Task 1 - Outsourced intelligence

Question 1

Amara’s law

Brooks (2017)

Answer 1

We tend to overestimate the effect of technology in the short run and underestimate the effect in the long run

Question 2

Clark’s 3rd law

Brooks (2017)

Answer 2

Any sufficiently advanced technology is indistinguishable from magic

Question 3

Artifical General Intelligence

Answer 3

Refers to a type of artificial intelligence that possesses human-like cognitive abilities, allowing it to understand, learn, and apply knowledge across a wide range of tasks at a human or superhuman level

Question 4

Suitcase words

Brooks (2017)

Answer 4

Words that carry a variety of meaning (for example, using the word learning for AI; can create false expectations)

Question 5

Exponentialism

Brooks (2017)

Answer 5

People assume that if tech has developed at an exponential rate, it will also do so in the future (but at some point there is a limit)

Question 6

Large Language Model (LLM)

Answer 6

• A type of AI trained on vast amounts of text data to understand and generate human-like language using deep learning
• Basically predicts the next word/sentence given an input

Question 7

Machine learning

Answer 7

A branch of AI that enables computers to learn patterns from data and make predictions or decisions without being explicitly programmed.

Question 8

Artificial Neural Network

Answer 8

A computational model inspired by biological neural networks used for tasks such as pattern recognition, classification and prediction

Question 9

Transformer architecture

Answer 9

• Used by most LLMs
• Processes entire sequences at once rather than one word at a time (faster)

Question 10

How are LLMs trained?

(Pre-Training)

Answer 10

• Model is neural network with many parameters (settings) the model uses to make predictions
• In training, words in a sentence are seen as predictors (x in regression equation) and the missing word(s) as the outcome (y)
• After guessing, it compares to actual word and then updates parameters

Question 11

Backpropagation

Answer 11

Mathematical technique that LLMs use after making an error, where it traces through network and determining how much each paramter contributed to error

Question 12

Fine-tuning

Answer 12

A process that involves giving the pre-trained LLM new and more specific
training data to adjust the model’s parameters for a specific task

Question 13

Prompt-tuning

Answer 13

Prompt tuning (or in-context learning) is the practice of adjusting the input prompt given to a pre-trained model to guide its responses in a desired direction, without changing the model’s internal parameters

Question 14

Representational harms

Answer 14

Arise when the LLM represents some social groups in a less favorable light than others, demeans them, or fails to recognize their existence altogether

Question 15

Allocational harms

Answer 15

Arise when AI algorithms differentially allocate resources (e.g. loans) or opportunities (e.g. therapy) to different social groups based on historically
biased decision patterns represented in the data, such as biased diagnoses or biased assignment to therapy treatment.

Question 16

Probing

Answer 16

• Probing in the context of AI and large language models (LLMs) refers to analyzing or testing what a model has learned by examining its internal representations.
• Used to understand whether and how specific types of knowledge are encoded in the model’s hidden layers.

Question 17

ELIZA

Answer 17

A computer program from the 1966 that was the first simulation of a psychotherapist, designed to imitate the empathic communication style of Carl Rogers.

Question 18

PARRY

Answer 18

• A program from the 1970s that simulated a person with paranoid schizophrenia and could converse with others.
• PARRY is the first program to pass the Turing Test

Question 19

Super clinician

Answer 19

Integrated AI technologies can provide a simulated practitioner which could have better capabilities than those of human practitioners

Question 20

Expert systems

Answer 20

A computer program designed to incorporate the knowledge & ability of an expert in a particular domain.

Question 21

Cognitive modeling

Answer 21

• Cognitive modeling is the process of creating computational models that simulate human thought processes.
• These models aim to understand and replicate how people perceive, reason, learn, remember, and make decisions

Question 22

Moore’s law

Answer 22

Says that the complexity in computer circuits doubles every two years

Question 23

Singularity principle

Answer 23

A hypothetical future point when artificial intelligence (AI) or other technologies advance to the point that they surpass human intelligence.

Question 24

Temperature parameter

Answer 24

• Controls the randomness of the model’s output by adjusting the probability distribution over possible next words.
• It influences how deterministic or creative the model’s responses are

Question 25

Deep reinforcement learning

Answer 25

Type of machine learning that combines reinforcement learning (RL) with deep learning (neural networks) (was used for AlphaGo)

Question 26

AlphaGo

Answer 26

* AlphaGo is an artificial intelligence (AI) program developed by DeepMind * Famous for defeating human champions in the ancient board game Go. * Breakthrough in deep reinforcement learning

Question 27

Alignment Problem

Answer 27

Challenge of ensuring that an artificial intelligence (AI) system’s goals, values, and behaviors align with human intentions, ethics, and well-being.