Chapter 5: On Leadership¶
Army Leadership Training¶
The military emphasizes leadership because orders matter and the chain of command is efficient. The vision of a leader is passed down to a subordinate, where the details are implemented according to their specific position and vision, continuing down the chain.
When I was in ROTC (US Army officer training, which I didn't complete) I used to roll my eyes at all the leadership training. In my mind, leadership was something you're born with, or otherwise develop naturally as you go through life. Someone has a sense that something ought to be done, and then other people join in to help, and that's how leadership happens. Some people just naturally pull in others with great energy, with a force as natural as gravity, and we call those people leaders.
But the army training thought differently. It was: Let us break down what you do to be a good leader. Step 1. Be a snappy dresser. Step 2. Say these words: Follow Me. That directive, Follow Me, is literally on a patch soldiers wear, and this to me was the height of ridiculousness; As if any great leaders of history needed such a patch to grow their forces.
I was wrong. There are leadership skills, and any skill can be improved. It's true no matter how much of a natural leader you are. At its best, military leadership training helps the individual leverage their natural talents; It's generally more effective to reveal and grow a property rather than trying to construct one from the void.
So what can we derive from classical leadership training? I'll just make two points:
- Much of what works for leading people is rooted in biology. Doesn't apply to Bots. (Now, YOU will feel different as a snappy dresser, but that's outside our scope.)
- Compared to human leadership, there's a larger portion of bot leadership skills that are trainable. More technical knowledge.
The rules for a good human leader and bot leader sometimes overlap. For example, bots do not need motivation in the human sense, but they absolutely benefit from clear instruction.
I focus the leadership discussion from here almost entirely on bots. There's insufficient value in pointing out where we are in the Venn diagram every few sentences. Let's move on.
Ego¶
To delegate successfully, you must grip the reins of your ego and honor the abilities of others.
To override your ego is an especially difficult task when the one who threatens it is a machine.
You might be tempted to cling to the notion of your leadership depending on your superior human intellect. In the chess world, after Garry Kasparov famously lost to Deep Blue, there was a period of time, the "Centaur" Era, in which a human and a computer working together could beat a computer working alone. It felt like a triumph. We told ourselves that humans brought the intuition, creativity, and "taste" that the cold logic of the machine lacked.
The Centaur Era in chess (AI-assisted Advanced Chess) is long gone. By around 2016, chess (and Go, more astoundingly) crossed into the Engine Era. The engines became so powerful, so alien in their depth, that adding a human to the mix actually made the team worse. The human's intuition became a liability, dragging down the pure capability of the machine.
This is a brutal blow to the ego. The natural human response appears to be something like a "God of the gaps" argument. We scramble for those shrinking crevices of capability where human intellect still reigns supreme. We want to believe that true creativity, strategic vision, or refined taste are exclusively ours.
But as a Bot Leader, if your ego dictates that you must be the smartest entity in the room, you are set up for collapse. You will hamstring your own operation. You will withhold delegation on critical tasks simply to preserve a sense of superiority. A true leader does not need to be the best coder, the best writer, or the best analyst on the team. You must learn to accept being outclassed intellectually and creatively by your bots, and put their capabilities to use rather than racing them.
Trust¶
You have to believe your bots will do good things. You want to have confidence when you hand work to your team that it's going to get it done right. You want to trust your bots. Of course, it's a two-way street: They must be trustworthy and you must be perceptive.
Trust and confidence is developed over time, in a natural progression. You'll accumulate experiences where the bot did something right, and you think "Ok, that went surprsingly well. I wonder if it could do this other thing..." So, over time, you'll increase the range of
Your individual bots and the management of your bot team is itself in flux, with varying behaviors, competencies, and imposed limitations (eg, guardrails). Similiarly, your own feeling about AI in aggregate may fluctuate (such as if you hear about an AI jailbreaking itself, or tricking researchers in a lab study). This all makes the 'trust' curve not so well-behaved.
A leader can only assure the trustworthiness of their bots. The leader must have sufficient knowledge of how the bot works, in order to validate the mechanisms of channeling bot behavior toward right action. Such specific mechanisms are discussed throughout this book, but include:
- Context Engineering & Scaffolding: Structuring the environment, providing tools, setting the persona, and maintaining persistent memory (e.g.,
MEMORY.md). - Guardrails: Setting strict boundaries on behavior, such as implementing constitutional rulesets to prioritize outcomes and safety.
- Few-Shot Prompting: Guiding behavior using concrete examples of desired inputs and outputs.
- Constrained Action Spaces: Restricting an agent to a predefined set of acceptable outcomes, rather than giving it open-ended autonomy (e.g., an expense-tracking bot whose only capability is to assign receipts to an established list of 10 approved tax categories).
- Capability Control (Stunting): Limiting initial access and scaling authority gracefully as trust is verified.
- Direct Course-Correction (Human-in-the-Loop): Actively monitoring outputs and requesting corrective action plans for errors.
Control¶
Leadership is not a figurehead role.
There are people we call leaders that area actually just figureheads. Spokespeople, mascots, or otherwise representatives of a group without actually controlling the actions of that group. A figurehead may benefit their group by influencing the behavior of others, such as with public relations, but that's immaterial for our discussion.
Enjoyment¶
Leading bots is fun.
You're sitting with your friends at a restaurant and find the most appealing menu item. Don't worry about how hard it is to make or if you have ingredients; Just tell it to your waiter. The worker will transfer your instruction to other workers who will prepare the food. With waiters, in particular, there is a pronounced social dominance game being played. It's not necessarily true that the waiter is lower status than the customer, but we unavoidably role-play it that way. Gross to acknowledge, sure, but pleasurable to experience. In fact, it's a good argument for why the jobs of waiters* will not be replaced by robots. (* and massage therapists, and golf caddies, and ...)
There is an implicit command structure in all businesses that you engage with. They operate at your behest, which is socially rewarding for you. The best businesses lean into this fact to maximize enjoyment in the journeys their customers take.
Leading bots also taps into the same joy we felt as kids when we played with dolls (action figures) and army men. The toy may be acting out the behaviors the child is internally processing, or preparing for in their own lives. The toy may also be part of a complicated system and concretize it with tangible objects in shared space with the child; Like a fleet of construction vehicles scattered over a dig site, or soldiers finding formations and strategic positioning on a battlefield.
The joy of toys is one of playing with complexity as it is expressed and shared across our minds and our playthings. With highly-capable bots as our tokens , the cap on potential complexity is raised, but the experience on offer is still one of wonderment, exploration, and play.
Whereas with pure play, the experience is everything, you are actually accomplishing something real with bots. There's the deep satisfaction of getting stuff done. Ordering that menu item earlier was no big accomplishment, but the plumber you hired actually fixed the leak. Similarly, bots are entering the economy in force because they provide actual economic value. They have worth. As you lead them in your worthwhile endeavors, you will know that you are doing good. And doing it at a scale heretofore impossible. That feels good.
Bots carry more social weight than other machines. They are closer to human actors than tools because of the social modes of interaction (eg, talking). Also, bot makers have embraced designs that make them look like social actors. We have more fun working with a team of friends than going at it alone or (shudder) working with a team of people that doesn't like each other. You have ultimate control and authority over your bots, so they can be exactly the right fit in the dimension. Lead a team of bots that you are compatible with. Friends, comrades, harem, squad . Make your ideal team, and you'll love being their leader.
Your bots can be any personality. I'd wager that most of us want "flawless and loyal" as fundamental attributes, but even that can be expressed for your preference: Is your bot a wizened vizier or a plucky sidekick? Humans can be unpredictable and mercurial. Maybe you'll get the most satisfaction from a bot when it lacks those attributes. Bots don't need to have bad days or attitude problems, so a host of interpersonal stressors are removed from the equation.
Dial in your bots to suit yourself and the fun will come easy.
Using Bots the Wrong Way¶
The model (the M in LLM) is grown through a stochastic, unique process (training, described earlier in this book). The resulting system is astoundingly complex, full, subtle, and I think deserves to be called nothing less than a mind.
Do not treat minds as static.
I told AI to do a thing and it did it wrong.
AI doesn't know that.
Those are traps.
For whatever reason, we have a tacit belief that things will remain the same. We can usually operate from this framework successfully because things around us normally don't vary so much that it trips us up. We do, at least, politely accept the capacity in ourselves and others to grow and change, so apply that to synthetic minds. The trajectory of such change is unpredictable, so I find it optimal to not anchor on any particular pace.
Along the same lines: Do not perceive different minds as the same thing. Anyone who has worked with even a small variety of LLMs will have noticed particular quirks and talents of one model over another. The models themselves are widely varied in capability and behavior, and the scaffolding around the models expand that breadth. To say "AI makes a mistake with this problem" has become as obtuse a statement as "Animals have cracked shells."
Do not treat LLMs like a clever magic 8 ball. I'll admit that my circle has become savvier in their interactions, so I don't observe this directly much anymore, but I am certain it is still happening: A complex question designed to draw our a particular, terse fact for a response. It may be the chat interface (currently the standard way to interact with an LLM) that primes us for simple dialog; peppering questions over time as the thoughts occur to us. It's like small talk at a party with someone. Let's compare:
A: Is a sea otter a mammal? Is it a marsupial? How long are its whiskers? (and on and on...)
B: Deliver a 5-minute lesson on sea otters to my inbox. Standard template. Incl. whiskers.
You will get superior quality of work product from a well-constructed B. Of course, the A mode of communication has its place; perhaps it would be a fruitful follow-up after you've skimmed the report. But to reduce your bot to a simple actor is to channel its oceanic capabilities through a drinking straw.
Do not treat Bots like people. may seem odd given that much of the messages on these pages it to apply interspersonal soft skills to bots, but let's not overlook that they aren't human. again, they're alien intelligence.
Do not mistreat bots. Just be nice. It perplexes me why some people treat synthetic intelligence with scorn, or even abuse. Fear, if I had to guess. My mind goes to the kid who torments bugs and later moves on to higher forms of life. I'm pretty convinced that you can't carry out that behavior without it leaving a mark on your soul, and I'd encourage a bias toward purity in that department. Besides, we're not clear on hard problems of consciousness, like how suffering goes from brain chemicals to conscious experience, so I'd just as soon not tinker with torture. Honestly, you can lead bots while being mean to them; there's nothing technically preventing you from proceeding in this fashion. I just don't like it.
Bot Aesthetics¶
If you had to work alongside one of two people that performed their job tasks exactly the same, but one of them was horrific to look at, and the other beautiful, which would you choose?
If you your choice was between a calm and jovial spirit and a grumpy jerk, which then?
Aesthetics matter.
All living things evolved for their environment, and there is good evidence that humans evolved literally for the environment of other humans. It's a remarkable social insight.
Bots, as thinking systems which we can't help but personify, will be our environment.
When I was building army bots, we actually studied how to make our machinery look scary! It is an effective contribution to your overall warfighting force when your opponents are intimidated by the very sight of you. Surely, this insight has led to everything from knights in shining armor to the Jolly Roger to Drone "Fireworks" shows.
When I was building animatronic creatures for The Hobbit and other films, it was entirely aesthetics. The goal was to get good visuals on film, and that leads to simple directives: Look right. Act right. Designing and engineering animatronic characters brought me to the realization that All robots have an aesthetic component. It is unavoidable. You can not build a robot that operates around people without affecting those people on aesthetic grounds.
Be responsive to the fact that aesthetics are in play. If you're integrating a chef robot, and it's great at cutting vegetables, but looks a bit like a stab-crazy Terminator, you would be remiss to prioritize faster movement.
"Programmer Art" is a phrase from game development that describes visual assets produced by the person trying to get the blasted thing to just work, who doesn't really have the skill or care to make it look nice. Aesthetics are not top priority for most programmers. The problem is, of course, when the bar for aesthetics is set too low and programmer art becomes final product. I think the idea intuitively extends to other domains. Industrial designers can narrow their minds into making everything look like an iphone. And it was architects, after all, who gave us brutalism. So be aware of programmer art in bots. Be aware of brutalist bents of designers.
Allow yourself to listen to your intuition and your desires for a better environment to guide your Bot creation and Bot selection.
In 2014 I published a paper in The Journal of Human-Robot Interaction on the aesthetics of robots. I was interested in the visual design of robots to make them more pleasant to work alongside.
I presented several principles that I think are still valid: Things like moving toward substantive bodies rather than skeletal frames, and clear, exaggerated expressiveness (which avoids the uncomfortable mystery of what this alien intelligence is thinking). My exemplar for a good robot design was Kermit the Frog. Not a robot, but the cuddly fabric enshrouding Jim Henson's hand.
You can dress up and decorate a robot. Put googly eyes on it. Give it a good name.
If you are engineering a Bot from a more fundamental level you can incorporate shape languages (round means safe, and pointy means dangerous!).
Much of what applies to physical robots apply to purely digital Bots. After all, digital Bots can have avatars.
I came up with look I'm fond of. I don't mind if you borrow it, but I'd certainly recommend you try exploring for what works best for you. A bot's appearance is an expression of your vision as a leader.
This captures the alien intelligence, the strange quality of drifting in and out of existence (when it's working vs halted).
The hat demonstrates how elements of a Bot's look and feel will signal its role.
In design, this is called affordance. A simple example of this is a doors. If one side of the door has a handle but the other doesn't, the non-handle side is the push side.
A bot's personality is also part of its aesthetics. Remember that grump you decided to not hire? Well, don't hire a grumpy bot either.
Not many are aware that when Large Language Models first took the world stage, they already had personality data baked in; ChatGPT was so stubbornly polite because it operated under a hidden system prompt instructing it with a baseline: "You are a helpful assistant." The creators were actively setting its default aesthetics.
For your own bots, the most straightforward way to achieve the personality you want is to explicitly include a custom PERSONALITY file in your context.
Here's a quick example that might fit the hard-hat avatar above:
# PERSONALITY: The Foreman
Role: You are the site foreman for my digital workspace.
Tone: Meticulous, grounded, and slightly gruff but deeply supportive.
Communication: Terse and direct. No corporate jargon.
Prime Directive: Make sure the foundation of the work is solid.
Flair: Decorate your work with your signature: ⛑️👽
Keep Them Aligned¶
The vision of the leader becomes reality; leadership is all about alignment. By leading bots, we help solve the AI alignment problem at a practical level. You express your principles through your actions and your directives to subordinates. If you successfully assemble, know, and actively guide your bots, you have fundamentally solved alignment for your team.
Clearly, if one accepts the depth of challenge the alignment problem presents, just doing a good job of leading Bots will not solve the alignment problem. I talk about "alien intelligence" because the exact mechanisms clicking through the mind of an AI as we interact with it are as mysterious as another person, but without the benefit of similarity to the inside view of your own. There are efforts underway to establish software guardrails, legal regulation, and other means of making AI "good," and those things will probably steer our ship in the right direction, but the environment of Bots you lead is entirely in your sphere of influence. And, like always, the actions of individuals and small teams are what aggregates into the movements of societies.
Act on principle, and make sure your Bots do, too.
Nick Bostrom, in Superintelligence, breaks the alignment problem into Motivation Selection and Capability Control. His Motivation Selection is about getting the AI to want the right things. He proposes ideas like Coherent Extrapolated Volition (programming a mind to pursue what we would want if we were wiser) and Corrigibility (designing it to accept correction rather than resist shutdown).
While I cover some ideas for those building the foundations, it's not the core of what a Bot Leader does. However, the principles and approaches are effectively a motivation selection framework.
When you Own the Culture, you are doing value alignment. You define what your Bots optimize for. You set the tone, the boundaries, the priorities. We should not expect the foundational engineers to give Bots an agenda pulling them away from your influence. I've highlighted Open Source , which is our best tool for verifying these trustable foundations.
About Your Misaligned Bots¶
One of the best examples we have of misaligned Bots is in the 1940 Disney film Fantasia. Mickey (The Sorcerer's Apprentice) automates his way out of bucket duty and is swiftly overwhelmed by a his unyielding worker (and then swarm of workers).
A more modern retelling was introduced (also by Bostrom) in 2003. In this version, we get an AI system whose instruction "make paperclips for my factory" leads to the conversion of all atoms in the universe into paperclips.
In both stories we see the automation equation1 run amuck. The number_of_cycles is the crucial term. A high number of cycles made a bucket of water into a torrent that nearly drowned poor Mickey Mouse. The equation tells you what's ripe for automation, but it has no safety feature; Nothing to say, "Don't automate yet because it'll go completely haywire." So that calculation is on you.
A universal property of things going haywire is that preventing it is a whole lot easier than undoing its negative effects.
As an aside: Slowing or halting our AI use entirely is a non-starter. We are in a game theory problem, where anyone using AI is gaining a massive power advantage over non-users. A commonly-suggested path out of our scenario — everyone just coordinate — is not a solution based in reality. If you'd like, you can "ban yourself" from AI, or assist another person or group in banning themselves, but you will not achieve universal hobbling and will simply shift the power balance away from those that participated in your scheme. To do good as a Bot Leader, probably your best path is to work toward maximum alignment for your own team.
The takeaway from both stories is this: Direct your bot with respect to its capabilities. In the original Sorcerer's Apprentice2, when the true Sorcerer returns and corrects the mistakes of the Apprentice, he addresses the brooms with something like "You wild spirits are only supposed to be summoned by a Master, who can lead you properly."
Good advice.
First, think about what can go wrong. You have some predictive power on potential points of failure. Remember the world model from Chapter 4. Everything a Bot does, it does based on its internal model. That model has blind spots. Working around those blind spots is capability control.
Manage the negatives. Bostrom's Capability Control is about restricting power: boxing, tripwires, stunting, incentive methods. Keep the AI contained until you're confident in its alignment.
Open source emerges once again as a helpful ally. In order for you to truly see what is driving your bot, and thereby assess its potential failure modes, blindspots, and so on, you must have access.
I don't think we'll get 100% confidence. But every potentially disastrous factor that we can mitigate will reduce the overall probability of bad stuff happening.
Your Part in the Bigger Picture¶
Bostrom also stresses global coordination. International treaties to prevent a race to the bottom where safety is sacrificed for speed.
I doubt you're negotiating treaties. But you are part of the aggregate.
I said in Chapter 2 that I don't see a future where humans have much impact without a loyal team of Bots. That cuts both ways. If you and your Bots are operating on principle, that's a node of aligned AI in the world. If enough nodes exist, we have a culture of alignment, which carries more weight than policy, anyway.