pt 1: tools serve goals

Humans have always used tools to achieve their goals faster. Fire accelerates calorie consumption. Calculators accelerate computation. Lasers accelerate all kinds of things, especially information transmission. Social technologies (i.e. culture) exist too, but that’s a separate essay.

Tools should be appropriate for the goal. A bucket excavator can be used in many ways. If your goal is to dig a large hole quickly, it’s an excellent tool. If your goal is to get strong from digging, it’s a terrible tool. If your goal is to dig a precise hole, it’s probably a good tool used in combination with other tools.

Sometimes people use the same tool for different goals. A child may be encouraged by his teacher to fact-check his math homework with a calculator to catch mistakes, and therefore have a deeper understanding of math. An engineer may already possess a deep understanding of math and use a calculator to accelerate computation, instead of doing the math by hand.

It is possible to be wrong about which goal a tool is serving. In 1986, teachers protested against calculator use in early grades, claiming that the students wouldn’t learn important math concepts if given calculators too early. It would be easy for young students to claim they want to learn math, while always using the calculator to get the answer. They could get the right answer every time and still have no idea how math works! This is also evident in college, when students claim they want to learn and still cheat on assignments. Failure is part of the process of learning. Not lying to yourself is a hard problem! The best answers I have right now are to 1. be aware this is possible and 2. monitor whether you’re actually progressing on the goal you claim to care about.

It is good to use all the tools available to achieve one’s goals. This includes one’s own mind and resources, and also the mind and resources of friends and AIs.

Humans have lots of competing goals. At our most primitive, we have goals to stay fed, hydrated, and get enough sleep. Sometimes goals clash: I will stay up late with my girlfriend on a date because I judge the growth of our relationship more valuable than the loss of sleep. Brian Johnson may judge this scenario differently. Modeling a human as having a single goal is too simple.

Problems and goals are two sides of the same coin. When something is a problem, it is preventing me from achieving a goal. Problems can be any size or proximity: small personal problems (my goal of wearing clean clothes is impeded by my problem of having a coffee stain on my shirt) ranging to vague large distant problems (my goal of no human suffering is impeded by civil war in Myanmar).

Not every goal is worth achieving. We’re all going to die one day; triaging what is important is good. I would like to be the best F1 driver in the world if I could press a button, but above that level of effort, many other goals dominate my attention.

pt 2: AI is the thinking tool

The quality of thinking is variable, not constant. Intelligence of the thinker improves quality. Context also improves quality. The more information you have about your values, likes/dislikes, and life situation, the higher-quality your thinking will be in service of your goals. This also means running experiments to gain information on your preferences is valuable! There are aspects that are helpful but don’t always improve quality, namely structured thinking and spending more time thinking. These can be helpful up to a point, but also fall into failure modes of obscuring what’s important and analysis paralysis respectively.

AI provides thinking on demand. They can do your thinking for you, if you’d like. The quality of that thinking will change based on the aforementioned factors. AI models are getting more intelligent over time. Also, the more you provide the AI with context about you and your life, the better it will understand your situation.

I notice I am confused: I would feel disgusted to use AI to write my wedding vows, and yet I feel perfectly fine using it to write software for silly games or demos. Both are generated outputs of words; why do they feel different? I propose it’s because my goals in these situations are very different.

What goal is the act of writing wedding vows intended to serve? For myself, the goal is to personally become the kind of person that can foster a meaningful relationship with my partner. Yes, this can be supported by friends and AI alike, but this is in service of making me a better partner. If I let the AI do the thinking for me, I would be missing out on an opportunity to honestly wrestle with the emotions I feel towards someone.

What problem am I wrestling with when deciding on which career to pursue? Perhaps I feel like I lack high-quality information to inform my decision. AI is excellent at researching and structuring information. I’m very happy to let the AI research and structure information at my behest. This feels very similar to the erstwhile calculator example: I already know how to research and structure information about careers; the AI just does it faster.

Perhaps I have plenty of information about possible career paths, and am emotionally unable to commit to a path. I would possibly be ok letting AI decide in this instance! If we had infinite time, it would be good for me to develop the fortitude to make these choices on my own. Unfortunately, we need to prioritize goals, so maybe it’s best to break analysis paralysis by giving all the information to an outside entity like AI to decide. However, if possible, I would personally want to make the decision myself. Thinking about tradeoffs and perhaps setting up a forcing structure to decide within a period of time seems like a valuable exercise to grow my own mind.

All these judgements are my personal ones. We are living through extraordinary times and trying to make sense of it quickly. Some of us will make different judgements.

Human growth is achieved through effort. If something is effortless, you’re probably not learning. If you seek to become a better partner or a better engineer or a better countryman, seek deliberate practice and failure. Try to understand deeper patterns from each failure. Read deeply and seek to understand. The process of learning how to learn is well-documented elsewhere.

If used well, AI is a tool to sharpen your own competence. AI can help at all of the tasks from the prior paragraph if used thoughtfully! Imagine how hard it would be for you to understand physics without a textbook. Tools can help structure your growth.

If used poorly, AI weakens the skills it is important for you to possess. It remains to be seen in the age of AI exactly which skills are important for humans to possess themselves, and which ones are outsourceable to AI with minimal consequences. You could argue that letting Claude drive you around like a meat mech might lead to good outcomes, but I would find that life very painful. See the fable of the whispering earring for additional reading.

When should we use AI? When it furthers our goals! Be careful to ensure you’re doing the thinking and growing that would maximally serve your goals, but you certainly don’t need to do all of it.

AI is the thinking tool. Which thinking would you actually like it to do?

pt 3: failure modes

There are many failure modes to thinking. They existed before AI and will exist after it. I’ll point out a couple here.

One: overdependence on external sources for decision-making. If I am overly influenced by the opinions of others and fail to wrestle with whether they’re true, I may never build the muscle of thinking about an issue for myself. A small solution to this is writing down my own thoughts and decisions as a first draft before consulting others. This includes asking AI for advice.

Two: analysis paralysis. Some people can struggle with making important decisions in a timely manner, and instead pick a path of inaction. The consequences of inaction can often be worse than the consequences of any chosen path. A solution to this is to precommit to a decision by a certain time. It also can help to auto-default to a decision by a certain date, so no decision is also a decision!

Three: giving each decision too much weight. It is a mistake to fully consider every decision in your life; this eats up your time. Most value in life comes from a few, important decisions: who to marry, what to work on, where to live. It is reasonable to delegate tedious math to calculators if it will speed up the work without impacting your understanding. It is reasonable to delegate your grocery list to AI if it will speed up your shopping while acquiring basically the same food. Humans do this without AI: hiring a cleaning lady is offloading the decisions of what and how to clean a house. A small solution is: try experimenting! Try to automate pieces of your life; notice the results and whether the trade is worth it.

Four: failing to consider long-term effects. Consider if you want to make piles of money on wall street to support your family. That’s your entire goal; you don’t intellectually care about finance and you don’t even like wearing suits that much. Tough! You might get an initial internship with some quick-witted interview answers, but long-term career success in finance depends on a deep intellectual understanding of finance. A deep understanding is an instrumental goal to achieve the terminal goal of money to support your family. (instrumental vs. terminal is a whole can of worms) (Failure to consider long-term effects is why utilitarianism and virtue ethics converge on long-enough time scales. People respond to virtue!)

pt 4: addendum

A couple brief additional thoughts.

The Eisenhower Matrix lays out a decision-making framework for how to bucket tasks. Consider your goals, and the actions you need to take in order to achieve your goals. How many of the actions fall into the delegate bucket? Within the past 3 years, you’ve been gifted a bright intern who never sleeps and works for a monthly salary of $20 (Some workhorse interns even charge $100-$200). This framework is useful at all power levels. You can use it today to delegate tasks to an AI, and a multibillion-dollar CEO can use it to delegate tasks to AI and people alike.

As you use more AI, you may fall into AI psychosis. This is ok, and is common in heavy users. The two main flavors of psychosis are “AI can do anything; my work now means nothing” and “AI can do anything; I need to do x100 more things”. Don’t beat yourself up if you feel this way. Continue sleeping enough and spending time with your loved ones. My hope for you, dear reader, is that you stabilize to the attitude you should always have: “I will diligently work towards my goals.” The power of AI tools should raise our ambitions, letting us reach for harder things and go faster. It shouldn’t change the underlying things we care about. Call your mom soon, ok?