I like building mental models of (interesting) people.
And the best instruments my partner and I have found for that goal are probing questions – asked with appropriate context, seeking to understand and not to judge (it helps to be a moral nihilist or constructivist), answered after a period of introspection, in a setting where people feel safe enough to be vulnerable. As a bonus, sometimes you learn things about yourself. This is partly what inspired the Humans of AI series.
Devi Parikh’s and my labs at Georgia Tech have a long-running tradition of an annual Lab Retreat. At the start of the academic year, our labs spend a long weekend together, typically in a small town with access to nature and a conference room, and little else. The goal of this retreat is to pause, step back from our daily grinds, think about the long term, and really get to know ourselves and our colleagues.1
I draft 2 new questions every year; everyone gets 2 weeks to mull over their thoughts, and we present our answers at the retreat. Answers to these questions generally elicit more questions, and the discussion expands to fill the entire weekend.
Over the years, people have told me that they enjoyed the questions and their associated context write-ups. Perhaps others will find value in reading these. In that spirit, here they are.
Caveat: these were written for an audience of machine learning researchers that tends to view the world through the perspective and vocabulary of machine learning and AI.
2018
Q1. Relatively speaking, what are you good at?
Consider an appropriate set of peers for you (e.g. vision/ML researchers 1-3 years out of their PhD, or senior/junior PhD students at GT, or MS students at a top-10 US university, or post-undergrad interns/folks). To be clear, these are just suggestions; you define your peer set; I don’t.
Now, within that peer set, what skill/knowledge/talent/quality do you believe puts you in the top-X percentile?
To be clear, that skill could be something unrelated to your professions or could be something that your peer set isn’t even trying to develop — e.g. among senior PhD students at GT, I’m the fastest runner. Perfectly acceptable answer.
It would be nice if the two (peer set and your skill) were aligned, but unaligned ones are acceptable.
Q2. When have you felt the most and least privileged?
Privilege in its various forms (where/when you’re born, who you’re born to, what opportunities came your way early on without you particularly trying, who you knew, who your parents/advisors/mentors/teachers were, what resources you had access to, which people you had access to, what someone taught you was within reach, etc.) has consequences. As does lack of privilege.
Note: the question isn’t about whether the world should be fair. That’s out of scope. The only thing we’re asking is – recount an experience where you felt most and least privileged. Note: not random luck/chance. Think of structural advantages/disadvantages and not lottery tickets.
2019
Q1. Name something you have recently (say in the last 1 year) changed your mind about.
Updating your beliefs is hard. We’d like to believe we’re rational Bayesians — we come in with priors, observe the world, and update our beliefs to incorporate the new evidence. Everything we know about cognitive biases suggests otherwise – from confirmation bias affecting how we consume new evidence depending on whether or not it aligns with our pre-existing views to entrenchment when presented with evidence that contradicts our current beliefs.
As social beings, we should praise and reward acts of changing our mind and updating our beliefs. Hence this question. Can be anything (small or big).
Q2. What are the latent variables that explain you and your trajectory?
You are a high-dimensional object. You have beliefs, desires, drives, quirks, principles, irrationalities, strengths, weaknesses, etc. Your trajectory (where/to-whom you were born, the choices you’ve made in life, the things you’ve done, the things that have happened to you, etc) is also a high-dimensional object with high degrees of randomness.
Consider creating a generative model for you and your trajectory. What are the latent variables? The low-dimensional variables that explain you and your trajectory. What drives you? What explains your state and how you act? Why do you do what you do? Why have you been where you’ve been? Why will you be where you'll likely be?
2020
Q1. What is some of the best advice you’ve ever gotten or given?
Reinforcement Learning (RL) is inefficient. Learning by trial and error (particularly without an accurate model of the world) takes a lot of time. What the framework of RL does do exceptionally well though, is to force us to confront the fleeting nature of time. Every day, we sense, we update our models (hopefully), we act; but whether we make progress towards our goals or not, time marches on. There is no way to go back and try different things. It’s the next day every day.
What’s worse, there does not appear to be a manual for this game of life. No explicit rulebook to consult, no dataset of experiences in identical circumstances to learn from. Our only hope appears to be that we are members of a population with generational transfer of accumulated knowledge. We get to create increasingly-less-wrong manuals to pass down.
So tell us – what piece of advice has made its way into your life manual? What would you want to contribute to the community’s manual?
Q2. What do you find yourself thinking about the most?
If you’ve been around the lab for a while, you know there are a few recurring themes in our questions – getting to know ourselves, getting to know our peers, and generally building accurate models of the self and the world.
In the spirit of getting to know ourselves better, tell us about the thoughts that occupy your mind. We understand you have insight only into your conscious thoughts and (by definition) not your subconscious thoughts. But even among your conscious thoughts, to the extent possible, try to exclude the intentional thoughts and focus on the “unintentional” thoughts that occupy your mind. Essentially, don’t tell us what you actively choose to sit down and think about (“How can I get embodied agents to navigate?“). Try to characterize what you find yourself mulling about without intending to. Obviously the boundary between the two is blurry because what you force enough times becomes a habit.
So tell us – what’s usually on your mind without forcing it to be there? What kinds of things pop into your head as you go about your day? Basically, if you’re a fly on the wall in the room of your mind, what do you observe?
2021
Q1. What are you optimistic about?
I got back into reading this year. Particularly, popular science and science fiction. I would say David Deutsch’s pair of books (Beginnings of Infinity and Fabric of Reality) have changed my worldview in ways I didn’t think was possible.
The core thesis of Beginnings of Infinity is one of profound optimism — problems are solvable, solving problems creates new problems, which in turn are solvable, and so on. We are at the beginning of an infinitely long journey in our understanding of the universe and we will always be at the beginnings of infinity. What we seek are increasingly better explanations to questions of “why”.
So in that spirit, tell us — what are you optimistic about?
This could be something small, something personal, or could be something at the scale of humanity. Tell us a problem that you are optimistic is solvable.
Q2. What role does regret play in your life?
I also read From Eternity to Here by Sean Carroll, which is about the nature of time. From Newtonian mechanics to thermodynamics to relativity to quantum mechanics and field theory, Sean presents an overview of different theories of physics, particularly how they view the concept of time. Turns out, we don’t understand why there is an arrow of time. All of our fundamental (and empirically successful) theories of physics are time-asymmetric, i.e. do not require an arrow of time (except famously, the 2nd law of thermodynamics, but thermodynamic arrow of time need not correspond to the perceived arrow of time; more on that on wikipedia). We appear to remember yesterday and not tomorrow, but it’s not clear why that should be the case. We have some explaining to do.
Sean Carrol’s proposal is that we should view time like space. There is no natural/canonical coordinate system for space. But when we are close to a large spatial object (e.g. Earth), we can easily orient ourselves in reference to that object (e.g. “up” is away from Earth). Sean’s claim is that the reason time appears to have an arrow is because we are close to a large temporal object (Big Bang) and we think of the future as away from the Big Bang. If we were sufficiently far from the Big Bang, we would again be back in a state where we would not have a natural direction.
Fascinating idea.
But right here right now, what we experience is the fleeting nature of time – time appears to have a direction and it appears to march incessantly, whether we like it or not. It is always the next day, whether you like it or not. There is no going back to yesterday. There is only now and a different now at every instant.
And that right there forces us to deal with regret. Online machine learning has a precise definition of regret – utility/reward/value of optimal actions in hindsight minus value of actions taken by us.
So tell us – what role does regret play in your life?
What does regret feel like to you? What kinds of things do you end up regretting? Do you end up regretting much? If so, how do you deal with it?
2022
Q1. What was your last epiphany?
The word ‘epiphany’ appears to have a specific religious meaning in Christianity, but I am interested in the broader secular sense of the word — that “aha” or eureka moment when something clicks into place, when a perspective shift happens, when a part of the world finally makes sense because it fits into a coherent mental model.
I have been listening to Sean Carroll’s podcasts this year and reading a number of his books. One of his guests related their philosophical journey to drug addiction – the first few epiphanies were life-changing experiences and came easy. And then the hits became harder to come by, requiring more effort. I thought the analogy was unnecessarily somber, but perhaps functionally accurate.
We can all remember the first few hits that got us doing what we do now. That first time something really clicked and you felt a sense of wanting to learn more. It’s good to keep that feeling going.
So tell us – what was your last epiphany? When was the last time you went “aha, that makes sense”? What did you finally understand?
Q2. What hurts right now?
John Green has a novel called The Fault in Our Stars, which describes a fictional novel called “An Imperial Affliction”, which contains a quote I like – Pain demands to be felt.
Obligatory self-congratulatory jokes about “how meta” aside, I really like that quote.
There is a certain clarity that pain provides. It quiets the cacophony of distractions and drives focus. You deal with the pain or you don’t deal with anything. You can’t procrastinate or negotiate with pain. You can’t delude yourself about the state of the world. No more uncertainty or doubt. Pain transitions the important into the urgent. If life is an episodic roll-out for an agent being trained with reinforcement learning, pain is a negative reward sufficiently large that gamma be damned.
To be clear, I don’t mean to glorify pain. Pain builds character. More accurately, overcoming pain builds character. But there isn’t anything intrinsically of value in pain itself – no pain for pain’s sake. “Try to minimize gratuitous pain for sentient creatures” is about as close to a moral axiom as I can get.
So, without condemning or condoning, without approving or disapproving, tell us – what hurts right now?
As you read this sentence or as you sit down to think about writing your answers, what do you think is providing you negative pleasure in your life? Could be something small and inconsequential, could be something more fundamental, could be something that builds character or otherwise has an expected positive return in the future, could be something utterly meaningless. Feel free to interpret the question as you’d like and share to the extent you are comfortable.
2023
Q1. What advice are you unable to follow?
One common metaphor for life is a massively open-ended video game. And what a bizarre game – one life, no respawns, saves or retries, no mission, no manual or help page, but beautiful graphics, intricate physics, and wild storylines. It’s not surprising then that we’re constantly looking for and passing tips around for how to navigate the game – we are crowdsourcing the manual.
In machine learning, we understand the value of imitation learning and the massive sample-efficiency gains that they provide over learning from trial and error. However, we’re also aware of the concept of Impossibly Good Experts. These are experts with access to information that the imitator does not possess. In navigation problems, shortest path oracles have access to a map that an agent does not. In salary negotiations, recruiters have access to statistics of offers being made that a single candidate typically does not. Such privileged information makes these experts essentially inimitable – they can tell a learner what the optimal action at each time is but the mapping from what the learner knows to the optimal action is unlearnable.
Teaching is fundamentally an exercise in empathy. And ironically, having privileged information or fine-tuned world models (that don’t include a specific learner) can make impossibly-good experts apathetically-poor teachers.
What makes this particularly challenging is that from the perspective of the learner, it is hard to tell the difference between impossibly good advice and bullshit (or noise).
So tell us – what advice do you find hard to follow?
It could be because it’s impossibly good wrt your mental model (but how would you know?) or it could be because it’s unempathetic (or off-policy) wrt you, or it could be you see it as straight up BS (but it is peddled frequently) or something else entirely.
Q2. Have your revealed preferences ever diverged from your stated preferences?
I am currently reading The Laws of Trading by Agustin Lebron. It is not a book about trading or finance (though it does use a lot of trading examples) but rather a book about how to think and make decisions. I am finding it most similar to Julia Galef’s Scout Mindset, which is also about seeing the world clearly and making good decisions.
The first law of trading is – know why you are doing the trade you are doing.
It might seem like we are not actively “trading” on a day-to-day basis, but we are. We trade our time to do the things we do, and our actions have opportunity costs (i.e. we could be doing other things with our time). So this law is really asking for introspection into our motives.
Unfortunately, introspection is hard because of self-deception. Some motives sound nobler than others and we want to cast ourselves (and our motives) in a positive light.
As Richard Feynman famously said: “Science is a way of trying not to fool yourself. The first principle is that you must not fool yourself, and you are the easiest person to fool.”
Feynman is articulating the problem, and also providing a hint for the solution – conducting experiments to infer what you actually believe, i.e. applying the scientific method on yourself.
The field of economics makes a distinction between stated preferences and revealed preferences. The former is what an agent says is important to them, the latter is what is inferred to be important to the agent based on their actions. In machine learning, this would be the difference between an agent’s description of its policy vs the reward function learned with inverse reinforcement learning on the rollout of their actions.
So tell us – when was the last time you discovered something about yourself by observing your actions rather than introspection?
There is a second part to the retreat that involves making falsifiable predictions about the future, but I’ll cover that in a different post.