読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
← 読解に戻る
ニューストピック · 梁文锋底层技术路线重大3 分で読めるDisconfirmAI 編集部

Why DeepSeek's Liang Wenfeng Rarely Signs Flagship Model Reports, Only Foundational Technique Papers

Lei Feng Net author Gao Yunyi surveyed the 11 papers Liang Wenfeng signed over the past three years, finding he consistently appears on foundational technique papers—MoE routing, attention mechanisms, network topology, inference operators, and sandbox infrastructure—while rarely appearing on flagship model reports that carry over a hundred author names. This signing pattern points to a coherent technical route: attacking cost and efficiency bottlenecks at the architectural substrate, rather than chasing model scale.

2026-09-25 18:48 GMT+8

Over the past three years, DeepSeek has released multiple flagship models, each technical report carrying over a hundred author names. But look at the author lists and you find an interesting pattern: Liang Wenfeng's name rarely appears on V3.2, V4, or V4.1-Flash reports. Instead, he consistently appears on a different class of papers—MLA attention mechanism, MoE routing strategy, mHC network topology, DSpark inference scheduling, and DSec sandbox infrastructure. These are the atomic technical units that constitute model capability, not the models themselves.

Lei Feng Net author Gao Yunyi compiled this observation into an article, calling it a technical sniping route, and counted the 11 foundational papers Liang Wenfeng signed. The signing pattern itself is verifiable: check the author lists on arXiv. What it reveals is that Liang Wenfeng's personal focus is not training a larger model but finding the deepest bottleneck that makes a model cheaper or stronger.

Take MLA (Multi-head Latent Attention) as an example. When large models process long text, they must cache massive key-value pairs (KV Cache), which directly determines memory usage and speed during inference. MLA's approach is to compress this cache with low-rank vectors, reducing KV Cache volume by approximately 93.3%. This is not optimization at the model level but a redesign of the data structure for one specific component within the attention mechanism. Similarly, MoE routing solves the scheduling problem of which experts each token should invoke; NSA solves the memory-access problem of how to align sparse computation with GPU hardware circuits; mHC solves the topological constraint problem of whether numerical stability will collapse during very deep network training.

ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

These papers share a common characteristic: they attack foundational bottlenecks that the industry acknowledges but has not effectively solved, and their solutions land at the architecture or mathematics level, not the data-volume or compute-investment level. Gao Yunyi cites a comparison in the article: OpenAI's GPT-4-class model training costs hundreds of millions of dollars, while the DeepSeek-V3 team reported completing training on 2,048 H800 GPUs for approximately $5.6 million. The article also relays SemiAnalysis's caveat—that the $5.6 million is merely the direct electricity bill for pure GPU training, excluding R&D losses, hardware depreciation, and total infrastructure investment. But even with the corrected scope, the order-of-magnitude difference still holds, and its source is precisely the foundational architectural choices described above.

However, what the signing pattern can prove requires caution. Liang Wenfeng's choice to sign foundational papers rather than model reports may reflect personal research preference, or it may be a natural distribution of authorship contribution—foundational technique papers have fewer authors, model reports have more. Gao Yunyi interprets it as a conscious sniping strategy, which is a reasonable inference but not a proof. What actually supports this inference is that these foundational choices repeatedly hit industry bottlenecks and produced verifiable efficiency gains, not the signing behavior itself.

For readers, the practical value of this observation is that it reduces DeepSeek's cost-performance label to a traceable technical route. If you care about AI infrastructure or model training efficiency, what deserves attention is not the parameter count of the next large model, but the next foundational bottleneck to be conquered—attention data structures, expert scheduling, network topology constraints, inference compute-memory access alignment, and agent training environment isolation. Liang Wenfeng's signing pattern at least suggests: someone has been consistently working in these directions.

ソース

  • leiphone.com↗
トピック

梁文锋底层技术路线

【背景】DeepSeek创始人梁文锋在过去三年中极少出现在动辄上百人署名的旗舰模型整体报告中,而是始终将名字签在MoE路由、注意力机制、网络拓扑、推理算子和沙箱基础设施等最底层的原子技术论文上。这一署名模式揭示了一条一致的技术路线:在架构最…

3 ステップ1 件更新日 2026-10-21 20:53

経緯

  1. 2024-01-11梁文锋作为作者之一发布《DeepSeekMoE》,提出细粒度专家动态路由与共享专家隔离,旨在实现终极专家专业化并降低计算成本。arxiv.org ↗
  2. 2025-07-01梁文锋参与撰写的《Native Sparse Attention (NSA)》入选ACL 2025,该论文提出硬件对齐的原生稀疏注意力机制,显著提升长序列处理速度。aclanthology.org ↗
  3. 2026-09-19DeepSeek发布新论文《DSec》,梁文锋为通讯作者。该论文首次展示自动化沙箱系统,用于大规模智能体强化学习与评测,每日可自动运行300万个沙箱环境。leiphone.com ↗

未解決の問題

  • 梁文锋团队在底层技术上的持续投入,是否会导致其在应用层或商业化产品形态上出现滞后?
  • 随着模型规模进入千亿级无人区,数学重构底座的具体下一步突破点是什么?
  • DSec沙箱基础设施如何平衡AI进化训练需求与防止AI失控的安全边界?
トピック全文を読む →
リサーチを続ける →