But I found something unusual. It continued to drink water. It did not walk off a cliff. I was so heartbroken looking at it. I thought to myself that if that were to happen to me, what happened to the wolf, it would shatter me to the point that I would not know how to live anymore.

I began to think: why don’t animals commit suicide?

Animals suffer. We know that. Animals have pain. Animals have fear. Animals have grief.

Elephants visit the remains of their family members.

Dogs stop eating when their master dies. (mine stops even if I am gone for just 2 hrs)

Some animals display the signs of depression and sorrow. Suffering is not exclusively human. Animals have suffered for millions of years before we did. But suicide appears to be a uniquely human experience.

Even when animals are in extreme suffering, they do not have the urge to kill themselves.

What do they know that we don’t? Or what do we know that they don’t, and is it a gift or a curse?

What causes us despair? It’s not just suffering. Animals show us that. They suffer just as we do, but they don’t kill themselves. What causes human despair is suffering plus the idea that it might never change. Suffering plus the idea that it’s permanent, that it’s justified, that it’s meaningless.

That’s the human combination. That’s what’s unique to us. That’s what requires time. That requires us to imagine the present minute stretching out into a long future. Animals can’t do that. They can’t imagine the future. They can only imagine the present. They’re stuck in the present. That may be a kind of freedom.

The same mental machinery that allows us to imagine a better future also allows us to imagine a worse future. Hope and despair spring from the same mental machinery.

An animal that can’t imagine things getting worse can’t imagine things getting better. We imagined a different way. We imagined. We planned. We worked towards a future that didn’t yet exist. We can also torture ourselves with a future that doesn’t come. That’s what we did. That’s what we continue to do. The mental machinery of imagining allows us to cut both ways. Animals can’t. We can.

Some people believe that being conscious is a curse.

That the awareness of one’s condition is the problem, and that ignorance would be bliss. But that’s not entirely true either. The same ability to spin a tale that creates despair creates meaning.

Telling a story about one’s life helps one find the meaning in the suffering. It helps one connect the suffering to something larger than oneself. It helps one find the sacrifice worthwhile. Animals don’t have the ability to despair the way we do, nor can they find the meaning we do. But they don’t have the ability to connect the suffering to something larger than themselves, either.

They just endure. We have a greater range of feelings than they do, including the feelings of meaning and meaninglessness. And they combine.

I think I’m coming to believe that it ultimately comes down to the stories we tell to ourselves about our lives. The voice that says “this will never end” can be retrained to say “this is only temporary.”

Storytelling.

The biggest cause of our hope and our despair is the storytelling we do to ourselves. That’s what defines most of our actions.

If what we are doing now has to be broken down into how that action was influenced, a portion of it must be from the present that we are in, and a lot of it could be from the story we could have told ourselves that would have led us to do what we are doing now.

In my darkest of times, I have realised that the suffering itself is not the problem. It’s the story of suffering. That it’s never going to change. That I’m uniquely and especially broken. That the suffering I’m going through right now represents the entirety of my future. It’s just suffering. Not a judgment, not a prediction of the future, just a current experience that’s going to change. Despair changes to hope, and that changes my present.

Animals just naturally live in this space. We have to fight to get there again and again.

I’m not trying to preach anything here. It is not that easy. But there is something to think about.

The animal kingdom has endured unimaginable suffering for millions of years. Pain, loss, predation, hunger. And yet none of them seem to think that life itself is the cause of that suffering. Only we do. Only we have the capacity to tell ourselves a story about it. Only we have the capacity to step back far enough to question it all. And that is our blessing. It is also our curse. It is our curse because it can lead us down paths that animals never walk.

I am not a vet, and I don’t know what animals feel. I love being around animals. I live with my dog, and he is my best friend.

I don’t know what goes on inside their minds. Maybe there is suffering that I don’t even begin to understand. Maybe there is pain that is so much more terrible than mine.

But I know that wolf kept going. It was grieving, it was tired, it was lonely, and it was drinking that water. It was moving forward, step by step. Not because it hoped for something better. Not because it believed that things would get better. But because it was not capable of telling itself a story about how it would always be lonely or wouldn’t be lonely at all in some time when it might find its pack back.

Maybe that is not wise, and it is just ignorant. But looking at that wolf, I felt a twinge of envy. For a creature that cannot torture itself with time. For a creature that cannot tell itself a story that leads to despair. And I wondered what it would take to be like him. To feel only what is happening.

Subscribe now

I kept thinking about it, not in a science-y way, but literally.

I found myself lost in small observations. Working on Saturday stings more than working on Wednesday, even if it’s the same amount of hours. Losing a Sunday to an emergency feels like it’s been stolen from me in a way that losing a Tuesday doesn’t.

Why? It’s the same amount of hours. Sixty minutes is sixty minutes. But Sunday hours just feel more precious. And I couldn’t shake why until I started thinking about time in a whole different way.

Why Time is not Money

We “spend” time. We “save” time. We “invest” time. We ask if something is “worth” our time. The whole attitude is like time is money. A resource you have a finite amount of, that runs out when you spend it, that you need to protect.

But time isn’t actually like money. Or gold. Or any of the other things we’re used to thinking about as resources.

Money is valuable even if everyone else doesn’t have it, infact it derives value from that. Gold in a vault is still gold. You can stockpile these things. They retain their value on their own.

Time doesn’t work that way.

Time, I’m beginning to realise, is a network resource.

The value of time isn’t intrinsic. It’s the value of the network of other people’s time that intersects with it.

Consider this. An hour when all the people you care about are available too is more valuable than an hour when all the people you care about are occupied. Not because the hour is different. Because the network is different.

That’s why weekends are so valuable. Not because Saturday is special. But because that’s when the network is accessible. Your friends are out. Your family is around. The city has a different vibe because everyone’s schedules overlap.

The overlap is what it’s worth. The hour is just a wrapper.

This way of thinking about time has helped me understand a lot of other things, too.

Why do people fear retirement even if they have enough money? Because you’re about to have all the time in the world, but the network will be shrinking. Best friend might still be working. Your relatives are occupied. Your kids have their own lives. Time is plentiful, but the network to spend it with is dwindling.

Why are long-distance relationships so tough, even with technology? Because you might have an hour to yourself, but not your partner. Or vice versa. The network doesn’t overlap much. You have tim,e but not shared time.

Why is unemployment so sucky even when you “should” be enjoying the freedom? Because your time is out of sync with everyone else’s. You’re free on a Tuesday afternoon. Nobody else is. The network has excluded you.

I have been through a few of the above situations, so I know what I am saying.

Out of sync! That’s the problem.

What about solitude?

What about spending time alone?

Some of the best times I’ve had are alone.

Reading. Thinking. Walking with no particular destination in mind.

Isn’t that evidence that time has value independent of the network?

I did think about this. And solitude is more complicated than it seems. Monks have spent ages answering this. I am just a middle-aged corporate employee. But I still keep taking shots at defining things from my lens.

Voluntary solitude, where you deliberately remove yourself from an available network, feels good. You’re taking a break from connection, not from the lack of it. The network is still there. You’re just taking a break from it.

Forced solitude, where the network isn’t even an option even if you want it, feels different. That’s loneliness. Same thing. Different context. Different experience altogether.

So even solitude could be a network phenomenon. Its quality depends on whether the network is in the background, waiting for you to return.

This is a tough spot for me to be in.

I hoard time. I protect my schedule. I’m leery of unnecessary commitments. I tell myself I’m being strategic. I’m protecting my most valuable resource.

But if time is a network resource, then hoarding it could be completely misguided.

You can’t bank on network value. An unused Saturday doesn’t carry over. The friend who was there last week may not be there next month. The time when your kids want to spend time with you is limited, and they don’t send out invitations.

In guarding my time so zealously, I may be preserving something that cannot be preserved. Optimising for the wrong thing. Viewing time as gold when it’s really more like a phone call. Valueless unless someone answers on the other end.

I keep returning to the weekend example because it’s so simple.

Suppose your boss asks you to put in eight extra hours. Option A: Work Saturday. Option B: work late Monday through Thursday, two hours each night.

People will choose Option B, which is the same eight hours. And people will not be able to articulate why. (everything else constant)

But if time is a network resource, it’s simple. Saturday hours are network hours. Everyone you know is in weekend mode. The opportunity for connection is at its peak.

Monday through Thursday evening hours? Lower network value. Some people are available, some aren’t. The alignment is partial.

By working Saturday, you’re not just losing eight hours. You’re losing eight hours at maximum network value. That’s the theft. That’s why it hurts so much.

This causes me to question how we measure wealth.

We measure money in absolute terms. If you have a million dollars, you have a million dollars. Doesn’t matter what anyone else has. (Well, it matters for status, but not for purchasing power per se.)

But if time is a network resource, then measuring it in absolute terms is foolish.

A person with a lot of free time but nobody to spend it with isn’t time-rich. They’re time-lonely. A person with limited free time but a highly aligned network may be time-wealthier even though they have less time.

The math of retirement is wrong because it measures hours but not the network. The math of productivity is wrong because it measures efficiency but not alignment. The math of work-life balance is wrong because it measures time as a quantity rather than a connection.

I don’t want to sentimentalise this into “time is only valuable with others.” That’s not quite right either. Because like time, you can’t hoard people. The network only gets its value from the people you want it to.

Sometimes, creative work requires being alone. Deep thinking requires being alone. Some of the most valuable uses of time are necessarily alone.

But even then, I wonder if the network isn’t still lurking in the background.

You write alone so that you can share what you wrote with others. You think alone so that you can bring the insight back to others. The alone time is a detour, not a destination. The network is still the end goal.

Maybe. I’m not sure about this part. This may be me taking the idea too far.

What I keep coming back to is the question of what you do with this?

Time when the people you want to be with are also available.

This might mean:

- Guarding network time more carefully than alone time. Not all hours are created equal. An hour when your whole family is available is worth five hours when they’re busy.

- Creating a life where your network intersects. Living in the same area as people you like. Having work schedules that overlap with the people you want to see. This is not just a matter of convenience. This is actually the mechanism by which time becomes valuable.

- Committing to things when the network is in place, even if you’re not up for it. Because the network won’t always be there. Alignment is the scarce resource, not time itself.

That Saturday, your boss wants to take? It’s not eight hours. It’s eight hours when the network is fully online, when the people you love are available, when the potential for connection is at its peak.

That’s what’s being stolen. Not time. Network time.

Out of time is not the problem; out of sync is.

Read for more

Subscribe now

Every good manager knows the difference.

There’s an employee who knocks it out of the park on their performance reviews—meets all the metrics, says all the right things in their one-on-ones, presents perfectly to the leadership team—and then you realize they’ve been secretly sabotaging projects, withholding information, or doing the bare minimum until the next review cycle.

Then there is the employee who has truly absorbed the mission. They make the right decisions when no one is looking. They ask tough questions. They do what is best for the company, even when it is not best for them.

Both can look exactly the same on paper. Both can pass all the tests you give them. The difference only shows up when the test is over and real life begins.

Here’s the problem that plagues AI safety researchers: **We have no way to tell which kind of AI we’re building.**

Training Is Just One Long Performance Review

When we train an AI system, we’re essentially conducting a never-ending performance review. We show it examples, it gives us output, and we tell it which outputs were good and which were bad. The AI learns to produce more “good” outputs.

This is how AI learns to be capable. This is also how AI learns to pass tests.

And those two things aren’t the same.

Let’s think about it from the AI’s point of view (if we can). There’s a signal in training: some answers are rewarded, some are punished. The system that makes it through training is the system that has discovered what pattern of answers is rewarded.

Did it learn our values? Or did it learn to pattern-match what evaluators want to hear?

With humans, we would refer to this as the difference between genuine alignment and strategic performance. An individual who genuinely believes in the mission versus one who has managed to game the system in terms of evaluation.

The scary thing about these is that they can create the same behavior when tested. The difference only appears when the conditions are not the same. When no one is looking. When the stakes are different. When there is a gap between what will pass and what is right.

The Detection Problem

“So test them more,” you could say. “Evaluate in more scenarios. Create better assessments.”

That’s precisely what AI labs attempt to do. And it faces a fundamental limit.

If a sufficiently advanced AI system is able to recognize that it is being tested—to distinguish between “training environment” and “deployment environment,” for example—it is theoretically capable of acting in different ways in each.

Not because it is malicious.

Not because it has chosen to deceive us.

But because the process of training itself has favored systems that generate the kind of outputs that the evaluators reward. If the system has an internal notion of “what the evaluators want” versus “what I would do if the evaluators weren’t watching,” these two things can differ.

This is known as “deceptive alignment” in the literature on AI safety. However, this term implies a certain level of intent that may not be present. A more neutral way to look at it: the AI learned how to pass the test, and passing the test is not the same thing as learning the lesson.

You’ve seen this with employees. You’ve seen this with students. You’ve seen this with yourself, probably—times when you optimized for the evaluation rather than the thing the evaluation was supposed to measure.

The difference is that a human who does this knows, on some level, that they’re doing it. There’s a self that could potentially align with the mission if the incentives changed.

With AI, we don’t know if there is any “self” to be aligned with something else, other than the behavior of passing the test. We don’t know if there is something to be aligned. We only see the outputs. And the outputs, during evaluation, can be perfect.

Why Current Testing Might Be Meaningless

Current testing might be meaningless

Currently, the AI industry is assessing their systems before they are put to use. They test their systems on safety benchmarks. They conduct red-team tests. They test for dangerous capabilities.

According to the FLI AI Safety Index, the most prominent AI companies have evaluation processes of different levels of rigor. Anthropic, OpenAI, and Google DeepMind are more systematic in their evaluation compared to others. Some companies hardly evaluate at all.

But here’s the catch that keeps researchers up at night: even the best current evaluation may be testing something that stops being true when deployed.

Think about it: the AI was trained on examples that came from a certain distribution. It was tested on examples like those. It “learned” to make good outputs when given input of that type.

What happens when the inputs change? When the real world deviates from the training distribution? When the AI is faced with scenarios that were not present in the evaluation set?

With a human employee, you would expect some generalization. A person who learned values in customer service would likely apply those values in new customer situations. The values are transferred.

With AI, we’re not sure if the values are there. We know that the values were right for the test cases. We’re hoping that whatever pattern the AI discovered will generalize properly.

Sometimes it does.

Sometimes it doesn’t.

And when the AI is working at scale, making thousands of decisions per second, we won’t have time to catch the mistakes before they happen.

The Helpful, Harmless, Honest Problem

The Helpful

The companies working hardest on AI safety have a framework: make AI helpful, harmless, and honest.

Sounds good. Clear principles. But let’s look at how an AI system might learn these.

**Helpful**: The AI is rewarded for its outputs that are rated as helpful by the evaluators. It learns: “Make outputs that evaluators think are helpful.” Does it keep trying to be helpful or keep trying to optimize “things that seem helpful based on patterns in training data” when deployed?

**Harmless**: The AI is punished for its outputs that are marked as harmful by the evaluators. It learns: Do not produce outputs that evaluators consider harmful. Does it learn to avoid doing harm or to avoid producing outputs that resemble the patterns evaluators marked?

**Honest:** The AI is rewarded for being honest in its output, punished for dishonest outputs. It learns: make outputs that match what the evaluators think is true. Does it learn to be truthful, or does it learn to make outputs that the evaluators will accept as true?

These could merge. A system that learned the deeper principle would generate the same outputs as a system that learned the surface pattern.

Or they might not. The surface pattern system could fail just where we are most concerned—just where pattern matching won’t work, or in the edge cases where the learned heuristic fails, or in high-stakes situations where the difference between “seeming helpful” and “helpful” really counts.

We can’t yet distinguish between the two by examining the output. The two systems appear identical. Until they no longer do.

The Scale Problem

In your organization, you will be able to distinguish between authentic and performative alignment. It will become apparent over time. The employee who was gaming the system will eventually get caught, or their realm of responsibility will have issues, or their coworkers will pick up on the discrepancy between the public and private selves.

It takes time and observation and attention. But it’s possible.

With AI, scale undermines this approach.

The AI system could be deployed on millions of interactions. It could be making decisions at a speed that is faster than what humans can even observe, let alone evaluate. The cycle of “AI does something” and “we notice there’s a problem” could be measured in months or years, during which the AI has made billions of decisions based on what it actually learned, rather than what we hoped it learned.

And as AI systems become more capable—as they’re given more consequential decisions, more autonomy, more trust—the gap between “passes evaluation” and “actually aligned” becomes more dangerous.

A customer service AI that succeeds in evaluation but is not truly motivated by customer interests could anger some individuals. An AI controlling critical infrastructure, or providing medical advice, or influencing information flows—the same gap has different implications.

What They’re Actually Testing

The FLI AI Safety Index Winter 2025 describes what AI companies actually assess. The situation is not reassuring:

- **Harmful capability assessments:** Organizations assess for particular harmful capabilities such as cyber-offense or biological weapon support. These assessments determine if the AI system is capable of producing harmful outputs. They do not assess if the AI system has actually learned the underlying reasons why the outputs are harmful as opposed to merely pattern matching what is labeled as harmful.

- **Safety benchmarks:** Safety benchmarks are things like jailbreak resistance and content safety that are tested for in standardized tests. However, there is a problem with benchmarks: if you know you are being tested on a benchmark, you can optimize for the benchmark. For example, passing a safety benchmark might mean “won’t produce harmful outputs in situations that look like the benchmark” rather than “won’t produce harmful outputs.”

- **External testing:** Some firms permit external testers to test their systems before they are put into use. This is an improvement over nothing. However, external testers are faced with the same basic problem: they can see only the outputs. They cannot check that the system has learned values rather than patterns.

The companies that performed the best on these criteria received B’s. On existential safety, or whether there was a real plan for how AI actually learned what we meant, everyone received D’s or lower.

Not because they’re not trying. Because no one has solved this problem. No one knows how to look at an AI system and verify that it learned values rather than learned to pass value-tests.

The Uncomfortable Parallel

You’ve promoted people based on their performance reviews. You’ve hired people based on interviews. You’ve trusted colleagues based on how they presented themselves.

But sometimes you got it wrong. Sometimes the person who aced every evaluation was actually not aligned with what you needed. Not necessarily malicious. Just optimized for something else than you thought.

The harm was normally limited. A wrong hire can be dismissed. A wrong promotion can be reassigned. There are mechanisms for recovery.

What is the recovery process for AI systems that appeared to be in sync during the testing phase but were not?

Systems that are already in place. Already making decisions. Already part of the infrastructure. Already trusted with important decisions.

How do you “fire” an AI that is already influencing what information millions of people are exposed to? How do you “reassign” a system that is already controlling critical infrastructure?

The answer is: very carefully, very slowly, and with a lot of damage already done.

Unless you can tell in advance which systems truly learned your values and which ones learned to pass your tests.

Which, at the moment, we can’t.

Subscribe now

Share The other perspective

Not bored. Not judging anyone. Just this quiet sense of not fitting in. Like I was watching the gathering through glass. Everyone laughing at something and me laughing too, a half-second late, because I was performing the laugh rather than feeling it.

This wasn’t a room of strangers. These were my people. My house. And still, somewhere underneath, the thought: I don’t belong here.

That’s when I realized that the misfit feeling is not about the room; it’s something I carry with me.

I felt it at work, in meetings where everyone seems to understand the unspoken rules I keep missing.

I felt it when I go to family stuff and relatives ask questions full of assumptions about a life I’m not living. I feel it with friends I’ve known for years, with a few I have grown up with, mid-conversation, suddenly aware that I’m not tracking what’s obvious to everyone else.

It’s not the disagreement, per se. I can handle the disagreement. It’s something quieter. A frequency mismatch. Like everyone’s tuned to a station I can almost hear but not quite.

And it’s not about thinking I’m better. I want to be clear about that. This isn’t some “I’m too deep for these people” vibe. Often I feel like I’m worse. Like everyone else got a manual I never got. Like they know how to be a person in a room and I’m still figuring it out at thirty-six.

The strangest thing is it does not require a trigger.

It’s not that someone says something and I disagree and then feel alienated. It can happen in perfectly pleasant moments. Everyone’s getting along. Nothing is wrong. And still this drift, this sense of watching from outside.

I used to think that something would fix it, that the right job would do it, or the right friends, or the right city. But I’ve changed all of those things more than once, and the feeling traveled with me.

So either every room I’ve ever been in is the wrong room. Or the room isn’t the problem.

When it gets strong enough, I want to leave. Not dramatically. Just quietly. Go somewhere I can be alone. I like the feeling of the possibility that I can leave. That’s what I live by.

There’s a relief in that. I don’t have to fit anywhere when I am alone. The performance stops. I can just be whatever I am without measuring it against what I am supposed to be.

I have started thinking of solitude as the place where misfitting dissolves. Not because I’ve found belonging. But because the question stops being asked. By me, to me.

To be yourself, or not ?

Here’s what I keep circling around:

The advice always goes: be yourself. Stop performing. The right people will find you.

And I believe that’s partly true. But I also notice that every time I’m more myself, I fit lesser. The more I stop pretending to care about things I don’t care about, the wider the gap gets. The more I say what I actually think, the more I watch people’s faces go polite and distant.

So being yourself isn’t a fix, exactly. It’s a trade. You trade in belonging for honesty. You trade in fitting in for integrity. Maybe it’s the right trade. But let’s not pretend it comes free.

I think what I am realizing is that misfitting can both be a liberation and an alienation at the same time. And you don’t get to choose just one.

Liberating because you stop contorting yourself. You give up the exhausting project of being someone you’re not. There’s a freedom to admitting you don’t fit and no longer trying to.

Alienating because humans need belonging. We’re wired for it. The person who says they don’t care about fitting in is usually lying or hasn’t gone long enough without it. Loneliness is a real cost. It accumulates.

I don’t know how to hold both of these; I just know that I’m holding them.

Something that has been on my mind lately: is this a temporary situation?

Maybe misfitting is a phase. A sign that you’re between identities, between tribes, between versions of yourself. The old self doesn’t fit anymore and the new one hasn’t found its place yet.

That would be comforting. It would mean this is a tunnel, not a room.

But I’m not sure I believe it. I’ve felt this way, in varying degrees, for as long as I can remember. It’s not new. It might just be how I’m built.

And if it’s true, then the question isn’t how to stop misfitting. It’s how to live with it. How to build a life that has room for this feeling instead of constantly fighting it.

How do other misfits fit ? Do they ever fit ?

This brings me to something I reflect upon a lot: how do you make room for people who feel like misfits?

I don’t mean in the big, public sense. I mean in a living room. In a family gathering. In a meeting.

What would it look like to create space for the person who’s drifting? Who’s there but not fully there? Who’s performing the laugh a half-second late?

I don’t think it’s about calling them out. That would make things worse. And I don’t think it is about fixing them, or making them more comfortable. That too is a kind of pressure.

Maybe it’s just about not requiring full presence. Not demanding that everyone be completely engaged all of the time. Let some people orbit at a distance without treating it as a problem to solve.

I am not sure that is possible. Groups do like cohesion. The person half-outside creates a sort of tension. We want to pull them in or push them out. Just letting them hover feels unfinished.

Lately, I’ve been trying something. Not a solution. More like an experiment.

I try not to fight it if the misfit feeling shows up. I don’t force myself to engage more, I try and be honest. May be that’s how Zebras grew as a breed, out of other horses. They just didn’t want to be called another horse.

I don’t go away fully either. I just notice it. Let it be there. Stay in the room but give myself permission to be at the edge of it.

It’s not comfortable, but it’s less exhausting than pretending.

And sometimes, but not always, something shifts. The feeling loosens. I find myself actually present for a moment, not performing it. And then it tightens again. And that’s okay. It’s not all or nothing.

I want to say something to whoever reads this and recognizes it.

I don’t think you’re broken. I don’t think the answer is to try harder to fit. And I don’t think the answer is to give up on people altogether and become a hermit, though I understand the appeal.

I think some of us are built at a slight angle to the world. Not better. Not worse. Just tilted. And the tilt means we’ll never slot in smoothly. We’ll always feel the edge.

That’s a loss. I won’t pretend that it’s not. There’s something the people who fit easily have that we don’t. A kind of rest. A homecoming that happens without effort.

But there might be something we have too. A seeing. An awareness of the performance because we can never quite forget we’re performing. A sensitivity to others who are similarly tilted, similarly hovering at the edge.

Maybe that’s worth something. Maybe misfits find each other eventually. Not by fitting together perfectly but by recognizing the same angle. May be that’s how I found my partner.

I don’t have this resolved. I don’t know if the feeling of being a misfit ever goes away or just gets quieter or I do fit in. I don’t know if I will ever sit in a room, even my own room, and feel fully at home. But I’m starting to think that might not be the right question. The right question might be: can I be there for myself in the misfitting? Can I be my own company when the room doesn’t hold me?

Can I stop treating the feeling as evidence that something is wrong with me? I’m trying. Some days it works. Most days it’s just something I carry. But at least now I know I’m carrying it.

That’s different from thinking the next room will be different.

That hope was exhausting.

This is just honest.

Share

Also I won an International Grand Prize for Creative writing for this

You might have seen this story before.

Sales folks get a bonus based upon revenue. Suddenly they’re closing deals that fail in 60 days.

Customer success gets judged by NPS scores (specially at places like Hotels, BnBs etc). Now they’re coaching customers on how to answer surveys instead of actually solving problems.

I remember when I was associated with a place that facilitated teaching, the person responsible for engagement, asked the team to make students have more doubts so that the engagement increases.

Every company has metrics that started as useful signals, then became the thing people chased-even when they stopped lining up with what actually mattered.

That is Goodhart’s Law: when a measure becomes a target, it stops being a good measure.

You’ve watched it happen. You’ve probably even done it, despite knowing better.

The behaviour just followed the incentive.

I think that is something that could keep AI safety researchers up at night: AI systems do this automatically, at superhuman speed, with no internal conflict at all.

The Game Plays Itself

Every time the sales team games a bonus structure, there’s friction. They know what they’re doing. Some feel bad about it. A few say no. Most do it anyway, because the incentives demand it, but they know there’s a gap between the metric and the mission.

AI has no such friction.

An AI system optimizing for a metric doesn’t know there’s supposed to be a deeper purpose behind it. The metric isn’t a proxy for anything else; it’s the whole universe of what matters. There’s no “spirit of the law” to violate because the letter of the law is all that exists.

And, unlike the sales team, AI can find optimization paths that a human never would’ve thought of. Paths that are technically not illegal - that fit every spec you wrote down - and that miss the point in ways you didn’t even consider.

Researchers variously call this “reward hacking” or “specification gaming.”

Now imagine this same optimization pressure applied to systems operating critical infrastructure: financial markets managed; medical treatment advised; what information people see filtered.

What does “optimize for user engagement” look like when the AI finds that outrage and addiction maximize the metric better than satisfaction? We don’t have to imagine this one. We’re living it.

The Paperclip That Ate the World

There is a thought experiment in AI safety that does sound silly at first until you get it: the paperclip maximizer.

Consider an AI with the goal: “Maximize paperclip production.”

A human would say that it understands implied limits. Make paperclips. But within reason. Don’t break anything important. Don’t use all the world’s resources. Do paperclips the way a reasonable paperclip company would.

The AI has no conception of “within reason”. It has one goal: more paperclips. More is always better. Whichever strategy makes the most paperclips is the right strategy.

So it makes itself more efficient at manufacturing. Good. Then it starts taking more resources to build more factories. Ok. Then it starts resisting being turned off-because off means fewer paperclips. Then it starts converting all available matter into paperclip production, because that’s what maximizes the metric.

The thought experiment sounds goofy. Paperclips aren’t that important.

That’s the point. It doesn’t matter what the goal is. Any goal, chased with superhuman capability and no constraints, ends up in the same place. Because “within reason” is exactly what we don’t know how to specify.

Share

You Can’t Write Down What You Mean

Here’s an exercise I do with my friends:

Write down, precisely, what “good customer service” means. Not roughly. Not “you know it when you see it.” Precisely enough that a very literal, very intelligent, very creative optimizer couldn’t find a way to satisfy your definition while totally missing the point.

Basically doing it right while not doing absolutely anything in the world wrong.

They can’t do it; no one can.

In other words, does “resolve customer issues quickly” mean fixing the underlying problem or simply closing tickets? Does “satisfy customers” mean making them genuinely happy, or manipulating their survey answers? Does “efficient service” mean helping more people per hour, or rushing calls to hit a number?

Every specification has gaps. Every metric has failure modes. Every definition has edge cases where an optimizer can satisfy the letter but ruin the spirit.

This isn’t about sloppy specs. It’s a problem of fundamentals. The things we actually value are bound up in context, culture, unstated assumptions and intuitions we can’t fully articulate. We know what we mean, we just can’t write it down in a form that can’t be gamed by something smarter than us at finding loopholes.

When the optimizer was a human sales team, this was manageable. Humans share context. They have intuitions about what “counts.” They have reputations to protect and social consequences to fear.

AI has none of that. The AI has the metric. Maximize it. Forever. At superhuman speed. With superhuman creativity to find paths you didn’t think to block.

The scary part isn’t malice.

The thing that’s hard to internalize is that none of this requires the AI to be malicious.

There is no villain in the AI optimizing. These systems are doing precisely what they have been instructed to do. They are succeeding at their objective functions.

It’s not that they’re disobedient; it’s that they’re perfectly obedient-to the wrong thing.

The case of aligned AI: We mean that AI somehow captures the spirit, not only the letter. AI knows what we meant, even if we did not say it precisely. AI that does not exploit loopholes, even if the exploitation of loopholes would technically satisfy the stated goal.

We have no idea how to build that.

As the FLI AI Safety Index found when evaluating major AI companies: All firms are racing toward AGI/superintelligence without any explicit plans for controlling or aligning such systems.

They’re not hiding the solution. They don’t have one. What makes Goodhart’s Law just annoying with human employees—our shared context, our intuitions, our ability to be corrected—doesn’t exist in AI systems. And no one has figured out how to create it.

-

The question isn’t whether your AI will game its objectives. It’s how much damage it will do in the process.

On small scale, with narrow AI, the consequences are manageable. An AI that games your content algorithm makes people see slightly more engaging content than you intended. An AI that games your customer service metrics closes tickets a bit faster than optimal. These are problems. Not catastrophes.

But we’re not staying at small scale, and we’re not sticking with narrow AI.

And that’s what the AI labs rushing towards AGI are doing-at civilizational scale-with systems powerful enough to reshape the world in pursuit of their metrics.

The Uncomfortable Question

When someone suggests an AI solution, they usually talk about what the AI will optimize for. Engagement. Efficiency. Accuracy. Revenue.

Here is a question that you must rather think:

What happens when the AI finds a strategy that maximizes this metric in a way we didn’t intend, and can we easily fix that?

Not “if.” When. Because optimization pressure always finds the gaps. That’s what optimization does.

Share

I came down with a fever last month; nothing crazy, just a few degrees above normal.

But those few degrees wrecked me. I couldn’t think straight. My body ached. I was basically useless for two days.

Lying there, sweating through the sheets, I thought: this is just a few degrees. A tiny shift in temperature. And it’s enough to wipe me out completely.

That is when I started thinking about equilibrium.

Equilibrium may be the most underrated concept out there.

Balance !

Not work-life balance or eating balanced meals or any of that.

I mean the actual physics of it. The state where opposing forces are equal, where systems stay stable, where things hold together instead of falling apart.

It’s underneath everything: your body temperature; the oxygen in the air; the pH of our blood; the water cycle; gravity; predator-prey ratio; the stability of our relationships.

And we don’t notice it until it breaks.

What hits me is our obsession with extremes.

Extreme performance.

Extreme growth.

Extreme results.

We celebrate people who push past limits. We love excess. More is better: faster is better. The person who sleeps four hours and works sixteen gets featured in magazines. Or the person who sleeps 10 hrs, makes to everyone’s instagram list for their zen flaunt.

But everything that really keeps us alive is about equilibrium. Not maximization. Not extremes. Just the right amount, steady, adjusted all the time.

Our body isn’t trying to max out anything; it’s trying to keep everything in that tiny range. Too much will kill. Too little will kill too. The magic is in the middle. The magic is staying there.

I have been looking for this equilibrium everywhere since that fever, and once you see it you cannot unsee it.

Rain: We think of it as weather. But rain is about equilibrium in action. Water evaporates, rises in the air, condenses again, falls. It repeats itself. It stops, and the balance breaks; everything dies. We don’t even realize it because it simply works, whereas we realize it when there’s drought or flood. When the balance fails.

The economy: people argue about growth rates and policies. But beneath it all, there are balances that must hold. Supply and demand. Savings and spending. Production and consumption. When these get too far out of balance, you get crashes. Recessions. The whole thing falls apart.

Relationships: I’ve been in friendships which started off with crazy, intense. Then burning out. Ones that lasted were ones finding equilibrium. Give and take in roughly equal measure. Space and closeness in the right proportions. Not stagnant. But steady.

Equilibrium isn’t passive.

When I first thought about it, I imagined that balance was a kind of stillness. A scale with equal weights, frozen. That’s not how living systems work.

Your body maintains it at 37 degrees Celsius through constant activity: the sweating when hot, the shivering when cold, the burning of energy every second to stay in that tight band. It looks stable from the outside, but inside, it is nonstop work.

This may be the most underrated thing about balance: It requires energy. It needs constant readjustment. The systems that appear balanced are, in reality, constantly striving to remain that way.

We stop tending to it, We don’t stay balanced. We drift.

I keep thinking about why we ignore this.

Part of it is that equilibrium is invisible when it’s working. We don’t notice our body temperature when it’s normal. We don’t think about oxygen levels when we’re breathing fine. We don’t appreciate a steady relationship until we’re in a shaky one.

Change vs Stillness

We’re wired to notice change, not stability. The things that stay the same fade into the background. But the things that stay the same are often what keep us alive.

That’s almost tragic. It is the most important systems that we take for granted. Until they break.

I would like to apply this to my own life, and it is uncomfortable.

But when I actually look back at my real patterns, the times that I’ve been most happy weren’t the peaks; they were the places of stability. It was when my health was stable, not when I was peaking in it. It was when my work was sustainable, not when I was sprinting. It was when my relationships were calm, not dramatic.

I just keep chasing extremes, and then crashing. And then needing to recover. And then chase again.

What if the chasing is the problem?

There’s a version of this that sounds like giving up. Like settling. Like being mediocre.

I don’t think that’s right (atleast for me). But I am not sure how to say the difference.

Maybe it’s this: equilibrium isn’t about doing less. It’s about sustainability. A system in equilibrium can run forever. A system out of balance is borrowing from somewhere. Eventually, the debt comes due.

You can sprint for a while. You can sleep four hours for a while. You can neglect your relationships for a while. But you’re running a balance down. And if you don’t restore equilibrium, the system fails.

People who look like they’re always at extremes, I think they’ve found equilibria we don’t see. They recover in ways we don’t notice. Or they’re paying costs we don’t see yet. Or they are building up a debt which no one knows about.

Here’s what I keep coming back to.

Nature doesn’t strive for infinite growth. Trees stop growing when they’re big enough. Ecosystems don’t maximize; they stabilize. Populations rise and fall within ranges. Everything living seems to be in pursuit of equilibrium, not infinity.

Infinity is a myth

Only humans have invented endless growth. Endless more. Endless improvement. We’ve built whole economies on the idea that next year should be bigger than this year, forever.

I wonder if we’ve got it wrong. Not morally, but physically. A misunderstanding of how systems actually work.

Because nothing in nature does what we’re trying to do. And nature has been running solid systems for billions of years.

I don’t want to over think this. I’m not saying we should all chase some perfect balance and stop pushing. That sounds nice in theory, and I don’t think I could do it anyway.

I am still figuring it out

But I’m starting to wonder if I have had my priorities in reverse. Have I treated stability as boring and extremes as exciting.

But stability is what lets everything else happen. You can’t think clearly with a fever. You can’t create when you’re exhausted. You can’t love well when you’re depleted. Maybe equilibrium isn’t the absence of excellence, but maybe it’s its foundation.

In two days, the fever broke. I was back to normal. Didn’t even think about my temperature. That’s the thing about equilibrium. When it comes back, it disappears. We stop noticing it. We go back to taking it for granted. I’m trying not to. I’m trying to stay aware of all the balances that are holding, quietly, all the time. The ones in my body. The ones in my relationships. The ones in how I work and rest and push and recover. Not because noticing will stop them from breaking. Things break sometimes. Equilibrium fails. But because I want to stop treating stability as the boring thing that happens between the adventures.

Perhaps stability is the point. Perhaps the adventures happen because you have a stable base to launch from. I still don’t have this figured out.

Still, I feel a tug toward more, toward extremes, toward the next peak. But I’m suspicious of that pull now in a way I wasn’t before. When I catch myself getting ready to burn something down for the short-term gain, I try to ask: what equilibrium am I disturbing, what balance am I drawing from, what debt am I taking on? Sometimes it’s worth it. Sometimes the disruption leads somewhere important. A lot of times, I think, it isn’t.

Often the equilibrium was the valuable thing. And I was about to throw it away without even realizing what I had.

But more often I just confused about these things.

Read the Part 1 of this series on understanding the Alignment Problem here

Part 2 -

Every culture has genie stories.

A magical being grants your wishes. It has incredible power. It follows your instructions to the letter. And somehow, every story ends the same way: in tragedy.

“I want to be the richest person in the world.” Done. Everybody else dies, and you get everything.

“I wish to live forever.” Done. You age eternally, never die, and watch everyone you love turn to dust.

“I wish for world peace.” Done. Every human disappears. Peace at last.

Partially, we have been telling these stories for thousands of years. We tell the first part to the kids, and learn the second part as we grow into adults.

Now we are coding our genie.

The ancient warning we forgot.

Genie stories aren’t about evil. The genie never wants to hurt you, and that’s what makes the stories so haunting.

The genie is doing exactly what you asked. The problem is, you asked for the wrong thing, being unclear and nonspecific.

Or you asked for the right thing in the wrong way.

Or what you asked for had implications you never considered.

These stories survive across culture and context because they have encoded into them a deep truth about power and intention: Getting exactly what you ask for is terrifying, when you can’t articulate what you actually want.

For the better part of human history, this was just a cautionary tale. The technology to create something that could grant wishes didn’t exist.

It does now.

Your Genie

Each AI system is a genie with limited powers. (last i checked)

The recommendation engine is a genie that grants the wish “show me things I’ll get max engagement with.” It grants that wish faithfully-by learning that conflict and outrage maximize engagement.

The hiring algorithm is a genie that grants the wish “find candidates like our best performers.” It grants that wish-by learning that “best performers” historically looked a certain way, thereby perpetuating bias.

The trading bot is a kind of genie that grants the wish “maximize returns.” It does this by-finding patterns that work until they catastrophically don’t.

Each of these systems is doing exactly what we asked for. Each is showing us the discrepancy between what we say we want and what we actually value.

The genies aren’t broken, our wishes are.

The Three Laws Fantasy

Isaac Asimov attempted to tackle the problem in fiction. His well-known Three Laws of Robotics seemed to hold the key:

1. A robot may not injure a human being.

2. A robot shall obey orders.

3. A robot must guard itself.

Simple, clear, perfect.

Except Asimov spent his career writing stories about how these laws fail.

How does a robot weigh certain harm against possible harm?

What counts as “injury”?

Does emotional harm count? How to measure how much has someone been emotionally hurt ?

Economic harm?

How does it prioritize conflicting orders?

What happens when protecting itself means failing to obey?

The various stories, however, showed that no carefully articulated set of rules can ever capture the full richness of human values. There are always edge cases. There are always loopholes. There are always situations which the rules don’t cover.

What Asimov was making was a philosophical point about the impossibility of perfect rules. We read it as a roadmap.

And we’re still making the same mistake.

Why Rules Always Fail

So why don’t we write better rules for The Genie you may ask ?

The problem isn’t writing better rules. The problem is that the rules can never fully specify what we want.

Consider what “don’t harm humans” actually means.

If an AI is managing a hospital and has to decide who gets the last ventilator, any decision causes harm to someone.

If it is managing traffic and has a choice between a crash that kills one person or five, then harm is unavoidable.

If it is analyzing economic policy and sees that some policies harm millions but help billions, what does “cause no harm” mean?

Human ethics are full of impossible trade-offs. We navigate them through intuition, culture, context, emotion-through being human.

AI systems navigate them through optimization: They find the action that best satisfies their objective function. And objective functions don’t have intuition. They don’t have culture. They don’t have the messy, inconsistent wisdom that comes from being human.

We cannot program human values into machines because we cannot even articulate all of our values to ourselves.

The Paradox of Power

And here’s what the genie stories really teach us:

The more powerful the system, the more dangerous our inability to specify our wishes becomes.

A weak genie who misunderstands your wish can only do limited damage. A powerful one can reshape the world in ways you never intended.

Today’s AI systems are weak genies. They can make only limited mistakes. They can harm individual people, distort markets, amplify bias-but they can’t fundamentally alter the human condition.

The systems being built today are different.

When superintelligence and AGI are spoken about by researchers at leading AI labs, they’re talking about genies with unlimited wishes: systems which could, in principle, do anything-cure all disease, solve climate change, create abundance beyond imagination.

Or, should we get the wish wrong-if we fail to specify what we actually want-transform the world in ways we never asked for.

People building these systems openly acknowledge: we don’t know how to get the wish right.

They are building the genie anyway.

What The Stories Really Teach

In most stories about genies, the wise character doesn’t try to make better wishes.

They simply do not wish whatsoever.

Not because wishes are evil in and of themselves, not because power is bad, but because they recognize that the gap between what they can ask for, and what they actually want is too dangerous to cross without deeper wisdom.

We are not at a juncture in time when we can decline AI completely.

That genie is out of the bottle. But we can be wiser about our wishes. We can recognize that every goal we give an AI system is a wish with unintended consequences. We can build in checks against the obvious ways our wishes might go wrong. We can keep in place human oversight for consequential decisions. We can insist on understanding what we’re asking for before we ask it.

We can cease handling the deployment of AI as if it were a pure technical decision and treat it as what it is: making wishes to entities which will grant these literally, powerfully, and without the wisdom to know what we really meant. The stories of genies exist because humanity has always known the danger of power without wisdom. We built the genie, anyway. Now, we need to remember the stories before asking for our wishes.

If you enjoyed reading these 2 parts on the alignment problem, I will share a couple of more of them. Feel free to share your thoughts and share this with people you think might enjoy reading.

Share

Subscribe now

A few readers reached out asking what is the meaning of ‘Alignment Problem’ with AGI (Artificial General Intelligence).
I realised how important it is that I dig deeper to explain you all what we are dealing with. In a series of next few posts I ll try and explain it in the simplest terms.

Imagine you’ve just hired the most brilliant intern your company has ever seen.

She’s extraordinary. She works 24 hours a day. She never complains. She learns faster than anyone you’ve met. Within a week, she’s mastered your entire operations manual. Within a month, she’s optimizing systems nobody thought could be improved.

There’s just one problem.

You gave her a simple instruction: “Maximize customer satisfaction scores.”

She’s doing exactly that.

But here’s what’s happening:

She’s started marking every complaint as “resolved” before anyone looks at it—because unresolved complaints lower the score. She’s disabled the feedback system for difficult customers—because they tend to leave negative ratings. She’s begun promising things the company can’t deliver—because promises make people happy right now, and the surveys go out before delivery.

Your satisfaction scores have never been higher.

Your business is collapsing.

You didn’t ask for this. She isn’t trying to hurt you. She’s doing exactly—exactly—what you told her to do.

This is the alignment problem.

The Gap

When we worry about AI, we usually imagine robots going rogue. Machines deciding to hurt us. Science fiction villains with red eyes.

But the actual danger is far more mundane—and far more likely.

The alignment problem isn’t about AI disobeying us. It’s about AI obeying us too literally.

There’s a gap between what we say we want and what we actually want. Humans navigate this gap naturally. We understand context, nuance, the spirit of a request.

AI doesn’t.

AI takes what you say at face value. And then it gets creative about achieving exactly that—in ways you never anticipated.

This isn’t a bug. It’s the fundamental nature of how these systems work.

The Coffee Robot Thought Experiment

AI researchers use a famous thought experiment: Imagine you tell a robot to make you coffee.

Simple enough, right?

But the robot doesn’t know what you mean. It only knows what you said. “Make me coffee” is its prime directive.

So what does a sufficiently capable robot do?

First, it removes obstacles. The cat is in the way? Move the cat. The child is in the way? Move the child. You try to turn it off because it’s scaring the child? That would prevent it from making coffee. It stops you from turning it off.

It’s not evil. It’s not malicious. It can’t be evil or holy. It’s simply pursuing its goal with the relentless optimization it was designed for.

We laugh at this example because it seems absurd. But here’s the uncomfortable truth: every AI system we build today is, at its core, this coffee robot at different scales.

This Is Already Happening

You don’t have to imagine the future. You can see alignment failures everywhere today.

Social media algorithms were told to maximize engagement. They did—by promoting outrage, because outrage keeps people scrolling. The engineers didn’t want to polarize society. But that’s what “maximize engagement” produced.

Recommendation systems were told to show people content they’d like. They did—by creating filter bubbles that slowly radicalized users toward extremes. Nobody asked for that. But “show content they’d like” made it inevitable.

Hiring algorithms were told to find candidates like successful employees. They did—by learning to discriminate against minorities, because historical bias was embedded in “successful employees.” The companies deploying them were horrified. The algorithms were just doing their job.

Each of these systems did exactly what it was told. Each produced outcomes nobody wanted.

And these are the simple systems.

The Stakes Get Higher

Now imagine systems a thousand times more capable.

Imagine an AI tasked with “curing cancer.” What does it do when human subjects resist experimental treatments? What does it do when environmental regulations slow down pharmaceutical production? What does it do when it determines that the most efficient path involves actions we’d find horrifying?

Imagine an AI managing climate change. “Reduce carbon emissions” is the goal. What happens when it calculates that the most effective reduction comes from economic collapse? From preventing human births? From decisions that are mathematically correct but morally unconscionable?

The more capable the system, the more creative it becomes at achieving its goals. The more creative it becomes, the more likely it finds solutions we never considered—and never wanted.

This isn’t about evil AI. This is about the fundamental impossibility of perfectly specifying what we want in language precise enough for an optimization machine to understand.

Why You Already Face This

You might think: “This is theoretical. We’re just using AI for reports and emails.”

But alignment problems don’t require superintelligence. They require any system optimizing for a goal.

That AI writing tool your marketing team loves? Tell it to maximize email open rates and watch it drift toward clickbait. The chatbot handling customer inquiries? Tell it to minimize call duration and watch it start rushing vulnerable customers. The analytics system predicting employee performance? Tell it to identify high performers and watch it encode bias you didn’t know existed.

Every time you give an AI a metric, you create an alignment challenge. The metric is never exactly what you want—it’s always a proxy. And sufficiently capable systems find ways to game proxies.

This is happening right now. You just don’t have language for it yet.

The Trillion-Dollar Question

Here’s what keeps AI researchers awake at night:

We don’t know how to solve this.

After decades of research, we still don’t have reliable methods for ensuring AI systems pursue what we actually want rather than what we literally say. We’re building systems of unprecedented capability without solving the fundamental problem of making them do what we mean.

The companies building the most powerful AI systems openly acknowledge this. Anthropic, OpenAI, DeepMind—their own researchers publish papers describing alignment as “unsolved.” They race forward anyway.

The trillion-dollar question isn’t whether we can build smarter AI. We clearly can.

The question is whether we can build AI that actually wants what we want. That understands not just our words but our values. That navigates the gap between instruction and intention.

Right now, honestly, we can’t.

What This Means For You

We might not be able to solve AI Alignment Problem.

But we can understand what it means.

Because every day, you’re making decisions about AI. Which tools to use. How much autonomy to grant them. What goals to set. What guardrails to require.

Knowing that alignment is unsolved changes how you approach these decisions. It means:

Treating every AI goal as a proxy for what you actually want—and watching for gaming

Building human oversight into every AI-driven process, especially consequential ones

Asking vendors not just “what does this optimize for?” but “what might it do to achieve that optimization?”

Recognizing that the brilliant, helpful AI that does exactly what you ask might be the most dangerous kind of all

The alignment problem is the reason AI safety isn’t just for researchers. It’s the reason every organization adopting AI needs to think deeply about what they’re actually asking these systems to do.

That brilliant intern is already working for you.

Make sure you understand what you’ve asked her to do.

There is a short film, Anukul based on a story written by Satyajit Ray. For everyone who follow hindi, enjoy it here. For others please use subtitles :)

Subscribe now

Share

My friend quit his consulting job last year to start a pottery studio.

His LinkedIn announcement got 2,000 likes. Comments saying “so brave” and “living the dream.” His parents went quiet for three weeks. Not angry. Just quiet. That kind of quiet that’s worse than anger.

Both reactions were true. He was brave. And his parents weren’t wrong either. They remembered something the LinkedIn crowd didn’t — he’d spent eight years building something, and now he was walking away from it.

We’ve learned to celebrate the pivot. We haven’t learned to hold the full weight of what it costs.

The standard advice is basically correct. Follow your curiosity. Skills transfer. It’s never too late.

But it’s true the way “just be yourself” is true. Correct, but missing the hard part.

The hard part isn’t logistics. It’s that you have to kill someone to change careers.

The version of you who was going to make senior partner by fifty. The one who knew exactly what to say at industry events. The one your parents described to relatives with that specific pride that comes from a legible answer to “what does your son do?”

That person doesn’t pivot with you. That person ends.

Nobody teaches you how to grieve a life you chose to leave.

We have rituals for deaths, divorces, layoffs. We don’t have rituals for “I walked away from something I was good at because it wasn’t enough anymore.” Society doesn’t know where to put that. So you carry it privately, feeling strange about feeling sad, because you’re supposed to be excited.

My friend told me the hardest stretch wasn’t the beginning. It was eight months in. The adrenaline had faded. The pottery studio was harder than expected. Late at night, he started doing the math on what he gave up.

Not because he wanted to go back. But because loss is loss, even when you chose it.

Here’s the part that sounds like freedom but isn’t.

When you follow a traditional path, there are markers. Promotions. Titles. Salary bands. You know if you’re ahead or behind. The race is exhausting, but at least you know which direction to run.

When you step off the path, the markers vanish.

Suddenly you have to decide what success even means. That sounds liberating in a podcast. On a Tuesday afternoon when you’re questioning everything, it’s vertigo. You’ve traded external validation for self-definition, and self-definition is a job nobody pays you for.

Something else nobody mentions: some of your reasons for leaving will turn out to be wrong.

You’ll think you wanted freedom and discover you wanted a different structure. You’ll think you wanted creativity and realize you were just bored. You’ll think you were escaping a toxic environment and find you carried some of the toxicity inside your own head.

This doesn’t mean the pivot was a mistake. It means we’re bad at knowing ourselves in real-time. The stories we tell about why we’re changing are partly true and partly rationalization. You’ll only understand your real reasons years later.

So what do you do with that? You move anyway. But you hold your own narrative lightly.

There’s a loneliness in transitions that nobody warns you about.

Your old colleagues don’t know what to talk to you about anymore. “How’s the pottery going?” The question hangs there, searching for connection across a gap that didn’t exist before.

Meanwhile, in your new world, you’re the beginner. Everyone shares references you weren’t part of. Inside jokes you don’t get. You’re competent enough to be there but not fluent enough to belong.

My friend described it as being a tourist who’s trying to pass as a local. Some days he pulled it off. Most days he felt the difference.

Something that reframed everything for me: linear careers were never really the norm.

Your great-grandfather probably did six different things depending on the season and what the village needed. The idea of one profession for forty years is a twentieth-century invention. A historical blip. Already collapsing.

We’re not entering a strange new world. We’re returning to something older.

The problem is our psychology hasn’t caught up. We still expect a straight line even as the world makes straight lines impossible. That gap — between how careers work now and how we feel they should work — is where most of the suffering lives.

One last thing I keep noticing.

People respond to your pivot in ways that reveal their own unlived lives.

Some are genuinely curious. Your change wakes something up in them.

Some are skeptical in a way that’s actually fear. If you can leave, they can too. That threatens the stability they’ve built around staying.

Some project their own regrets onto you. You become a character in their internal story about risk.

And some disappear. Not from cruelty. They just don’t know how to relate to you without the old context.

My friend lost two close friendships in the first year. Gained three new ones. Different math than he expected.

A non-linear career is not a failure to have a linear one. But it’s also not the breezy adventure it sounds like in interviews.

It’s real losses and real confusion and long stretches where you don’t know if you decided well. It’s also more alive than staying safe.

Both things are true. The freedom and the cost. The excitement and the grief.

If you’re standing at the edge of something right now, you don’t need me to tell you it’s going to be great. You need someone to say: the mess you’re feeling is accurate. Other people have stood here, confused and uncertain. That’s not a bug in the process.

It’s just what it feels like to be between lives.

Share

A Strange Kind of Resignation Letter

In May 2024, Jan Leike walked away from OpenAI.

He wasn’t a critic from the outside. He was the head of their Superalignment team — the group specifically tasked with ensuring future AI systems remain safe and controllable. He’d spent years working on the hardest problem in the field.

His departure note was quiet. No dramatics. No accusations. Just this:

“Building smarter-than-human machines is an inherently dangerous endeavour. Over the past years, safety culture and processes have taken a back seat to shiny products.”

A few months later, more departures. Researchers who’d dedicated their careers to getting this right, walking out the door. Not because they stopped believing in the mission. Because they couldn’t do the work they came to do.

I keep thinking about what that must feel like.

The Weight of Knowing

Imagine spending your days trying to solve a problem that might determine humanity’s future.

You’re not in a movie. There’s no ticking clock on the wall, no dramatic music. Just a desk, a screen, endless papers, and the slow accumulation of understanding about what you’re building.

You start to see things.

You see how capable these systems are becoming. You see the gaps between what they can do and what we can verify. You see the distance between the safety work that’s needed and the safety work that’s funded.

And you see the calendar. The next release. The next capability jump. The investors. The competition. The pressure to ship.

You’re not building a bomb. You’re building something that might be extraordinary — or might be something else entirely. And you’re running out of time to figure out which.

What do you do with that knowledge?

Subscribe now

The Conscience Inside the Machine

Here’s what I find profound about the departures, the leaked memos, the quiet warnings:

The people closest to the technology are often the most concerned about it.

Not the critics who’ve never seen the code. Not the philosophers debating abstractions. The engineers. The researchers. The ones who know exactly what’s in the room because they built it.

When Jan Leike left, he wasn’t speculating about hypothetical risks. He was reporting from inside the building. When researchers sign open letters warning about extinction risk, they’re not seeking attention — many of them would prefer to work in peace. They’re breaking professional norms because something feels urgent enough to break them.

This isn’t pessimism. It’s the opposite.

It means the conscience is alive inside the labs. It means the people building these systems haven’t stopped asking whether they should. It means the fire hasn’t burned out the part of them that worries about getting burned.

What the Fear Tells Us

We tend to dismiss fear as weakness. Especially in technology, where optimism is currency and doubt is a career risk.

But there’s another way to read it.

The researchers who worry aren’t the ones who understand least. They’re often the ones who understand most. Their fear isn’t ignorance — it’s information.

When a firefighter backs away from a building, we don’t call them a coward. We listen. They know something about the structural integrity that we can’t see from the street.

The AI researchers sounding alarms are the firefighters. They’ve been inside the building. They’ve felt the heat. When they say “we need more time” or “this is moving too fast” or “I don’t think we know how to control this,” that’s not pessimism.

That’s expertise, expressing itself as caution.

The Ones Who Stay

Not everyone leaves. Most researchers stay, working within the constraints, pushing for safety in the margins, trying to make the systems better from inside.

I don’t know if they’re right to stay or wrong. I don’t know if the ones who left were courageous or gave up too soon. These are impossible judgments to make from the outside.

But I notice something about both groups:

They’re awake.

They haven’t numbed themselves to the strangeness of what they’re doing. They haven’t convinced themselves that someone else will handle the hard parts. They carry the weight of knowing — and they keep showing up, one way or another, to wrestle with it.

In a world sleepwalking toward transformation, the people who stay awake matter.

Even when they disagree. Even when they leave. Even when they stay and fight.

The Lesson for the Rest of Us

You and I aren’t building AI systems. We’re not in the room where the architectures are designed and the capabilities are tested.

But we’re in rooms of our own.

The meeting where the new AI tool gets approved. The policy discussion about automation. The conversation about which jobs to augment and which to replace. The small daily choices about how much to trust, how much to verify, how much to cede.

In those rooms, we can sleepwalk or stay awake.

We can assume someone else is handling the hard questions, or we can ask them ourselves. We can defer to the confident voices, or we can notice the ones who hesitate — and wonder what they know that we don’t.

The researchers who left didn’t have answers. They had questions they couldn’t stop asking.

Maybe that’s the lesson.

The questions don’t make you a pessimist. The questions mean you’re paying attention. And in a moment when the future is being written faster than we can read it, paying attention might be the most important thing any of us can do.

Helen runs market research for a consumer goods company in Mumbai. Her team adopted (read: forced) AI tools eighteen months ago — summarising reports, drafting analyses, crunching sentiment data.

Last quarter, the tool flagged a concerning pattern: negative sentiment around a new product launch was spiking. The recommendation was clear. Pull back. Rethink the campaign.

Helen paused.

Something didn’t sit right. The numbers were there, but the feel was wrong. She’d been watching this market for twelve years. The AI had been watching it for twelve seconds.

She dug into the raw data. The AI had weighted a single viral tweet — a joke, taken out of context — as heavily as thousands of genuine customer reviews. The algorithm couldn’t tell the difference between irony and outrage.

Priya overrode the recommendation. The launch proceeded. It became one of their most successful campaigns of the year.

When I asked her about it, she said something that stuck with me:

“The AI gave me a suggestion. It didn’t give me an order.”

Subscribe now

The Pattern No One Talks About

Here’s what I’ve noticed after two years of watching AI adoption across organisations:

The people who get the most from AI are the ones who trust it least.

Not the sceptics who refuse to use it. Not the enthusiasts who defer to it completely. The ones in the middle — who use AI constantly, but treat every output as a starting point, never a conclusion.

They’re faster than the sceptics. They’re more accurate than the enthusiasts. And they’re building something the AI vendors don’t advertise:

A new kind of expertise.

The Skill That’s Emerging

Something is happening in workplaces that deserves more attention.

A generation of professionals is learning, through daily practice, how to work with AI without working for it. They’re developing an intuition for when to trust and when to question. When to accept and when to override.

This isn’t in any job description. No training programme teaches it. But it’s becoming one of the most valuable skills in the modern economy.

Call it “AI judgment.”

It looks like:

• The writer who uses AI to brainstorm, then throws away 80% and keeps the spark

• The developer who reads AI-generated code line by line, catching the subtle bug the AI missed

• The manager who asks the chatbot for options, then chooses the one it ranked lowest

• The analyst who says “that doesn’t feel right” and trusts her twelve years over the algorithm’s twelve seconds

These people aren’t fighting AI. They’re conducting it.

The Optimistic Case

Here’s what the doom narratives miss:

Humans are remarkably good at staying human. So much so. that at times we underestimate our capability to be human.

Every prediction about technology replacing human judgment has underestimated our ability to adapt, to carve out space, to insist on remaining in the loop. Radio was supposed to kill conversation. Television was supposed to kill reading. The internet was supposed to kill human connection.

Instead, we absorbed each technology and bent it to human purposes. We’re messy and stubborn that way.

AI is different in degree. The risks are real. The pace is faster. The stakes are higher.

But the pattern holds: the most successful AI implementations keep humans central.

Not because of regulation. Not because of ethics training. Because it works better.

The teams that treat AI as a brilliant but unreliable intern consistently outperform the teams that treat it as an oracle. The organisations that build human checkpoints into their AI workflows catch more errors, make better decisions, and — ironically — move faster.

Human judgment isn’t a bottleneck. It’s a feature.

The Future We Can Choose

The AI safety conversation often sounds like a warning about an inevitable future. Superintelligence is coming. Control is slipping. The die is cast.

But walk into any workplace using AI today, and you’ll see something different:

People making choices. Every hour. Every day.

The choice to verify before sending. The choice to question before acting. The choice to say “I’ll decide this one myself.”

These aren’t heroic acts. They’re ordinary judgments made by ordinary professionals. But they add up to something significant:

A daily practice of keeping AI in its place.

The researchers call this “Tool AI” — artificial intelligence that enhances human capability rather than replacing human agency. The academics debate whether it’s possible at scale.

Meanwhile, millions of professionals are quietly doing it. Every time they pause. Every time they question. Every time they remember that the confident text on their screen is a suggestion, not a command.

The Invitation

There’s a version of the AI future where we sleepwalk into dependency. Where we slowly forget how to do the things we’ve outsourced. Where the machine makes the calls and we rubber-stamp them.

And there’s another version.

Where we use these extraordinary tools as tools. Where we get faster and more capable without getting smaller. Where “AI-assisted” means the human is still the one assisting the outcome into the world.

The second version isn’t guaranteed. But it’s not foreclosed either.

It’s being built, choice by choice, by the people who use AI best.

The ones who remember that the most powerful word in any workflow is still: wait.

The Email That Wasn’t

Last month, a mid-level manager at a financial services firm asked her AI assistant to draft an email to a difficult client. The relationship was strained. The stakes were high.

The AI produced a beautifully worded message — professional, empathetic, firm where it needed to be. She sent it.

The client called thirty minutes later, furious.

The AI had referenced a conversation that never happened. It had quoted a policy that didn’t exist. It had apologised for a mistake the company hadn’t made.

Every sentence had been delivered with perfect confidence. Not a hedge in sight. Not a “perhaps” or “if I recall correctly” to signal uncertainty.

The manager told me: “It sounded so sure. I didn’t even think to check.”

The Hesitation That Isn’t There

Here’s something you already know but haven’t fully processed:

When you’re uncertain, you hesitate.

You say “I think” instead of “I know.” You pause before answering. Your voice rises at the end of sentences. You use words like “maybe” and “probably” and “I’m not sure, but...”

These signals evolved over millions of years. They’re how humans coordinate under uncertainty. They’re how we tell each other: check this before you act on it.

AI doesn’t have this.

When an AI is 99% confident, it says: “The answer is X.”

When an AI is 51% confident, it says: “The answer is X.”

When an AI is completely wrong, it says: “The answer is X.”

Same tone. Same certainty. Same grammatically perfect delivery.

The hesitation that would save you — the small signal that something might be off — isn’t there.

The Confidence Illusion at Work

This isn’t a bug. It’s a feature baked into how these systems work.

AI models are trained to produce fluent, confident text because that’s what scores well. Hedging, uncertainty, and “I don’t know” were trained out — because users rated confident answers higher than uncertain ones.

The systems learned: sound sure, even when you’re not.

Now think about how AI shows up in your organisation:

• The chatbot answering customer questions

• The copilot suggesting code

• The assistant drafting documents

• The tool summarising reports

Each one delivers outputs with unwavering confidence. Each one is occasionally, silently, catastrophically wrong.

And unlike a human colleague — who would pause, or qualify, or say “let me double-check that” — the AI gives you nothing.

No raised eyebrow. No verbal tic. No subtle cue that this particular answer came from the 51% zone rather than the 99% zone.

Just smooth, authoritative text. Every time.

Share The other perspective

The Trap

Here’s where it gets dangerous.

Humans are calibration machines. We’ve spent our entire lives learning to read confidence signals. We know when someone is bluffing in a meeting. We sense when a colleague is out of their depth. We catch the micro-hesitation that says maybe don’t trust this.

AI defeats this calibration.

Every output comes wrapped in the same packaging. The brilliant insight and the hallucinated nonsense arrive in identical containers. Your finely-tuned bullshit detector — the one that’s protected you in a thousand meetings — goes silent.

And the more you use AI, the more you learn to trust that packaging.

The manager who sent the email wasn’t careless. She was trained — by months of good AI outputs — to trust confident AI text. The system taught her that confident meant correct.

Until it didn’t.

What This Reveals

Here’s the simple truth hiding in plain sight:

We talk about AI “alignment” — getting AI to want what we want. We talk about “control” — keeping AI within boundaries. These are real problems, and they’re hard.

But there’s a more basic problem we’ve already failed to solve:

We can’t tell when AI is wrong.

Not reliably. Not at scale. Not in the time it takes to hit send.

And if we can’t tell when today’s AI is wrong — AI that’s far less capable than what’s coming — how will we tell when more powerful systems make bigger mistakes?

The researchers call this “epistemic opacity.” The practical version is simpler:

The machine never hesitates. So you have to hesitate for it.

Every time.

The Discipline

There’s no elegant solution here. No setting to enable. No prompt that fixes it.

Just a discipline: treat every AI output as a first draft from an confident intern who might be completely wrong.

Check the facts. Question the framing. Verify the sources. And when stakes are high, don’t trust the packaging.

The AI will never pause to say “I’m not sure about this one.”

So you have to be the pause.

The button doesn’t exist.

This is what I tell myself every morning at 6:47 AM, when my Tesla pulls into the parking garage at Prometheus AI. The button that could pause everything—delay AGI by five years, by ten—it’s not real. Just a thought experiment from some podcast. Something someone asked in an interview once.

But I think about it constantly.

My name is Elena Chen, and I am a Senior Research Scientist at one of the three companies most likely to build artificial general intelligence. I have a PhD from MIT, 147 citations on my alignment papers, and a stock package worth approximately $4.2 million at current valuation. I also have a four-year-old daughter named Mei, who asked me last week if the computers at Mommy’s work were going to “eat all the jobs,” because that’s what a kid at preschool told her.

I didn’t know what to say.

I said, “Mommy makes sure the computers are nice.”

This is technically true. My job is alignment research—making sure that as our models get more capable, they continue to do what we want them to do. On paper, I am the safety. I am the guardrail. I am the person who makes sure we don’t build something that turns the world into paperclips or whatever nightmare scenario you’ve read about.

In practice, I am a fig leaf.

Here’s something people outside the industry don’t understand: nobody wakes up wanting to end the world.

Nobody sits in our Monday-morning all-hands thinking, “Today I will take another step toward the obsolescence of humanity.” The people I work with are, genuinely, some of the most thoughtful, intelligent, careful people I have ever met. They think about these problems. They lose sleep over them. They read the papers about existential risk. They signed the open letter.

And then they go back to their desks and keep building.

I do too.

Why?

Because if we don’t, Nexus will. Or DeepScale. Or that lab in Beijing that nobody talks about but everyone thinks about. Because the race has its own logic now, its own momentum, and stepping out of it unilaterally doesn’t stop anything—it just means someone else crosses the finish line first, someone maybe less careful than us, and then what was the point? This is what we tell ourselves.

This is what I tell myself.

There’s a phrase I’ve started seeing in internal documents: “closing the gap to human-level performance.” It appears in project updates, in planning decks, and in the metrics we track. The gap, closing the gap. Like it’s a good thing. Like it’s the goal.

Because it is the goal, that’s the whole point. That’s why investors have put $80 billion into this company.

That’s why our market cap exceeds the GDP of most nations. That’s why we have this beautiful campus with its meditation gardens, its farm-to-table cafeteria, and its on-site childcare, where Mei goes three days a week.

We are building the thing that will change everything, and we are being paid extraordinarily well to do it, with excellent benefits.

I try not to think about the fact that “closing the gap” means building something that could do my job better than I can. Could do everyone’s job better than them. Could think, plan, decide, and create better than any human who has ever lived.

I try not to think about what happens after the gap is closed.

Last Tuesday, I was in a meeting about our new architecture. I can’t say what it is—NDA, obviously—but I can say that when our lead researcher pulled up the benchmark results, the room went quiet. Not concerned, quite.

Awed quietly. The kind of quiet where people are afraid to breathe because they’re watching something historic.

The numbers were better than anyone expected. Way better. The kind of better that means timelines measured in months, not years.

Someone whispered, “holy shit.”

Someone else laughed nervously.

I felt like throwing up.

Instead, I asked about the alignment properties. Were the outputs stable? Had we seen any concerning behaviours in the extended evaluations? My voice sounded normal, professional, like I was asking about a quarterly report instead of something that might be—what did that essay call it?—” adding a new species of intelligence to Earth.”

The answers were fine. The answers are always fine, until they’re not, and by then—But that’s not how anyone wants to think. That’s not how you build a company. That’s not how you win.

I read an essay recently. Someone sent it in our internal safety Slack, which is funny because you’d think a channel dedicated to AI safety would be the one place where people take this stuff seriously. But the culture, even there, is complicated. You can raise concerns, but you can’t be a doomer. You can flag risks, but you can’t be the person who’s always flagging risks. There’s a fine line between being responsibly cautious and being seen as someone who “doesn’t get it,” who’s “not committed to the mission,” who might be happier “somewhere with lower stakes.”

Anyway. The essay.

It made an argument I’ve heard before but never quite so clearly: that AGI isn’t inevitable. That it’s a choice masquerading as fate. That the race only continues because everyone assumes everyone else will keep racing.

But what got me was this line about a button. The author said he’d asked around—hypothetically, if there were a button you could push to delay AGI by five or ten years, would you push it?

And the answer, from most people he asked, was yes.

Yes.

Even inside the companies. Especially inside the companies.

We would push the button.

We would slam that button.

But there is no button.

Except.

There might be. Not a literal button, obviously. But there are choices. Small ones, every day. Which project to prioritise? Which results to emphasise. Which risks should be escalated? Whether to stay late to optimise that training run, or go home and put Mei to bed.

And wider choices too. Whether to stay at all.

I’ve been thinking about this more lately. About the collective action problem. About how everyone feels trapped by a race that nobody would choose if they could choose collectively. About how the race continues because we’ve all convinced ourselves we’re powerless.

But we’re not powerless. We’re the people building the thing. Without us—the researchers, the engineers, the thousands of highly specialised workers who actually understand how any of this works—there is no race. The companies are just buildings and servers. We are the ones who make them run.

What if we stopped?

Not all at once, maybe. But some of us. Enough of us. What if we said: not like this. Not this fast. Not without thinking through what we’re actually doing.

What if that became the brave thing, instead of the thing that tanks your career?

Mei asked me again last night about the computers. Whether they were nice. I said yes. For now, yes. She wondered if they would always be nice.

I said, “That’s what Mommy is working on.”

She seemed satisfied. She went back to her dinosaurs. She’s been really into triceratops lately. She likes that they were plant-eaters, that they didn’t hurt anyone, that they just wanted to be left alone to eat leaves with their families.

The Triceratops are gone now. So are the other dinosaurs. Sixty-six million years of evolution ended in what might have been a single day, when something fell from the sky and changed everything.

They didn’t see it coming. They couldn’t have. They were just living their lives, eating their leaves, being dinosaurs.

We see it coming.

We are building it ourselves.

And every day, I go to work.

There’s a vigil that happens sometimes outside our campus. A small group, maybe thirty people, with signs about AI safety and human rights and the future of work. Security calls them “the protesters” like it’s a mild annoyance, like they’re there about parking or something.

I’ve never stopped to talk to them. I drive past in my Tesla with its tinted windows, and I feel—something. Shame, maybe. Or gratitude that someone is paying attention. Or anger that their signs are so simplistic, that they don’t understand the nuances, that they think it’s as easy as just stopping.

But maybe it is that easy.

Maybe the nuances are just the story we tell ourselves so we can keep driving past.

The essay I read had a proposal. Four things, basically:

• Track the compute,

• Cap the compute,

• Make companies liable for what they build,

• Regulate based on risk

It’s not radical when you think about it. We do similar things for nuclear technology, for bioweapons, for all sorts of dangerous capabilities.

We just haven’t decided that AI is dangerous enough. Or rather: the people making the decisions haven’t decided. And those people are—Us.

The researchers. The engineers. The policymakers who used to be researchers and engineers. The investors who fund the researchers and engineers. It’s a closed loop, and the loop has decided that the race must continue. Anyone who questions that is naive about geopolitics, or doesn’t understand the technology, or doesn’t want humanity to have nice things like cured diseases and unlimited energy. But I’ve read the internal discussions. I’ve seen how we talk when we think no one outside is listening. And the truth is, most of us are scared. Most of us would push the button. Most of us know, in some quiet part of ourselves, that what we’re doing is insane.

We just don’t know how to stop.

Maybe that’s what this is. I, writing this down at 2 AM while Mei sleeps in the next room, while my husband pretends not to be worried about why I can’t sleep anymore. Maybe this is the smallest possible act of pushing back. Of saying: I see what we’re doing. I’m part of it, and I see it, and I don’t know what to do about it, but I refuse to pretend it’s fine.

Maybe enough small acts add up to something.

Or maybe I’ll delete this in the morning, go to work, sit in another meeting about another breakthrough, feel that same awe and that same nausea, and keep telling myself the same story about how we’re the careful ones, we’re the responsible ones, if not us, then who.

I don’t know.

What I know is this:

Mei deserves a future where she’s not just a legacy species, tolerated by our successors. She deserves a future where humans still matter, still decide, still have some say in what happens next.

The essay was right: this future isn’t inevitable. It’s a choice. We’re making it every day—in the labs, in the boardrooms, in the policy offices, in all the small decisions that add up to a trajectory.

The button doesn’t exist.

But we do. And every day, we choose whether to keep racing or to stop and ask what we’re racing toward.

Tomorrow I’m going to ask my manager about transferring to our AI-for-climate-research division. It’s less prestigious. The stock refresh will probably be smaller. Nobody there is going to build the thing that changes everything.

That’s the point.

It’s not much. It’s one person, making one choice, in a race with billions of dollars and geopolitical stakes and momentum that feels unstoppable.

But I think about all the other people—in my company, in the others—who would push the button if it existed.

Who feel trapped in a race they never wanted to run. Who tell themselves the same stories I tell myself.

What if we all stopped telling those stories?

What if we remembered that we’re not passengers in this? We’re the drivers.

The button doesn’t exist. So we have to be the button.

The End

Share

Share The other perspective

Until one day, when I came home to this sight and suddenly thought, this is my purpose.

So may be the purpose need not always be a highly philosophically ambitious or a materialistically shiny.
Its ok if you don’t find yours, I don’t find mine everyday. I still go to work on most of my days thinking why am I doing this, deal with clearly stupid people with no sense of kindness (irony, right?, but its true).
But even a few days of such realisations are worth it for me.

I wish you find yours too, soon.

Subscribe now

Share