Embedded Files
Weekly Learning Goals
Weekly Learning Goals
Understand current discussions on how to align AI algorithms' goals with human values
Discuss how AI goals should be aligned with the Common Good.
This week, we're investigating some fundamental ethical dilemmas that artificial intelligence system developers face. As we've seen, AIs are special because they create models to optimize ways of achieving various objectives. Developers create AI to recognize images with cancerous structures, to make more profit in financial transactions, or to maximize the time you spend on social networks. Behind such diverse applications, all AIs are structured around objective functions that determine their operation. Thus, one of the fundamental questions from the developers' perspectives is what principles we can use to design AIs to always aim to benefit human beings. This will be our challenge for the week.
Paperclips
Game Time!!!
Game Time!!!
I know you have been working hard, so I wanted to give you a break :) Your first task of the week is to play this game for as long as you want to. The intention is for you to play long enough to understand what the game is about and how it might connect to this week's topic. You will get the most out of this experiential activity if you just play the game without looking for online descriptions or commentaries about it. We will have plenty of time to talk about it in class.
To make things even funnier, if you beat my score, I will waive the next FA for you.
Once you are done playing, please answer these brief questions.
Reading
Consider While Reading
As you read this (short) chapter, please pay close attention to how the author frames the central tension between the power of advanced AI systems and the challenge of keeping their goals aligned with human values. Ask yourself: What makes alignment different from traditional questions of programming or control? Think about the real-world stakes—why is misalignment not just a technical bug but potentially an ethical and societal crisis? Come prepared to discuss whether you believe the alignment problem is primarily a technical issue, a governance issue, or a moral issue—and why.
Tuesday Class
Tuesday Class
Reading
Consider While Reading
In the second reading of this week, Stuart Russell, one of the most important voices in the AI technical world, lays out three guiding tenets for the development of AI systems. I recommend that we approach this material in a two-phase manner. Initially, let's fully engage with Russell's arguments to grasp how he aims to align AI goals with human values. Subsequently, we can take a step back to consider the gaps and unanswered questions in his principles. For example, are there other principles that should be considered in the creation of AI systems? What challenges might surface when implementing these principles in specific AI contexts? Moreover, what additional social, economic, and political frameworks are required to ensure that AI systems are in harmony with human goals? It's important to note that we chose this reading not because it provides a definitive answer, but because it illustrates the complexity and subtlety of the issue at hand.
Thursday Class
Thursday Class
Additional Reading (Optional)
1. Gabriel - Artificial Intelligence, Values, and Alignment. A complementary take on how to address the alignment problem from an ethical standpoint.
1. Gabriel - Artificial Intelligence, Values, and Alignment. A complementary take on how to address the alignment problem from an ethical standpoint.
3. Nick Bostrom. The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents ( A seminal work on the philosophical aspects of the alignment problem)
3. Nick Bostrom. The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents ( A seminal work on the philosophical aspects of the alignment problem)
4. Russell and Andrew Ng. Algorithm for Inverse Reinforcement Learning (if you want a more precise mathematical formulation of this algorithm)
4. Russell and Andrew Ng. Algorithm for Inverse Reinforcement Learning (if you want a more precise mathematical formulation of this algorithm)
Page updated
Report abuse