Post

HKUDS00000: StudyMap3 - HKUDS020 之后的八条后续路线

HKUDS00000: StudyMap3 - HKUDS020 之后的八条后续路线

这是 PENGYI_HKUDS_STUDYMAP 的第三张学习地图。

1
HKUDS00000 -> StudyMap3

前两张地图分别是:

MapRole
HKUDS000第一张总览地图,确认 HKUDS 本地 repo 宇宙和第一批优先级
HKUDS0000中场地图,把 HKUDS000-013 归纳成四大主线
HKUDS00000第三张地图,确定 HKUDS020 之后的后续路线

现在我们已经实际完成到:

1
HKUDS019 -> Paper2Slides

所以从这里开始,编号必须重新收束。

这张 StudyMap3 的作用就是:

1
明确 HKUDS020 之后按什么顺序继续走。

这不是随便排项目。

它服务于我们的主目标:

1
2
3
4
Pengyi Research OS
+ Pengyi Quant Research OS
+ AI Scientist Stack
+ Open-source research portfolio

当前标准顺序

后续以这张表为准。

编号Repo系列为什么看
HKUDS020FutureShowQuant / Forecasting预测、forecast agent、prediction market,接我们的量化判断系统
HKUDS021VideoRAGRAG / Knowledge把访谈、课程、讲座、视频资料进入知识系统
HKUDS022FastCodeCoding / Research Engineering代码理解加速,接 DeepCode 和我们的工程能力
HKUDS023OpenSpaceAgent Workspaceself-evolving agent space,接 AgentSpace / OpenHarness / AutoAgent
HKUDS024+GraphAgent / OpenGraph / GraphGPT / HiGPTGraph / Knowledge Graph图智能、知识图谱问答,接 QuantMind / Research OS 记忆层
HKUDS025+RecLM / XRec / AutoCF / KGRecRecommendation / Finance-adjacent推荐系统、用户行为、金融/交易信号抽象可借鉴
HKUDS laterUrbanGPT / OpenCity / EasyST / AutoSTSpatiotemporal / Urban AI时空预测、城市智能,和量化时间序列也有相通处
HKUDS laterClawTeam / ClawWork / FastAgent / LitewriteAgent Product / Workspace多 agent 协作、工作流、写作与产品化

一句话版本:

1
2
3
4
5
6
7
8
9
HKUDS020 之后进入八条后续路线:
Forecasting
Video / Multimodal RAG
Code Intelligence
Agent Workspace
Graph / Knowledge Graph
Recommendation / Finance-adjacent
Spatiotemporal Intelligence
Agent Product / Workspace

这就是第三阶段地图。

为什么现在需要 StudyMap3

前面我们已经完成了一段很强的系统闭环。

HKUDS001HKUDS019,大致形成了:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
knowledge ingestion
-> research memory
-> personal agent shell
-> tool action
-> quant workflow
-> live trading platform
-> organizational workspace
-> agent factory
-> research-to-code
-> autonomous research
-> idea generation
-> deep research
-> report evaluation
-> AI scientist self-training
-> agent runtime
-> skill distillation
-> universal tool-use
-> lightweight memory
-> research presentation artifact

也就是:

1
2
3
4
5
6
7
8
9
能读
能记
能做
能研究
能写报告
能评估
能学习
能调用工具
能表达成 slides / poster

这已经不是零散学习了。

现在缺的是第二层扩展:

1
2
3
4
5
6
7
8
预测能力
视频知识能力
代码理解加速能力
更完整的 agent workspace
图智能和知识图谱能力
推荐/用户行为/金融信号抽象能力
时空预测能力
产品化 agent workspace 能力

所以 StudyMap3 的目标不是“多看几个 repo”,而是:

1
补齐 Research OS 和 Quant OS 还缺的能力面。

已完成到 HKUDS019

当前已完成的 numbered posts:

IDRepo / TopicSystem Position
HKUDS000Study Map第一张 HKUDS 总览地图
HKUDS001LightRAGgraph-based research memory
HKUDS002Vibe-Tradingagentic quant research workflow
HKUDS003nanobotpersonal always-on agent shell
HKUDS004CLI-Anythingagent-native software action layer
HKUDS005AI-Traderagent-native live trading platform
HKUDS006AgentSpaceorganizational agent workspace
HKUDS007RAG-Anythingmultimodal document ingestion
HKUDS008AutoAgentagent framework / agent factory
HKUDS009DeepCodepaper-to-code / research-to-code
HKUDS010AI-Researcherautonomous scientific discovery workflow
HKUDS011DeepInnovatorscientific idea model training
HKUDS012Auto-Deep-Researchpractical deep research assistant
HKUDS013DeepResearch-Evalreport evaluation and factuality checking
HKUDS014DeepTutorAI scientist self-training layer
HKUDS015OpenHarnessagent harness runtime
HKUDS016UpSkillfailure-to-skill self-improvement layer
HKUDS017AnyTooluniversal tool-use / capability routing
HKUDS018MiniRAGlightweight graph RAG / local memory
HKUDS019Paper2Slidesresearch-to-presentation artifact generation

这张表确定了一个事实:

1
HKUDS020 不是继续补基础层,而是进入能力扩展层。

Series 1: Quant / Forecasting

1
HKUDS020 -> FutureShow

FutureShow 放在 HKUDS020 是合理的。

因为它接的是我们的量化主线:

1
2
3
4
5
forecasting
prediction
judgment
prediction market
future event reasoning

我们做 Quant Research OS,不能只做:

1
2
3
4
data ingestion
factor implementation
backtest
report generation

还要能回答:

1
2
3
4
5
系统如何形成预测?
系统如何表达不确定性?
系统如何记录 forecast history?
系统如何比较不同 agent / model 的判断质量?
系统如何把预测变成可评估对象?

这和量化非常接近。

一个 factor hypothesis 本质上也是 forecast:

1
2
given features today
predict future return distribution

一个 macro / news / event-driven 判断也是 forecast:

1
2
given event + market context
predict price reaction / volatility / regime shift

所以 FutureShow 对我们的关键词是:

1
2
3
forecast as artifact
prediction as trackable object
judgment as evaluable system output

我们做 HKUDS020 时要重点看:

QuestionWhy
它如何定义 forecast object?这能迁移到 quant signal / research hypothesis
是否有 prediction market 或 crowd judgment 结构?这能迁移到多 agent research debate
如何记录预测、结果和评分?这能变成 research track record
是否支持 agent 参与预测?这能接 R&D Agent
是否有 UI / product layer?这能接 public demo 和网站展示

FutureShow 是第三阶段的第一站。

Series 2: RAG / Knowledge

1
HKUDS021 -> VideoRAG

前面我们已经看了:

1
2
3
LightRAG
RAG-Anything
MiniRAG

这三者分别解决:

RepoRole
LightRAGgraph-based RAG
RAG-Anythingmultimodal document ingestion
MiniRAGlightweight / small-model-friendly RAG

但我们还有一个重要知识源没有系统化:

1
video

视频对我们很重要。

因为很多高价值知识不在 paper 里,而在:

1
2
3
4
5
6
7
8
访谈
课程
讲座
seminar
podcast
公司分享
研究者 talk
硅谷 101 这类高质量对谈

我们最近看田渊栋访谈就是典型例子。

这些内容如果只靠“看过了”,就很容易丢。

VideoRAG 可以接到我们的个人知识系统:

1
video -> transcript -> chunks -> visual / temporal grounding -> retrieval -> notes -> Research OS memory

HKUDS021 时要重点看:

QuestionWhy
它如何切分长视频?决定是否能吃课程和访谈
它是否保留 timestamp?方便回看证据
它如何处理视觉内容?slides、图表、白板内容可能是关键
它如何 query 视频?影响后续个人学习系统
是否能生成 structured note?直接接网站学习笔记

对我们的目标来说,VideoRAG 是:

1
让视频资料进入 Research OS 的入口层。

Series 3: Coding / Research Engineering

1
HKUDS022 -> FastCode

前面我们已经做了:

1
HKUDS009 -> DeepCode

DeepCode 更像是:

1
research-to-code / paper-to-code implementation layer

FastCode 从名字和主线定位上,更适合放在:

1
2
code understanding acceleration
research engineering speed layer

我们现在最核心的生产力就是 coding。

所以第三阶段必须补一个代码理解加速方向。

未来 Research OS 不只是调用 LLM 写代码,还要做到:

1
2
3
4
5
6
快速读 repo
快速定位核心模块
快速理解依赖关系
快速找到 bug / PR 机会
快速把 paper idea 转成 runnable code
快速生成 study map / architecture map

FastCode 可以和这些项目连起来:

RepoRelation
DeepCode深度代码实现和研究复现
CLI-Anythingagent-native software action
AnyTooltool routing / execution
OpenHarnesscoding agent runtime
FastCodecode understanding acceleration

HKUDS022 时要重点看:

QuestionWhy
它加速的是 coding 哪个环节?read / search / edit / test / review
是否有代码索引结构?可迁移到 repo study automation
是否有 benchmark?判断不是纯 demo
是否适合接 Codex workflow?直接影响我们的日常生产力
是否能形成 PR workflow?接 open-source contribution

我们要成为 AI scientist,code speed 是基本盘。

FastCode 就是这个基本盘的一块。

Series 4: Agent Workspace

1
HKUDS023 -> OpenSpace

前面我们已经看过多个 agent workspace / runtime:

1
2
3
4
5
6
nanobot
AgentSpace
AutoAgent
OpenHarness
AnyTool
UpSkill

它们分别覆盖:

RepoRole
nanobotpersonal always-on agent shell
AgentSpaceorganizational agent workspace
AutoAgentagent framework / agent factory
OpenHarnessagent runtime / harness
AnyTooltool-use layer
UpSkillskill distillation

OpenSpace 放在 HKUDS023,是因为它可以作为第二阶段 agent workspace 的总合流。

关键词是:

1
2
3
4
self-evolving agent space
multi-agent workspace
agent collaboration
agent memory and skill evolution

对我们的 Research OS 来说,workspace 很关键。

因为一个 AI scientist 系统不能只有一个 agent。

它应该有角色分工:

1
2
3
4
5
6
7
8
Researcher Agent
Developer Agent
Evaluator Agent
Reviewer Agent
PM Agent
Data Agent
Writing Agent
Presentation Agent

OpenSpace 可以帮助我们思考:

1
2
3
4
5
多个 agent 在哪里协作?
任务如何分配?
状态如何共享?
成果如何沉淀?
agent 如何进化?

HKUDS023 时要重点看:

QuestionWhy
它的 workspace model 是什么?决定能否接 Research OS
agent 是否有 memory / skill / role?接 UpSkill / MiniRAG
是否支持 multi-agent collaboration?接 R&D team 模型
是否有 UI / product surface?接网站和个人工作台
是否能跑真实任务?判断工程成熟度

这是从 agent 框架走向 agent organization 的一站。

Series 5: Graph / Knowledge Graph

1
HKUDS024+ -> GraphAgent / OpenGraph / GraphGPT / HiGPT

这一组可以作为 HKUDS024+,不必只做一篇。

原因是 HKUDS 的 graph 系列很多,而且和我们的目标高度相关。

我们已经多次提到:

1
2
3
4
5
6
QuantMind
knowledge graph
Research OS memory
factor relation graph
paper idea graph
organization knowledge graph

图智能是一个底层能力。

尤其对 quant 来说,很多对象天然是图:

Quant ObjectGraph View
股票nodes
行业 / 供应链 / 股权关系edges
因子feature nodes
新闻事件event nodes
研报观点claim nodes
策略strategy nodes
回测结果evidence nodes
portfolioweighted subgraph

对 Research OS 来说,很多对象也是图:

Research ObjectGraph View
papernode
methodnode
datasetnode
experimentnode
claimnode
evidenceedge
limitationedge / node
follow-up ideanode

所以 GraphAgent / OpenGraph / GraphGPT / HiGPT 应该成为一个专题系列。

HKUDS024+ 时要重点看:

RepoFirst Question
GraphAgentagent 如何操作 graph / KG?
OpenGraph是否提供开放图智能框架?
GraphGPTgraph 和 language model 如何对齐?
HiGPTheterogeneous graph intelligence 如何做?

这一组最终要服务:

1
2
3
Research OS memory graph
QuantMind knowledge graph
factor / event / strategy relation graph

Series 6: Recommendation / Finance-adjacent

1
HKUDS025+ -> RecLM / XRec / AutoCF / KGRec

推荐系统看起来不是量化,但其实很有借鉴意义。

因为推荐系统和量化有很多结构相似:

RecommendationQuant
userinvestor / account / strategy
itemstock / asset / factor / signal
interactiontrade / click / return / exposure
rankingstock selection
feedbackrealized return
cold startnew stock / new factor
preferencerisk appetite / style exposure
collaborative filteringcross-sectional relation

所以 RecLM / XRec / AutoCF / KGRec 可以放在 Finance-adjacent。

我们不一定要直接做推荐系统,但可以借:

1
2
3
4
5
6
ranking
retrieval
preference modeling
feedback loop
user-item graph
knowledge-aware recommendation

这些方法可以迁移到:

1
2
3
4
5
6
stock ranking
factor recommendation
research idea recommendation
paper recommendation
strategy allocation
PM preference modeling

HKUDS025+ 时要重点看:

QuestionWhy
它如何建模 user-item interaction?可迁移到 signal-return interaction
是否使用 KG?接 QuantMind
是否有 LLM recommender?接 research idea recommender
如何做 evaluation?接 backtest / ranking metrics
是否有 cold-start 处理?接新因子 / 新标的

这一组可以帮助我们把“推荐”转成“研究对象排序”。

Series 7: Spatiotemporal / Urban AI

1
HKUDS later -> UrbanGPT / OpenCity / EasyST / AutoST

这一组可以稍后看。

它看起来偏城市智能,但和量化时间序列也有相通处。

关键词是:

1
2
3
4
5
6
spatiotemporal prediction
traffic forecasting
urban data
temporal graph
sequence modeling
region interaction

量化里也有类似问题:

1
2
3
4
5
6
time series
cross-sectional relation
market regime
sector rotation
lead-lag effect
event propagation

城市里是区域之间的流动。

市场里是资产之间、行业之间、事件之间的信息流动。

所以这组 repo 的价值不是“做城市项目”,而是学习:

1
2
3
时空数据如何建模
多区域/多节点关系如何进入预测
temporal graph 如何和 LLM / agent 结合

做这一组时要重点看:

RepoPossible Value
UrbanGPTLLM + urban intelligence
OpenCitycity-scale data / platform 思路
EasySTspatiotemporal baseline / framework
AutoSTautomated spatiotemporal modeling

这一组可以作为后续时间序列和图预测补课。

Series 8: Agent Product / Workspace

1
HKUDS later -> ClawTeam / ClawWork / FastAgent / Litewrite

这一组更偏产品化。

前面 agent framework 更多是:

1
2
3
4
5
runtime
tool
workspace
skill
agent loop

但如果要真的把个人系统做成可用产品,还需要:

1
2
3
4
5
team collaboration
work management
writing workflow
fast agent execution
user-facing interface

ClawTeam / ClawWork / FastAgent / Litewrite 可以作为 agent product 系列。

对我们来说,它们可能服务:

1
2
3
4
5
个人 Research OS 工作台
私有 pitch workspace
网站内容生产
团队式多 agent 协作
RA / PhD / job application material management

这一组不是最急,但很适合后面做产品化参考。

做这一组时要重点看:

RepoWhy
ClawTeamteam-based multi-agent collaboration
ClawWorkwork/task workspace
FastAgentlightweight fast agent execution
Litewritewriting / content production workflow

这条线会把 Research OS 从“自己用的工程系统”推向“可以展示、可以交互、可以复用的产品”。

新的总路线

从现在开始,HKUDS 后续可以按这个顺序推进:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
HKUDS020 FutureShow
  -> forecasting / prediction / judgment

HKUDS021 VideoRAG
  -> video knowledge ingestion

HKUDS022 FastCode
  -> code intelligence / research engineering speed

HKUDS023 OpenSpace
  -> self-evolving agent workspace

HKUDS024+ GraphAgent / OpenGraph / GraphGPT / HiGPT
  -> graph intelligence / knowledge graph memory

HKUDS025+ RecLM / XRec / AutoCF / KGRec
  -> recommendation / ranking / finance-adjacent signals

HKUDS later UrbanGPT / OpenCity / EasyST / AutoST
  -> spatiotemporal intelligence

HKUDS later ClawTeam / ClawWork / FastAgent / Litewrite
  -> agent product / workspace / writing

这条路线有一个很清楚的结构:

1
2
3
4
5
6
7
8
Predict
Know from videos
Code faster
Coordinate agents
Reason on graphs
Rank / recommend
Model time-space systems
Productize agent work

对应到我们的系统:

CapabilityOur System Need
Predictquant judgment / forecast record
Video Knowledgeinterviews, lectures, courses as memory
Code Intelligencefaster repo study and PR output
Agent WorkspaceR&D agent organization
Graph IntelligenceQuantMind / research memory graph
Recommendationidea / factor / paper / strategy ranking
Spatiotemporaltemporal graph and market regime modeling
Agent Productwebsite, private workspace, public demo

和 Quant Research OS 的关系

这张 StudyMap3 对 Quant OS 的影响非常直接。

我们可以把后续 repo 映射到 Quant OS:

Quant OS LayerHKUDS Future Repo
forecast engineFutureShow
video/news/course knowledgeVideoRAG
research engineering speedFastCode
multi-agent research roomOpenSpace
factor/event/strategy graphGraphAgent, OpenGraph, GraphGPT, HiGPT
signal ranking / idea recommendationRecLM, XRec, AutoCF, KGRec
time-series / regime / cross-market relationUrbanGPT, OpenCity, EasyST, AutoST
PM pitch / writing / workflow productClawTeam, ClawWork, FastAgent, Litewrite

这说明 HKUDS 后续不是偏离量化,反而是在给量化系统补底层能力。

真正的 Quant OS 需要:

1
2
3
4
5
6
7
8
data
forecast
ranking
graph relation
time-series model
research workflow
agent organization
presentation artifact

前面我们已经有 data / research workflow / artifact 的基础。

后面要补 forecast / graph / ranking / spatiotemporal。

和 AI Scientist 的关系

对 AI Scientist 来说,StudyMap3 也很关键。

AI Scientist 不是只会读 paper 和写代码。

它应该能:

1
2
3
4
5
6
7
8
预测研究方向
吸收视频和讲座
快速理解代码库
组织多个 agent 做任务
在知识图谱上推理
推荐下一篇 paper / 下一个 idea
理解时间演化和复杂系统
把成果产品化和表达出来

所以第三阶段路线本质上是:

1
2
from research workflow
to research intelligence stack

前面 HKUDS001-019 帮我们建立了 workflow。

后面 HKUDS020+ 要帮我们建立 intelligence。

一句话总结

HKUDS00000 StudyMap3 的核心判断是:

1
2
HKUDS020 之后不要散着看。
按 Forecasting -> VideoRAG -> FastCode -> OpenSpace -> Graph -> Recommendation -> Spatiotemporal -> Agent Product 的顺序推进。

这样每一步都有明确意义:

1
2
3
4
5
6
7
8
FutureShow gives us forecasting.
VideoRAG gives us video memory.
FastCode gives us coding speed.
OpenSpace gives us agent organization.
Graph series gives us knowledge graph reasoning.
Recommendation series gives us ranking and signal abstraction.
Spatiotemporal series gives us temporal-graph prediction.
Agent product series gives us workspace and productization.

这张地图之后,我们后续就不乱。

下一篇正式进入:

1
HKUDS020 -> FutureShow

它会把我们重新拉回 Quant / Forecasting 主线。

This post is licensed under CC BY 4.0 by the author.