Thread by @EmilWallner

François Chollet’s core point: We can't measure an AI system's adaptability and flexibility by measuring a specific skill.

With unlimited data, models memorize decisions. To advance AGI we need to quantify and measure ***skill-acquisition efficiency***.

Let’s dig in👇

In the 1970s, many thought that chess reflected the entire scope of rational human thought. Solving chess with computers would lead to major leaps in cognitive understanding. But after IBM's DeepBlue, they realized they didn't have a better understanding of human thinking.

IBM's DeepBlue is not intelligent, but we consider humans who master chess intelligent. It's because we associate chess with a meta-skill, an aptitude for logical tasks such as math and reasoning. We often anthropomorphize AI systems in a similar way: task mastery = AGI.

Similarly, Chollet argues that AlphaZero, DeepMind's program synthesis engine for board games, is not flexible and general. He compares it to a hashtable that uses a locality-sensitive hash function. With unlimited simulation, you can map board positions with actions.

Chollet thinks our task-centric idea of intelligence is a bottleneck to advance AGI. Instead, we should adopt Hernandez-Orallo take:

“AI is the science and engineering of making machines do tasks they have never seen and have not been prepared for beforehand”.

When humans are tested in different cognitive tests the results correlate with each other. This indicates that we an underlying meta-skill to learn skills, the g-factor. These are the abilities Chollet want to measure in the context of AI, starting with Broad Generalization.

In section II.2, Chollet formalizes his central idea:

"The intelligence of a system is a measure of its skill-acquisition efficiency over a scope of tasks, with respect to priors, experience, and generalization difficulty."

Below is an overview of skill-acquisition efficiency. A task could be chess, the situation is a board position, and the skill program is a frozen chess engine. Think of the intelligent system as a program synthesis engine for different tasks.

Chollet uses Algorithmic Information Theory to quantify the programs and interactions, expressed below.

'The intelligence is the rate at which a learner turns its experience and priors into new skills at valuable tasks that involve uncertainty and adaptation.'

To avoid local-generalization systems that artificially "buy" performance on a specific task, Chollet restricts priors to 'Core Knowledge' found in developmental science theory: such as elementary physics, arithmetic, geometry and a basic understanding of intentions.

Chollet created a dataset according to the best practice he outlined, the ARC. The dataset mimics the abstraction and reasoning portion in an IQ test (fluid intelligence).

Examples:

The ARC dataset has 400 training tasks and 600 evaluation tasks.

Key features:

- Only novel tasks in the evaluation set
- Highly abstract
- Similar to human IQ tests
- 3 demonstrations per task
- Fixed/limited training data
- An explicit set of priors

github.com/fchollet/ARC

'The Measure of Intelligence' contradicts the compute-is-all-we-need narrative, adds historical context, and is accessible to a broad audience.

Hats off to @fchollet for creating clear and actionable suggestions to advance AGI. It's a must-read!

arxiv.org/pdf/1911.01547…