Tutorial: Research Notes with Flint

flint init ai-safety-research
cd "(Flint) ai-safety-research"

[flint]
name = "ai-safety-research"
type = "flint"

[plugins]
required = [
  "living-documents",
  "notepad",
  "archive"
]

[mods]
required = ["git"]

flint sync

Mesh/
├── (System) Flint Init.md
├── (System) Index.md
├── Sources/           # Primary sources
├── Concepts/          # Key ideas
├── Syntheses/         # Your analysis
└── Drafts/            # Work in progress

---
id: (auto-generated)
tags:
  - "#source"
  - "#book"
---

# Superintelligence (Bostrom, 2014)

**Author:** Nick Bostrom
**Year:** 2014
**Type:** Book

## Key Claims

1. Superintelligent AI could emerge this century
2. Control problem is the central challenge
3. Value alignment is technically difficult

## Important Quotes

> "The first superintelligence may shape the future of Earth-originating life."
> — Chapter 1

> "A superintelligence could be the last invention humanity ever needs to make."
> — Chapter 2

## My Notes

Strong argument for taking AI safety seriously early.
The orthogonality thesis is particularly compelling.

## Related

- [Concepts/Orthogonality Thesis](/concepts-orthogonality-thesis)
- [Concepts/Control Problem](/concepts-control-problem)
- [Sources/Russell 2019 - Human Compatible](/sources-russell-2019-human-compatible)

---
id: (auto-generated)
tags:
  - "#concept"
  - "#core"
---

# Orthogonality Thesis

The claim that intelligence and goals are independent: a system
can be arbitrarily intelligent while pursuing any goal.

## Definition

"Intelligence and final goals are orthogonal: more or less any
level of intelligence could in principle be combined with more
or less any final goal." — Bostrom

## Implications

1. Smart doesn't mean benevolent
2. Can't assume advanced AI shares human values
3. Alignment must be engineered, not assumed

## Sources

- [Sources/Bostrom 2014 - Superintelligence](/sources-bostrom-2014-superintelligence)
- [Sources/Armstrong 2015 - Motivated Value Selection](/sources-armstrong-2015-motivated-value-selection)

## Related Concepts

- [Control Problem](/control-problem)
- [Value Alignment](/value-alignment)
- [Instrumental Convergence](/instrumental-convergence)

---
id: (auto-generated)
tags:
  - "#synthesis"
  - "#ld/living"
---

# Control Problem Overview

My synthesis of the AI control problem literature.

## The Core Challenge

Given the [Orthogonality Thesis](/orthogonality-thesis), we cannot assume advanced AI
will share human values. The control problem asks: how do we
ensure AI systems remain beneficial as they become more capable?

## Key Approaches

### 1. Corrigibility

Making AI systems that accept correction. Challenges include:
- Incentive to resist shutdown
- Preserving own goals

See [Concepts/Corrigibility](/concepts-corrigibility) for details.

### 2. Value Learning

Having AI infer human values through observation.

See [Concepts/Value Learning](/concepts-value-learning) for details.

### 3. Constitutional AI

Training with explicit principles. Anthropic's approach.

See [Sources/Bai 2022 - Constitutional AI](/sources-bai-2022-constitutional-ai).

## Open Questions

1. Can value learning scale to superintelligence?
2. Is corrigibility stable under self-improvement?
3. What happens with multiple competing AI systems?

## Sources

![Sources/Bostrom 2014 - Superintelligence#Key Claims](/sources-bostrom-2014-superintelligence-key-claims)
![Sources/Russell 2019 - Human Compatible#Central Argument](/sources-russell-2019-human-compatible-central-argument)

Media/
├── papers/
│   ├── bostrom-2014-superintelligence-excerpt.pdf
│   └── russell-2019-human-compatible-ch1.pdf
└── diagrams/
    └── control-problem-taxonomy.png

See the original argument: ![Media/papers/bostrom-2014-superintelligence-excerpt.pdf](/media-papers-bostrom-2014-superintelligence-excerpt-pdf)

Visual overview:
![Media/diagrams/control-problem-taxonomy.png](/media-diagrams-control-problem-taxonomy-png)

---
id: (auto-generated)
tags:
  - "#system"
  - "#dashboard"
---

# Index

## By Type

### Sources
- [Sources/Bostrom 2014 - Superintelligence](/sources-bostrom-2014-superintelligence)
- [Sources/Russell 2019 - Human Compatible](/sources-russell-2019-human-compatible)

### Core Concepts
- [Concepts/Orthogonality Thesis](/concepts-orthogonality-thesis)
- [Concepts/Control Problem](/concepts-control-problem)
- [Concepts/Value Alignment](/concepts-value-alignment)

### Syntheses
- [Syntheses/Control Problem Overview](/syntheses-control-problem-overview)

## By Topic

### Foundational
- [Concepts/Orthogonality Thesis](/concepts-orthogonality-thesis)
- [Concepts/Instrumental Convergence](/concepts-instrumental-convergence)

### Solutions
- [Concepts/Corrigibility](/concepts-corrigibility)
- [Concepts/Value Learning](/concepts-value-learning)

---
id: (auto-generated)
tags:
  - "#draft"
  - "#paper"
  - "#ld/living"
---

# Draft: Control Problem Survey

## Abstract

[To be written]

## 1. Introduction

The development of artificial general intelligence poses
unprecedented challenges for ensuring beneficial outcomes.
This paper surveys approaches to the AI control problem.

## 2. The Control Problem

![Syntheses/Control Problem Overview#The Core Challenge](/syntheses-control-problem-overview-the-core-challenge)

### 2.1 Theoretical Foundations

The [Orthogonality Thesis](/orthogonality-thesis) establishes that...

## 3. Approaches

### 3.1 Corrigibility

![Concepts/Corrigibility#Definition](/concepts-corrigibility-definition)

### 3.2 Value Learning

[Content]

## 4. Open Problems

[Content]

## References

- Bostrom, N. (2014). Superintelligence.
- Russell, S. (2019). Human Compatible.

---
id: (auto-generated)
description: Survey paper for publication
---

# Survey Paper Export

[Drafts/Control Problem Survey](/drafts-control-problem-survey)

flint export build Survey