Our Mission

Likely impacts of machine learning

In the near future, ML has the potential to reshape any area of life where we have clear metrics and plenty of data, probably mostly for the better:

• We’ll see better automated traders, better recommendations for what to read or watch, as well as high-quality automated translation and text-to-speech transcription.
• We’ll build control systems that require sophisticated world models, including self-driving cars, autonomous drones, and advanced industrial robots; but also systems that require detailed models of human behavior, such as apps that find interventions that make us happier and healthier, help us learn more effectively, and “help” us decide what stuff to buy.
• We’ll have better tools for semi-automated programming and engineering, and for generating visual, audio and textual content.

Where might machine learning not help by default?

In matters of thought, it is less clear that ML can be of substantial help:

• Can it improve our understanding of the world? Can it help us see more clearly why some cities and countries prosper while others fail, why humans age, or how a bicycle works?
• As individuals, can it help us reflect on our personal goals and how to achieve them in ways that take into account our flaws and limitations? Can it help us become the kinds of people we want to be?
• As a society, can it help us make progress on the problems that we collectively face, such as efficiently aiding those most in need or coordinating the wishes of different groups?

To a limited extent, the answer is “yes” for all of these. ML can reveal patterns in arbitrary data, including data relevant to these questions, so we should expect some support; and there are indirect effects such as automation expanding our leisure time and making our society run more smoothly.

But will ML help substantially in matters of thought, reflection, and understanding? Will it help to a degree that rivals the way we expect it to help with driving cars, and beyond, ultimately resulting in superhuman performance? So far, there is little evidence that it will.

Why might machine learning fail to help us think?

ML works best in domains where we have well-defined tasks (play a game of Go) with plenty of existing data (train on recorded human games) and clear objectives (win the game). None of these are obviously present for supporting thinking:

• What is the task? How do you structure the interaction between a human and an ML system if the goal is to support the human’s thinking?
• What is the data? Thinking happens mostly in our minds. We can record inputs and outputs, but what happens in between is more complex than current ML algorithms can infer.
• What is the objective? We don’t care about thinking for its own sake, but because it helps us form plans or understand ideas that lead to better outcomes. This leads to sparse, delayed rewards, a setting that current ML algorithms are not suitable for.

Why it matters

Why should we want to get better at thinking in the first place? Here’s a partial list of reasons:

In the short term, many of the issues we care about the most—“How can I be healthier?” “How can I find a romantic partner?” “What should I do with my life?” “How can I make more money?” — are seriously constrained by our ability to gather the right information, think through relevant considerations, and make robust plans. We can sometimes get help from experts, but experts don’t exist for all topics, tend to be expensive, and have limited time and ability to deeply consider individual situations.

In the medium term, going beyond issues that can be addressed by helping individuals, we could probably make progress on many conflicts in our society if we better understood our own and others’ values, the space of potential policies, and their likely consequences. While some conflicts may be unavoidable, it seems plausible that there is lot of zero-sum competition that could be replaced with collaboration if all sides could see how. There’s a positive vision of the future where broadly accessible cognitive technology is the tide that lifts all boats and allows us to make considered decisions together.

In the long term, it’s difficult to see a path where we make good use of our cosmic endowment — the long future ahead, the vast space available — without figuring out how to leverage AI to get better at reflecting about what we want things to be like, and how to get there. There may always remain some fundamental uncertainty about what we truly value the most, but it does seem that not all of our value judgments are created equal. Reflecting on our options and plans does tend to help. If the space of possibilities is really big, we’ll need to improve beyond our current abilities, eventually.