跳到主要内容Skip to main content

学 AI · 2026 补丁包 · 2026-11-18

四个词,补上四条最常被跳过的线Four terms for four lines that are usually skipped

词库从 256 到 260。不是凑数:训练(GRPO)、推理(MLA)、表征(JEPA)、生成(Flow Matching)各有一个常被提到、但站内没有独立词条的概念。The glossary moves from 256 to 260 — not padding: training (GRPO), inference (MLA), representation (JEPA) and generation (Flow Matching) each had a frequently mentioned concept without a standalone entry.

每个词条都写出处、边界与「它没主张什么」;读完能判断一个产品页说的是哪一层,而不是多背一个缩写。Each entry gives sources, boundaries and what it does not claim; after reading, you can tell which layer a product page is talking about instead of memorising another acronym.

词条Terms 260新增New 4指南更新Guide update 1

四个新词Four new terms

按「训练 → 推理 → 表征 → 生成」排,不是按字母。每个都链到独立词条页。Ordered training → inference → representation → generation, not alphabetically. Each links to its own term page.

训练TRAININGGRPO
推理INFERENCEMLA
表征REPRESENTATIONJEPA
生成GENERATIONFlow Matching
01训练Training

GRPOGRPO →

用同一道题的一组回答互相比较,省掉 PPO 的 critic。看懂「推理能力是训出来的」到底在训什么。Compare a group of answers to one question instead of training a PPO critic. Understand what “reasoning is trained” actually trains.

边界Boundary

不是无需奖励模型;组大小与奖励设计直接影响稳定性。Not reward-model-free; group size and reward design drive stability.

02推理Inference

MLAMLA →

把 KV 缓存压成低秩潜向量,长上下文的第一道墙因此松动。Compresses KV cache into a low-rank latent, relaxing the first wall in long context.

边界Boundary

省的是显存与带宽,不是 FLOPs;不是无限上下文。Saves memory and bandwidth, not FLOPs; not infinite context.

03表征Representation

JEPAJEPA →

在表示空间预测被遮住的部分,而不是重建像素或 token;2026 的 JEPA-PRH 用它检验跨模态对齐。Predicts masked parts in representation space rather than reconstructing pixels or tokens; JEPA-PRH uses it to test cross-modal alignment in 2026.

边界Boundary

对齐分数不是生成质量;预印本规模有限。Alignment is not generation quality; the preprint is limited in scale.

04生成Generation

Flow MatchingFlow Matching →

学一个从噪声到数据的速度场,用 ODE 积分采样;少步生成的主流底层之一。Learns a velocity field from noise to data and samples by ODE integration; one mainstream basis for few-step generation.

边界Boundary

不是一步生成;步数与蒸馏有关,FID 不等于人类偏好。Not one-step generation; steps depend on distillation, and FID is not human preference.

2026 变化线:补到 11 月 18 日The 2026 timeline: updated to 18 November

指南从 8 条扩到 15 条,补进 Byteification、JEPA-PRH、Muse Spark、Mistral Large 4、Agent Skills、MI450X。The guide grows from 8 rows to 15, adding Byteification, JEPA-PRH, Muse Spark, Mistral Large 4, Agent Skills and MI450X.

10-07Byteification 登上 Nature;49.1B token 把子词模型改字节级。
10-15Meta Muse Spark Open:816B / 42B 激活,Muse Model License。
10-22Agent Skills 捐给 AAIF;技能成为跨平台标准。
10-29Mistral Large 4:1.05T / 49B 激活;JEPA-PRH 预印本同日。
11-12AMD MI450X 上市:288GB HBM4 / 23.3TB/s。

读完整指南 →Read the full guide →