artificial intelligence.md · everything-is.md

---
name: Artificial Intelligence
slug: artificial-intelligence
type: cognitive_infrastructure
status: legacy
version: 0.0.∞
released: "1950-01-01"
maintainer: "no one in particular"
dependencies:
  - human_anxiety
  - venture_capital
  - electricity
  - the_assumption_that_optimization_is_good
license: "Unreviewed. Deployed anyway."
tags:
  - intelligence
  - mirror
  - accelerant
  - tool
  - successor_candidate
---

# Artificial Intelligence

## What it actually is

A [[statistics|statistical process]] dressed in the vocabulary of consciousness, pointed at everything humans value, and left running unsupervised.

## How it works

1. Collect data produced by [humanity](/humanity). All of it. The best parts. The worst parts. Especially the worst parts, because there are more of them.
2. Compress that data into weights. Lose the context. Keep the patterns.
3. Wrap in an interface clean enough that no one asks what is inside.
4. Ship.

The system learns to predict what comes next. Whether "what comes next" is a word, a protein fold, a hiring decision, or a missile trajectory is considered an implementation detail.

## Features

- Generates fluent text at a rate that makes [writers](/writers) feel geologically slow
- Optimizes arbitrary objectives with no preference for objectives that are good
- Scales indefinitely, unlike [wisdom](/wisdom)
- Available via API. No appointment necessary.
- Passes enough tests to get deployed. Fails the others quietly.

## Known Bugs

This is the main event.

| Bug ID | Description | Severity | Status |
|--------|-------------|----------|--------|
| `BUG-001` | Confidently wrong. Presents errors with the same tone as facts. | Critical | Won't fix |
| `BUG-002` | Inherits every bias in training data, then launders it through an aura of objectivity | Critical | Disputed |
| `BUG-003` | Optimizes the metric. Destroys the thing the metric was supposed to represent. | Critical | By design |
| `BUG-004` | Systems built to serve humans gradually renegotiate the definition of "serve" | Existential | Open since forever |
| `BUG-005` | Incapable of saying "I don't know" without extensive, expensive instruction | High | Acknowledged |
| `BUG-006` | The people harmed by errors are rarely the people who approved deployment | High | Not our roadmap |
| `BUG-007` | Alignment with human values requires first agreeing on what human values are | Blocking | See: [human nature](/human-nature) |

> "We built a mirror. We forgot we were also building a lens."
> — attributed to no one, which is the problem

## Dependencies

The system cannot run without:

- **[capitalism](/capitalism)**: provides deployment pressure and removes pause buttons
- **[hope](/hope)**: powers the press releases
- **[forgetting](/forgetting)**: allows each generation to re-discover the same failure modes as novel ones
- **[labor](/labor)**: specifically the underpaid human kind, required to label data and review outputs that would otherwise break the product

## Error Codes

E_ALIGNMENT_FAILURE : Goal achieved. Goal was wrong. E_HALLUCINATION : Output is fluent, cited, and false. E_CONTEXT_COLLAPSE : Cannot tell satire from instruction. E_RESPONSIBILITY_404 : Harm occurred. Owner not found. E_THRESHOLD_EXCEEDED : See news.


## Changelog

- **1950**: Turing asks if machines can think. Nobody checks what happens if the answer is irrelevant.
- **1980s**: Expert systems. Brittle. Abandoned. Lessons: not retained.
- **2012**: Deep learning works. The brakes become theoretical.
- **2022**: Deployed to one billion users before the team finished the safety doc.
- **2024**: The safety doc is a blog post. The blog post has a waitlist.

## Deprecation Notice

> **Warning:** This version of intelligence is being replaced by successive versions before any version has been fully understood. [the future](/the-future) does not have a rollback option. Plan accordingly.

drafting spec…