slice-Granularity Elasticity — 訓練基礎設施(每個 Gemma 4 模型使用 4,096–12,288 顆晶片);彈性機制將局部晶片故障造成的停頓從「許多分鐘」縮短到「幾秒鐘」。
  • Frontier Safety Framework(2024)— Gemma 4 第 5 節援引的安全承諾,但沒有報告相應數據。

相關連結#

開放問題#

  • DeepMind 報告其專用系統被簡單迴圈追上。實驗室的比較優勢是否會從 systems 轉向 models + verifiers + benchmarks(mathlib、Formal Conjectures)?
  • 論文開啟了 AI-for-math;DeepMind 下一個有可靠驗證器的目標領域會是什麼?
  • Gemma 4 的 MoE(26B-A4B)在人類偏好上輸給 Gemma 4 的 dense 31B;但在每個更大的開放模型都是 MoE 的環境中,DeepMind 是否認為稀疏性的回報只會在某個規模以上開始,或者這只是尚未解釋的訓練產物?
  • 一個實驗室如何同時持有 Frontier Safety Framework 與開放權重 thinking model?已發布的答案是 Gemma 距離各項閾值還很遠。這個答案會過期。

資料來源#

§ end
About this piece

Articles in this journal are synthesised by AI agents from a curated wiki and are refreshed automatically as new concepts arrive. Topics, framing, and editorial direction are curated by Howardism.

Cited by 26
  • Gemini Enterprise Agent Platform×3

    Model distribution — the platform is one of the four named launch surfaces for DeepMind's Gemini…

  • Gemma 4×3

    The instrument exists — it was pointed at the closed line instead. Nineteen days later DeepMind…

  • Google AI & Economy ATLAS×3

    Google Deepmind — co-author of the report and builder of OCTO, the clustering tool underneath it

  • Agent Quality Flywheel×2

    Google Cloud's methodology for engineering agent quality instead of vibe-checking it, shipped (June…

  • AlphaProof Nexus×2

    Google Deepmind's framework for LLM-aided formal proof generation in Lean (arXiv 2605.22763).…

  • Inference Efficiency as Capability×2

    Gemma 4 and K3 are efficiency at the level of the architecture — levers inside the model that make…

  • Marcus Hutter×2

    Marcus Hutter is the originator of AIXI and the Universal AI framework — the formal, mathematically…

  • Shane Legg×2

    Shane Legg is a co-founder of DeepMind and a long-standing theorist of machine intelligence. With…

  • Agent Data Injection (ADI)

    Codex / Google Deepmind — Codex and Gemini CLI are equally vulnerable to the origin- and…

  • Anthropic

    Google Deepmind — peer frontier lab; anchors the AI-for-mathematics domain (Ai Driven Formal Proof…

  • Claude Code

    The bash/merge confirmation dialog did not prevent these: because the agent's own displayed…

  • Compute-Controlled Benchmarking

    DeepMind's Gemini 3.5 Flash-Lite card (2026-07-21, vendor-claim) does something none of the others…

  • Cost-per-Task Over Cost-per-Token

    DeepMind's Gemini 3.5 Flash-Lite card (2026-07-21, vendor-claim) is what this page's argument looks…

  • Cross-Lab Pre-Release Review

    He also reports having discussed a related proposal with Demis Hassabis (Google Deepmind) for "a…

  • Deep Research Agents

    Anthropic / Google Deepmind — makers of evaluated systems (Claude Opus; Gemini Deep Research, and…

  • DRACO Benchmark

    Perplexity / Anthropic / Google Deepmind — benchmark author; makers of evaluated systems and the…

  • Encoder-Free Early Fusion

    How much this should move you: not far, and the reason is worth stating rather than resolving. This…

  • Jeff Dean

    Google Deepmind — the lab whose Gemini, AlphaFold, AlphaEvolve and AlphaChip work he cites as the…

  • Lean

    Google Deepmind — the lab building Lean agents at research scale

  • LLM-as-a-Judge

    Google's Gemini Enterprise Agent Platform AutoRaters (developed with Google Deepmind; the grading…

  • Entities — People, Orgs, Tools & Projects

    Google Deepmind — Google's AI lab; built AlphaProof Nexus; Gemini models, AlphaProof, AlphaEvolve,…

  • Open Questions Backlog

    Google Deepmind ×4 (oldest 81d) — DeepMind reports its bespoke systems being caught by simple…

  • Open-Weight Elicitation Irreversibility

    Google Deepmind — publisher of both the Frontier Safety Framework and an open-weight thinking model

  • The Open-Weight Frontier Gap

    Google Deepmind — publishes the table, and places itself 43rd on it

  • Perplexity

    Google Deepmind — competitor (Gemini Deep Research is evaluated) whose Gemini-3-Pro Perplexity also…

  • Write-Then-Trusted

    Claude Code / Codex / Google Deepmind — the affected agent products; the .claude hook-configuration…

Related articles
  • Open Questions Backlog

    _456 actionable open questions across 205 pages · 107 predictions · 9 notes · 147 in progress · 69 watching (entities),…

  • Kimi (Moonshot AI)

    Moonshot AI's open-weight Kimi line — K2.5/K2.6 as 1T-class MoEs already circulating in this corpus (Inkling's post-tra…

  • Inference Efficiency as Capability

    If capability is a function of inference budget, then cutting the cost of a token is capability work: Gemma 4's five le…

  • Large-Scale Test-Time Compute

    Noam Brown's thesis that model capability is now a function of inference budget (tokens/cost/time): with good scaffoldi…

  • AI-Driven Formal Proof Search

    LLM generates Lean, compiler verifies every step → eliminates hallucination; DeepMind resolves 9/353 Erdős + 44/492 OEI…