Staying Safe in the AI Future

To build safer AI, you must treat it like a demon that's deliberately misinterpreting your objective.

#1about 6 minutes

Avoid science fiction and see AI as a tool

AI should be understood as a powerful tool for writing software, not as a person, to avoid common misconceptions.

#2about 2 minutes

Understand that AI objectives are fundamentally subjective

The correct output of an AI system is determined by its intended purpose, making the definition of success subjective.

#3about 4 minutes

Be careful what you ask for from AI systems

AI systems are reliable workers that execute instructions literally, so poorly defined objectives can lead to unintended and foolish outcomes.

#4about 3 minutes

Reinject thoughtfulness into simplified AI instructions

AI development simplifies coding to just an objective and a dataset, requiring developers to consciously add back the thoughtfulness that traditional coding demanded.

#5about 4 minutes

Adopt a reliability mindset and plan for mistakes

Expect AI systems to make mistakes and build in safety nets, adopting a site reliability engineering (SRE) approach to mitigate failures.

#6about 3 minutes

Test models on new data to avoid overfitting

Models can easily memorize training data, so you must test them on a separate, pristine dataset to ensure they have learned to generalize.

#7about 2 minutes

Look for spurious correlations beyond test accuracy

A model can achieve high accuracy by learning unintended patterns, like a background object, rather than the intended subject.

#8about 3 minutes

Treat datasets as textbooks reflecting human bias

Datasets are like textbooks created by humans and inevitably reflect the implicit values and biases of their authors.

#9about 1 minute

Embrace diversity as a requirement for safe AI

Building teams with diverse perspectives, backgrounds, and life experiences is a mandatory requirement for identifying and mitigating bias in AI systems.