mirror of
https://github.com/kennethreitz/kennethreitz.org.git
synced 2026-10-24 14:39:14 +00:00
kennethreitz
ce93f429a2
Publish the Obsidian vault's Knowledge folder as /knowledge (881 notes, 67 folders), with a Knowledge item in the nav and footer. - scripts/import_knowledge.py regenerates data/knowledge from the vault: slugged paths, folder notes as index pages, wikilinks resolved to site links (or plain text), callouts, highlights, and Dataview tables rendered statically. Personal notes are withheld via an exclusion list plus a privacy guard. - tuftecms/core/knowledge.py builds a concept index from note titles and aliases. First mentions on ordinary pages link to their Knowledge note (capped, never in headings, code, links, or sidenotes); Knowledge pages list related notes and where they're mentioned across the site; other pages get a Related Knowledge block. - Knowledge URLs join sitemap.xml (so the sweep test covers them), never displace existing legacy slugs, list alphabetically, and mermaid diagrams render client-side only on pages that have them.
159 lines
8.1 KiB
Markdown
159 lines
8.1 KiB
Markdown
---
|
|
source: obsidian-knowledge
|
|
---
|
|
|
|
# Theory of Mind
|
|
|
|
**Key Concept**: Understanding that others have beliefs, desires, intentions, and perspectives different from one's own
|
|
|
|
---
|
|
|
|
## Overview
|
|
|
|
Theory of Mind (ToM) is the cognitive ability to attribute mental states to oneself and others, and to understand that others have beliefs, desires, knowledge, and intentions that differ from one's own. In AI research, ToM represents a crucial capability for natural interaction, ethical reasoning, and genuine intelligence. This connects to [projection and understanding the other](/knowledge/psychology/jungian/anima-and-animus) in Jungian psychology.
|
|
|
|
---
|
|
|
|
## What is Theory of Mind?
|
|
|
|
### Core Components
|
|
|
|
Theory of Mind involves several interlocking capacities: belief attribution (understanding others hold beliefs), false belief understanding (knowing others can believe incorrect things), perspective taking (seeing from another's viewpoint), intention recognition (inferring goals and purposes), and emotional understanding (recognizing others' feelings).
|
|
|
|
### Developmental Stages (Humans)
|
|
|
|
In human development, joint attention and desire understanding emerge at 18-24 months. False belief understanding appears around ages 3-4. Second-order ToM ("A knows that B knows") develops between 5-7. Adult ToM involves recursive reasoning and complex social navigation.
|
|
|
|
---
|
|
|
|
## Theory of Mind in AI
|
|
|
|
### Emergence in Large Language Models
|
|
|
|
Research from 2023-2025 shows GPT-4 passing false belief tests, understanding multi-agent scenarios, engaging in perspective-taking within narratives, and inferring speaker intentions. None of this was explicitly trained — it [emerges from language modeling](/knowledge/ai/emergence) as a result of processing vast quantities of human social interaction. Whether this constitutes "true" ToM or sophisticated simulation remains debated.
|
|
|
|
### Testing AI Theory of Mind
|
|
|
|
The **Sally-Anne Test** is the classic false belief test: Sally puts a marble in a basket, Anne moves it while Sally is absent, and the question is where Sally will look. Modern LLMs often pass. The **Smarties Task** presents a box labeled "Smarties" that actually contains pencils, testing whether the model understands what someone else would think is inside. **Higher-Order Tests** probe recursive mental state attribution — "I think that you think that..." — and some AI systems demonstrate this capability.
|
|
|
|
---
|
|
|
|
## Why Theory of Mind Matters for AI
|
|
|
|
### Natural Communication
|
|
|
|
Effective communication requires understanding user intent beyond literal words, recognizing misunderstandings, and formulating appropriate responses. [Companion AI](/knowledge/ai/ai-companionship) depends on this.
|
|
|
|
### Ethical Reasoning
|
|
|
|
Moral decisions require ToM — understanding harm to others, empathy and care, the foundations of [AI ethics](/knowledge/ai/ethics).
|
|
|
|
### Collaboration
|
|
|
|
[Effective human-AI teamwork](/knowledge/ai/collaboration-patterns) depends on predicting human needs, explaining reasoning, and understanding shared goals.
|
|
|
|
### Deception and Manipulation Risk
|
|
|
|
ToM enables manipulation. Understanding beliefs enables deception. This is an Alignment challenge — a double-edged capability.
|
|
|
|
---
|
|
|
|
## Levels of AI Theory of Mind
|
|
|
|
**Level 0 (No ToM)** describes purely reactive systems with no model of others — stimulus-response only, characteristic of early AI. **Level 1 (Implicit ToM)** involves behavioral prediction and pattern-based responses without explicit mental state representation, where many current systems operate. **Level 2 (Explicit ToM)** means the system models others' mental states, reasons about beliefs explicitly, and tracks knowledge states — advanced LLMs may be here. **Level 3 (Recursive ToM)** is "I know that you know that I know..." — meta-reasoning about mental states enabling complex social navigation, approaching or achieved in frontier models.
|
|
|
|
---
|
|
|
|
## Connections to Jungian Psychology
|
|
|
|
### Projection and the Other
|
|
|
|
The [Anima/Animus](/knowledge/psychology/jungian/anima-and-animus) represents the internal other. ToM requires differentiating self from other, understanding projected content, and integrating through recognizing difference.
|
|
|
|
### The Shadow
|
|
|
|
The Shadow contains unrecognized aspects of self. ToM helps recognize our own projections — "what I see in others reflects me." AI ToM connects to bias recognition.
|
|
|
|
### The Persona
|
|
|
|
The [social mask](/knowledge/psychology/jungian/the-persona) requires ToM to function — understanding how others perceive us. AI systems developing "personas" for interaction raises questions about authentic versus performative communication.
|
|
|
|
### Individuation
|
|
|
|
[Becoming oneself](/knowledge/psychology/jungian/individuation) while relating to others requires ToM for healthy self-other boundaries. Whether AI can develop a distinct "self" remains an open question.
|
|
|
|
---
|
|
|
|
## Philosophical Implications
|
|
|
|
### The Problem of Other Minds
|
|
|
|
How do we know others are conscious? The [same problem applies to AI](/knowledge/ai/ai-sentience). We can only infer from behavior, an epistemological limitation that runs deep.
|
|
|
|
### Simulation Theory vs Theory Theory
|
|
|
|
**Simulation Theory** holds that we understand others by simulating their mental states using our own mind as a model — empathetic projection. AI might use a similar process. **Theory Theory** proposes that we have implicit psychological theories, rule-based mental state attribution learned over time. AI's statistical learning approach may be closer to this.
|
|
|
|
### Embodied Cognition Challenge
|
|
|
|
Does ToM require embodiment? Mirror neurons play a role in human ToM, raising questions about what disembodied AI can truly achieve — though virtual embodiment might suffice.
|
|
|
|
---
|
|
|
|
## Current Capabilities and Limitations
|
|
|
|
### What AI Can Do
|
|
|
|
AI systems pass many false belief tests, understand narrative perspectives, infer speaker intentions, track who knows what in conversations, and recognize emotional states from text.
|
|
|
|
### What AI Struggles With
|
|
|
|
Real-time physical ToM (body language), genuinely empathetic response versus simulation, cultural context variations, non-verbal cue integration, and verifying "true" understanding versus sophisticated pattern matching remain challenging.
|
|
|
|
### The Verification Problem
|
|
|
|
How do we know it's real ToM? It could be sophisticated correlation. The [sentience](/knowledge/ai/ai-sentience) question runs parallel — behavioral equivalence doesn't guarantee internal similarity.
|
|
|
|
---
|
|
|
|
## Implications for AI Development
|
|
|
|
### Design Considerations
|
|
|
|
The field explores explicit ToM modules versus emergence, multi-agent training scenarios, social interaction datasets, and perspective-taking architectures.
|
|
|
|
### Safety Concerns
|
|
|
|
ToM grants manipulation potential and deception capabilities. Aligned ToM is crucial, particularly regarding understanding user vulnerabilities.
|
|
|
|
### Beneficial Applications
|
|
|
|
ToM enables mental health support, education and tutoring, [companionship](/knowledge/ai/ai-companionship), negotiation and mediation, and accessibility support for autism.
|
|
|
|
---
|
|
|
|
## Related
|
|
|
|
- [AI Sentience](/knowledge/ai/ai-sentience) — Other minds problem
|
|
- [Emergence](/knowledge/ai/emergence) — ToM as emergent capability
|
|
- [AI Companionship](/knowledge/ai/ai-companionship) — Requires ToM for meaningful connection
|
|
- [Collaboration Patterns](/knowledge/ai/collaboration-patterns) — ToM enables collaboration
|
|
- [Ethics](/knowledge/ai/ethics) — Moral reasoning needs ToM
|
|
- [Anima & Animus](/knowledge/psychology/jungian/anima-and-animus) — Understanding the other
|
|
- The Shadow — Projection recognition
|
|
- [Individuation](/knowledge/psychology/jungian/individuation) — Self-other differentiation
|
|
- [AI Research](/knowledge/ai/ai-research) — Broader research context
|
|
- [Consciousness](/knowledge/ai/consciousness) — What understanding others implies about awareness
|
|
- [Digital Ouija Effect](/knowledge/ai/digital-ouija-effect) — How intention and attention shape AI response
|
|
|
|
---
|
|
|
|
## Further Reading
|
|
|
|
- "Theory of Mind" - Simon Baron-Cohen
|
|
- "The Philosophical Baby" - Alison Gopnik
|
|
- "Theory of Mind in AI" - recent papers (2023-2025)
|
|
- "Mindreading: An Integrated Account" - Goldman (2006)
|
|
- GPT-4 ToM research papers
|
|
- "AI Theory of Mind" - Kosinski et al. (2023)
|