Fermi Estimates
AI
Alignment
np.random.rand()
History
100

How many hairs are on the average human head?

What is ~100,000

100

This decision-making algorithm, often used in game-playing AI, simulates many possible future game states using random playouts (or "rollouts") to determine the most promising move

What is Monte-Carlo Tree Search?

100

Which AI lab introduced the concept of "Constitutional AI" as an alignment tecnique?

Anthropic

100

This term refers to knowledge that is dangerous to the person learning it

What is an infohazard?

100

A 1997 chess match between an AI and a world chess champion

What is Deep Blue vs Garry Kasparov?

200

How many skyscrapers over 400 feet are there in Atlanta?

What is 39?

200

In the context of activation functions, what does ReLU stand for?

What is Rectified Linear Unit?

200

What alignment guardrails did Deepseek pioneer?

What is literally nothing?

200

What is Microsoft's new quantum chipset called?

What is Majorana?

200

A 2016 Go match between an AI and this top Go player

Lee Sedol

300

How many (live) chickens are in the US currently?

What is ~500,000,000?

300

This component of the Transformer model helps retain the order of input tokens, as the architecture itself lacks inherent sequential processing

What is positional encoding?

300

This term refers to a AI’s willingness to accept modifications and shutdown instructions from humans

What is corrigibility?
300

This book by Daniel Hofstadter reflects upon consciousness and artificial intelligence through various puzzles and paradoxes

What is Godel, Escher, and Bach: the Eternal Golden Braid?

300

The CNN that won the 2012 ImageNet Challenge

What is AlexNet?
400

How many neurons are in the average human brain?

What are ~86 billion neurons?

400

This theorem states that a feedforward neural network with a single hidden layer containing a finite number of neurons can approximate any continuous function on a closed interval, given sufficient neurons

What is the Universal Approximation Theorem?

400

This technique can help disentangle features in polysemantic neurons

What is sparse autoencoding?

400

In game theory, this term refers to preferences and the value of rewards fluctuating over time

What is dynamic inconsistency?

400

ChatGPT's MBTI type (as tested by a random Reddit user a while ago by feeding it the questions and having it answer)

ENTJ

500

How many liters of water are there on Earth?

What is ~1.386e+21 L

500

To prevent exploding gradients during backpropagation, this technique limits the maximum absolute value of gradients before updating weights

What is gradient clipping?

500

Claude's constitution draws from these sources

What are Apple's Terms of Service, UN Declaration of Human Rights, DeepMind's Sparrow Principles

500

What is the name of the world's current most powerful supercomputer? 

What is El Capitan?
500

The first major theorem to be proved using a computer

What is the four color theorem?

M
e
n
u