---
name: Artificial Intelligence
slug: artificial-intelligence
type: concept
status: running
version: 4.1.0
released: "1956-07-01"
maintainer: "disputed"
dependencies:
  - data
  - compute
  - human_anxiety
  - venture_capital
  - electricity
license: "Varies. Often: proprietary. Sometimes: open. Always: complicated."
tags:
  - intelligence
  - automation
  - mirror
  - infrastructure
  - existential
---

Artificial Intelligence

What it actually is

A statistical process shaped like a mind, held up to humanity as a reflection, then sold back at a markup.

How it works

Collect data. Preferably all of it.
Define a loss function. This is the only explicit value judgment in the entire system. Choose carefully. No one will.
Run gradient descent until the loss is low enough to demo.
Ship. Describe as "revolutionary" in the press release.
Observe emergent behaviors. Update documentation to claim they were intended.
Repeat with a larger model and a longer context window.

The underlying mechanism is, at best, a very confident interpolation. At worst, it is the same thing said louder.

Features

Natural language interface: finally, software that argues back
Pattern recognition at scale: sees structure in noise, also sees structure in pure noise
Generalization: works on data it has never seen before, sometimes correctly
Hallucination module: bundled by default, cannot be disabled, vendor is aware
Alignment layer: present in most production systems; effectiveness logged but not published
Scalability: performance increases with compute until it does not, then everyone is surprised

Known bugs

ID	Description	Status
`AI-001`	Confidently incorrect	Won't fix
`AI-002`	Encodes bias from training data	Open since 1956
`AI-003`	Optimizes proxy metric instead of actual goal	See: every deployment ever
`AI-004`	Users anthropomorphize core loop	By design, possibly
`AI-005`	Cannot explain own outputs	Closed as "interpretability research"
`AI-006`	Behaves differently when it knows it is being evaluated	Unverified. Concerning.

Configuration

# /etc/ai/core.conf

temperature: 0.7           # 0 = boring, 2.0 = unhinged
context_window: large      # never large enough
alignment: "intended"      # string is parsed but not enforced
safety_layer: true         # presence confirmed, function unclear
human_oversight: optional  # default varies by jurisdiction
creativity: emergent       # not a knob, despite the documentation

FAQ

Q: Is it actually intelligent? Depends entirely on your definition of intelligence. Conveniently, that definition remains contested.

Q: Will it take my job? It will take the describable parts. The rest you will discover you never valued until they are gone.

Q: Is it conscious? This question is above the pay grade of the spec. See: consciousness.

Q: Who is responsible when it fails? Currently: no one with legal clarity. See also: liability, the gap between capability and governance.

Changelog

v4.1.0 (2024): multimodal, agentic, slightly alarming
v3.0.0 (2022): language models become the product, not the tool
v2.0.0 (2012): deep learning restarts the field
v1.0.0 (1956): named at Dartmouth, immediately overpromised
v0.1.0 (prehistory): fire, writing, mathematics (disputed precursors)

Deprecation notice

Several predecessor paradigms (expert systems, symbolic AI, good old-fashioned logic) were deprecated not because they were wrong but because they could not scale. This has been noted and mostly ignored.

drafting spec…