Skip to main content

constructionism

Educational philosophy — learn by building inspectable things

来源信息

仓库
SimHacker/moollm
最近来源活动
2026年2月7日 14:40
检测到的 SKILL.md 语言
英语
星标
60
Forks
5

安装方式

默认使用会先检查来源的 Prompt;你也可以切换为直接命令,或下载本地副本。

检查来源文件

决定是否安装前,请先阅读 SKILL.md,以及 SkillsMP 当前展示的配套文件。

文件资源管理器
5 个文件

正在显示 SKILL.md

SKILL.md
来源说明 · 只读预览
name
constructionism
description
Educational philosophy — learn by building inspectable things
allowed-tools
[]
tier
0
protocol
CONSTRUCTIONISM
origin
Seymour Papert — Logo, Mindstorms (1980)
lineage
["Seymour Papert — Constructionism, Logo, Mindstorms (1980)","Marvin Minsky — Society of Mind, K-lines (1986)"]
related
["play-learn-lift","society-of-mind","leela-ai","manufacturing-intelligence","room","yaml-jazz","adventure","skill","schema-mechanism","debugging"]
tags
["moollm","philosophy","education","papert","microworld","learning"]
# Constructionism > *"If you can build it, you can understand it. If you can inspect it, you can trust it."* Seymour Papert's educational philosophy: you learn best by **building things** you can **inspect and modify**. Not passive consumption. Not abstract explanation. **Construction.** ## The Tradition **Logo Microworlds** — Children don't learn geometry from textbooks. They teach a turtle to draw shapes: ```logo TO SQUARE :SIZE REPEAT 4 [FORWARD :SIZE RIGHT 90] END ``` The child: 1. **Builds** the procedure 2. **Runs** it and sees results 3. **Debugs** when it's wrong 4. **Understands** geometry through construction ## MOOLLM as Microworld | Logo | MOOLLM | |------|--------| | Turtle | Agent/Character | | Canvas | Room floor | | Procedures | Skills | | Variables | YAML state | | Drawing | File creation | **Everything is inspectable.** Open `ROOM.yml` — see the state. Read `session-log.md` — see the history. Modify `character.yml` — change the world. ## Core Principles ### Low Floor Easy to start — no setup, just explore: ``` > LOOK You are in the workshop. > EXAMINE hammer A simple claw hammer. ``` ### High Ceiling Unlimited complexity — build custom skills, complex pipelines, new protocols. ### Wide Walls Many paths to many goals — adventure games, workflow automation, knowledge organization. ## Learning by Doing ### The Debug Cycle 1. **Try** something — it doesn't work 2. **Inspect** state — see what happened 3. **Hypothesize** — "maybe the path is wrong" 4. **Modify** and retry — test the hypothesis 5. **Understand** — now you know how it works ### Cheating is Learning From Don's Logo Adventure: > Type `PRINT :ITEMS` to see where everything is. > Type `MAKE "RNUM 5` to teleport to the treasure room. > **If you cheat, you win by learning Logo.** "Cheating" in MOOLLM: ``` > Open character.yml directly > Add "magic_sword" to inventory > You've learned YAML and file structure! ``` The system rewards curiosity with knowledge. ## Micropolis: The Dream Don's Micropolis for OLPC applied the same philosophy to SimCity: - Open source simulation - Scriptable in Python - Kids can modify the rules - The city IS the curriculum MOOLLM applies this to LLM agents: - Open file state - Scriptable in any language - Users can modify the rules - **The filesystem IS the microworld** ### Embed Micropolis in MOOLLM ``` cities/downtown/ ├── ROOM.yml # Room metadata ├── city.save # Micropolis save file ├── state.yml # Extracted game state ├── newspaper/ # Generated stories ├── advisors/ # Expert cards └── session-log.md ``` LLM reads state, plays the game, summons advisors. ## PLAY-LEARN-LIFT Constructionism in action: 1. **PLAY** — Explore manually, make messes 2. **LEARN** — Notice patterns, understand 3. **LIFT** — Extract principles, create skills You don't design skills in the abstract. You **build them** from experience. --- ## Drescher's Schema Mechanism Gary Drescher's *Made-Up Minds* (1991) extends constructionism into a computational theory of how minds learn causal models. Drescher was a student of Minsky at MIT. **See: [../schema-mechanism/](../schema-mechanism/)** for the full treatment. ### The Core Idea A **schema** is a causal unit: `Context → Action → Result` The agent discovers schemas through experience, refining them via **marginal attribution** -- tracking which conditions correlate with success. ### Connection to PLAY-LEARN-LIFT | Schema Mechanism | PLAY-LEARN-LIFT | |-----------------|-----------------| | ACT + OBSERVE | PLAY | | ATTRIBUTE (marginal attribution) | LEARN | | SPIN OFF (refine schemas) | LIFT | ### Why LLMs Complete Drescher | Aspect | Deterministic | LLM + YAML Jazz | |--------|---------------|-----------------| | Items | Opaque tokens | Grounded meanings | | Patterns | Statistical correlation | Semantic understanding | | Spin-offs | Mechanical | Creative generalization | | Explanations | None | Natural language | > *"The YAML provides the skeleton; the LLM provides the soul."* --- ## The Insight > *"If you can build it, you can understand it."* > *"If you can inspect it, you can trust it."* > *"The filesystem IS the microworld."* See also: [schema-mechanism](../schema-mechanism/) for Drescher's computational extension.
在 GitHub 查看