I’ve been lucky to work for bosses who trusted me with autonomy I hadn’t yet earned. Dave was the most important — and not just at Google. He’d hire me again years later to help build Upstart, and the same pattern played out there. Looking back, the trust he extended is the single thing that made the biggest difference in my career.
It’s also the framework I find most useful for thinking about how careers actually progress. Not title, not comp, not team size — those are the trappings of seniority. The real measure is trust. Specifically: how much does your organization trust you to operate autonomously? And that trust comes from one thing — demonstrated judgment.
There’s a spectrum of autonomy I think captures how careers actually progress. I’ll break it into five levels for the sake of explanation, but think of it as a continuum — the transitions between levels matter more than the labels.
Level 1: You can complete a task with guidance. Someone tells you what to do, and with some help along the way, you get it done. This is where everyone starts. The task is defined for you and the path is mostly laid out.
Level 2: You can complete a task without guidance. Same situation — someone hands you a project — but you can run with it. You don’t need check-ins or hand-holding. You hit a wall, you figure it out.
This is where a lot of career frameworks stop. They focus on increasing the complexity of the tasks you can handle at level 2. Can you do a harder project? A bigger project? A more ambiguous one? That progression is real, but I think the most important transitions happen next — and they’re a fundamentally different kind of growth.
Level 3: You can identify what needs to be done. This is the first big shift — from execution to identification. Instead of waiting for someone to hand you a project, you’re the one surfacing problems, opportunities, things that aren’t working. You come to your boss and say, “Hey, I noticed this issue. What should we do about it?”
Level 4: You identify the problem and propose a solution. Now you’re not just bringing problems — you’re bringing plans. “I noticed this issue, and here’s what I think we should do about it.” The conversation with your manager shifts from “what should I do?” to “here’s what I’m thinking — does that seem right?” You’re doing the thinking. They’re sanity-checking.
Level 5: You identify the problem, develop the plan, and start executing. You come to your boss and say, “I found this challenge, here’s the approach I’m taking, and I’ve already started.” Sometimes they just see the results. At this level, you’ve internalized what matters to the organization well enough to move without waiting for direction.
The jump from level 2 to level 3 is where careers really break open, and it’s worth understanding why. Levels 1 and 2 are about execution skill — getting better at doing the work. Level 3 and beyond are about something different: understanding the broader context of the organization well enough to know where the important work is.
That’s what I mean by judgment. It’s not just “I spotted a problem” — it’s “I spotted the right problem.” The one that actually matters to the business, the team, the customers. The one that’s high leverage. Anyone can generate a list of things that could be improved. The hard part — the part that earns trust — is knowing which ones are worth spending time on.
I had a moment like this early at Upstart. I’d been hired to build out university partnerships — that was the strategy when I joined. After a few months in the role, I’d become convinced that university partnerships weren’t going to be a winning channel for us. I went to Dave to tell him, half-expecting I’d just talked myself out of a job. Instead, he said, “Well, your job is to figure out what we need to do instead.”
That’s the thing about high-judgment people. They don’t just do what they were hired to do — they figure out what actually needs to be done, even when that means changing the plan entirely. And the bosses who hire that way know the most valuable thing they’re paying for isn’t the execution of a strategy. It’s the judgment to recognize when the strategy needs to change.
Here’s the part I had to learn the hard way: autonomy without communication isn’t really autonomy — it’s just operating in the dark.
Earning the trust to run on your own doesn’t mean disappearing. The opposite, actually. The people who get more autonomy over time are the ones who keep the people around them well-informed about what they’re seeing, what they’re doing, and where they’re heading. Not as a check-in or an approval-seeking exercise — just as a normal part of how the work happens. That’s how trust compounds. Your boss doesn’t get blindsided. Your peers stay aligned. People learn to trust your direction because they’ve been watching your judgment in action all along.
This is also where the best mentoring and collaboration actually happen — not in scheduled coffees or formal programs, but as a byproduct of communicating about the work itself.
If I had a career do-over, this is the thing I’d push myself to do better. I’ve always been more comfortable putting my head down and operating than narrating what I’m doing. But the more I look at the people who’ve successfully built careers around real autonomy, the more I notice that they over-communicate, not under-communicate. The trust to run comes from the people around you having clear visibility into where you’re running.
If you’re trying to grow, stop thinking about career progression as “doing harder things.” Think about it as earning more autonomy. And be honest with yourself about where you actually are on the spectrum.
Moving up isn’t just about being more proactive, though. It’s about investing in context. Learn how the business works beyond your team. Understand what your boss cares about, what their boss cares about, what the organization is actually optimizing for. The better you understand that broader landscape, the better your judgment gets about where to focus.
And then communicate as you go. The people who earn the most autonomy are the ones who make it easy for everyone around them to trust where they’re heading.
That’s what seniority actually is. Not a title on a leveling rubric — it’s your organization trusting you enough to let you run, because you’ve shown the judgment to run in the right direction, and the discipline to keep everyone else in the loop while you do.
Dave didn’t hand me a job description at Google. He handed me trust. Almost everything I’ve learned about careers since then has been about how to earn that, over and over.
]]>The screens debate has settled into a pretty tired binary — devices good or devices bad, embrace or remove. The AI debate is shaping up identically: let kids use it, or block it, detect it, punish it. Both miss what we should actually be talking about, and they miss it the same way.
When schools started rolling out one-to-one device programs, the case for them was mostly made in language like modernizing the classroom and preparing kids for the twenty-first century — phrases that did a lot of rhetorical work without committing to much. The drivers were a mix of vendor pressure, federal funding tied to tech deployment, equity arguments about the digital divide, and a sense that schools couldn’t keep using chalkboards while everything else moved on. Personalized learning got named sometimes, but the devices usually arrived first and any theory of how they’d improve learning came later, if at all. L.A. Unified’s 2013 iPad rollout became the cautionary tale precisely because nobody could quite say what the devices were supposed to make better.
What was real underneath the modernization rhetoric was a broader sense that our schools weren’t performing as well as we wanted them to, paired with the hope that the technologies transforming every other industry could transform education too. The hope was big but the plan was vague — and most of it hasn’t panned out. But the underlying concern was real — and still is. Twenty-five kids in a room, different speeds, different starting points, different learning styles, one teacher trying to move them all through the same material at the same pace. That isn’t the most effective way to teach kids, and we’ve known it for decades. Fast learners get bored. Slow learners fall behind. The case for trying something better was obvious long before anyone put a tablet in a classroom, and it persists after we take the tablets away. The device-skeptics keep skipping past this: removing the laptops doesn’t address the learning challenges that made us reach for new tools in the first place. The kids the one-size-fits-all model already wasn’t serving are still not being served — now without the tools that might have helped, if anyone had been clear about how.
That’s the argument I want to make about AI too. And the same “modernize the classroom” rhetoric is already coming back, just with AI swapped in for tablets: kids are going to use AI in the real world, so schools have to teach them how. I don’t buy it. Kids learn the tools that matter to their lives without school teaching them — that’s how Google, YouTube, smartphones, and every consumer technology of the last twenty years got absorbed by anyone under twenty. They’ll figure out AI the same way. The reason to bring AI into classrooms is not exposure. It’s whether AI can make teaching and learning better, and what we have to do to keep the obvious downsides from swamping the gains. Those are the questions worth arguing about: what improvement are we aiming at, what risks are we taking on, and how will we know if we got the trade-off right?
Bloom’s 2 sigma study from 1984 is one of the most-cited pieces of education research for a reason: in those experiments, one-on-one tutoring with mastery-based instruction produced outcomes roughly two standard deviations better than ordinary classroom teaching. The numbers have been argued about for forty years, but the broad direction holds — high-quality personalized instruction works dramatically better than the lecture-and-worksheet model most of us grew up with. The reason every kid doesn’t have a tutor isn’t that we don’t know it helps. It’s that high-quality tutors are expensive and there aren’t enough of them.
That’s the hypothesis behind what places like Alpha School are doing with AI — that the bottleneck on personalized, mastery-based instruction has always been supply and cost, and AI uniquely changes that math. I don’t know whether Alpha is right. They’re small, their student population is unusual, their published outcomes are limited. They could be wrong about the model, or about the implementation, or about whether AI tutors actually replicate what made human tutoring work in the first place.
What I like is that they’re asking the right kind of question. They have a clear theory of how AI can make learning better — give every kid the personalized, mastery-paced instruction that we have decades of evidence for but have never been able to scale — and they’re testing whether that theory holds. That’s the conversation I’d like to see more of. Not “AI yes or no,” but “where do we believe this can improve the learning experience, and is it doing that?”
I don’t think the lesson of one-to-one device programs is technology doesn’t belong in classrooms. It’s something more specific.
Some of the harms we’ve seen from devices in classrooms weren’t fully foreseeable at the time of deployment. When laptops first showed up in schools, social media barely existed. TikTok wasn’t a thing. YouTube was small. Recommendation algorithms hadn’t yet figured out how to pull a kid from an educational video into a four-hour rabbit hole. The balance of upsides versus downsides shifted as the surrounding technology evolved — and the deployment didn’t shift with it. The devices we put in kids’ hands also matter: a Chromebook locked to school apps is a very different proposition from an iPad with social media one tap away. The shape of the technology, not just its presence, drives the trade-off.
Two things I take from that.
First, the trajectory of new technology is hard to predict. With screens, the upsides looked clear and the harms emerged later — some unforeseen, some flagged early and dismissed because the upside was more exciting and the incentives all pointed at deployment. With AI the order is partly reversed: the cheating headlines and attention costs are arriving before we’ve seen any clear improvement in learning outcomes. The pattern doesn’t generalize, but the unpredictability does. Anyone telling you they already know how AI in schools lands is wrong.
Second — and this is the one I keep underweighting myself — it is much easier to add technology to a school than to remove it. Once one-to-one became universal, once homework lived on Blackboard, once digital was the default, the cost of pulling back was enormous. That asymmetry should make us more cautious about going all in, not because the tech is bad but because reversibility matters and we usually only learn what’s working a few years in. The version of “be adaptive and iterate” that I find most honest is: pilot small before you go big, measure outcomes you actually care about, and treat the initial deployment as a hypothesis you’re testing, not a future you’re committing to.
Stop asking whether AI belongs in classrooms. The yes/no framing is a trap, the same way the screens yes/no framing was a trap.
Ask instead: how do we believe this technology can make learning better? Where is it well-suited to deliver that improvement? What guardrails do we need to put around it? What outcomes are we measuring, and on what timeline are we willing to revise the deployment if the data tells us we got it wrong?
Those are slower questions. They don’t make for a good X argument. But they’re the questions where there’s real work to be done, and they’re the only ones that have a chance of producing schools that actually serve the kids in front of them.
]]>My career has been built on being a generalist. I know that sounds like faint praise — most people hear “generalist” and think “jack of all trades, master of none.” Someone who knows a little about a lot. That’s not what I mean.
The kind of generalist I’m talking about knows a lot about a lot. You won’t have the depth of a true subject matter expert in any one area — but to people outside that area, you should look and sound like one. And in every field you’re engaging with, the actual experts should think of you as a quasi-insider. That’s the bar.
Getting there requires one thing most people find uncomfortable: admitting what you don’t know — constantly. Every time you hear a term or concept you don’t fully understand, ask about it. Write it down. Look it up later. The game of being a generalist is a game of relentless learning, and it starts with being honest about the gaps.
And by the way — it has never been easier to do this than right now. With tools like Claude and ChatGPT, you can go from “I don’t understand this concept” to “I understand it well enough to talk about it” in minutes. The barrier to filling your knowledge gaps has basically collapsed. If you want to be this kind of generalist, AI is an enormous unlock.
But learning across domains isn’t just about personal development — the real payoff is what it lets you see. Every organization is made up of parts that interact in ways that are hard to predict if you only understand one of them.
Here’s an example from my time at Upstart. When we updated our ML models, it could change the mix of borrowers who got approved — shifting the credit spectrum of our loan portfolio. That shift might or might not fit the capital market structures we had in place, which affected how much funding was available, which in turn changed which marketing campaigns and channels were bringing in the right borrowers. Model changes rippled from data science to capital markets to marketing in ways that weren’t obvious from inside any one of those teams.
It wasn’t that people in marketing or capital markets weren’t paying attention — it’s that it wasn’t really anybody’s job to look across all of it. It’s easy to get siloed into your part of the business without seeing how changes elsewhere are affecting yours. And it’s not just seeing those connections after the fact — it’s being able to anticipate them. To look at a proposed model change and think: how does this impact our capital partners? Our funding availability? Our marketing costs? Sometimes that’s a knowledge problem, sometimes it’s an information flow problem. Either way, the habit of stepping back to ask “how does this ripple through the rest of the organization?” is rare and incredibly valuable.
If you’re the kind of person who’s curious about everything and has never been able to pick a lane — that might be a superpower, not a weakness. Lean into it.
Being a generalist isn’t the right path for everyone. This next one is: learn to Get. Shit. Done.
I know — that’s vague. I wish I had a magic playbook. But it starts with a mindset, and it’s one of the most consistently valuable traits I’ve seen in the people I’ve worked with. I think of it as the ownership mindset.
Here’s what I mean. There are people who do their part. They complete their tasks, hand things off, and move on. And then there are people who own the outcome. They don’t just do their piece — they own the end-to-end result. When something gets stuck at a handoff point, or a dependency falls through, or something just isn’t working — they don’t shrug and say “well, I did my part.” They figure out what needs to happen and they make it happen.
It’s the advice I give my kids all the time: finish the job. Not just your piece of it. The whole thing. Get it over the line.
Barack Obama gave similar advice — his number one career tip is “just learn how to get stuff done.” Be the person who projects “let me take care of that, whatever is needed, I can handle it.” He’s right. In any organization, there are people who seem perpetually busy but somehow never finish anything. And there are people who just make things happen. Learning to be the second kind of person is one of the most valuable career moves you can make.
The unsexy truth is that most of the value in execution lives in the last 10% — the follow-through, the loose ends, the stuff nobody wants to do. The person who chases down that last piece and actually ships a finished product is worth their weight in gold. That’s the GSD mindset: own the outcome, do the unglamorous work, and deliver.
The last lesson is about what you choose to take on. I borrowed this framing from Mark Suster, who wrote about hiring people who punch above their weight class — people willing to step into roles one level above their previous experience. The phrase stuck with me because it captures something I’ve found really important, not just in hiring but as personal career advice.
The idea is simple: don’t shy away from roles, projects, or challenges that feel too big for you. Actively seek them out. I think of it like skiing — if you’re not falling down every now and then, you’re not pushing yourself hard enough to actually improve. The same is true in your career. If every assignment feels comfortable, you’re probably not growing.
When I was hired at Google, I remember thinking, “I don’t actually think I’m qualified to do the thing they want me to do — but I think I’m qualified to learn how to do it.” That’s been the internal monologue for almost every job I’ve taken. The Google Apps for Education role, joining the founding team at Upstart, taking on business development for a company whose product I didn’t fully understand yet — none of those felt like things I was obviously prepared for at the time. I bet on my ability to figure it out rather than waiting until I felt ready.
This pairs with getting shit done. You can GSD without ever putting your hand up for the bigger challenge. You can volunteer for stretch roles without having the execution chops to deliver. Either one alone is valuable. But doing both — consistently delivering and choosing things that scare you a little — is more than the sum of its parts. That combination is what builds a career that accelerates.
People who punch above their weight tend to volunteer for tasks where they know how to do some of it but not all of it — and then they find a way to deliver the whole thing anyway. Their work is bigger than their title or experience would suggest. Over time, that’s how careers compound — not by waiting for someone to promote you into the next challenge, but by already doing the work before anyone asks.
Looking back at the moments that mattered most in my career, they were almost all moments where I said yes to something that felt too big — and then figured it out along the way.
So that’s it — three lessons, none of them the usual career advice. Learn a lot about a lot of things so you can see connections other people miss. Own outcomes end-to-end, especially the unglamorous last 10%. And say yes to things that feel too big. None of this is universal — but it’s the mindset that’s driven most of my success. Maybe some of it will work for you too.
]]>The concept is called regularization. In machine learning, when you’re training a model, there’s a constant temptation to make it more complex. Add more parameters. Capture more nuance. Fit the training data as closely as possible. And the model will get better at matching the data you trained it on. But there’s a catch: the more complex you make it, the worse it tends to perform on new data it hasn’t seen before. It memorized the past instead of learning the pattern.
Regularization is the fix. It’s a technique that deliberately penalizes complexity — it pushes the model toward simpler explanations, even if that means a slightly worse fit to the training data. The insight is counterintuitive: a simpler model that’s a little bit wrong about the past will usually be more right about the future.
I think this is one of the most underappreciated ideas I’ve encountered. Not because of what it tells us about machine learning, but because the exact same mistake — over-optimizing for complexity — is one we make constantly in public policy.
Here’s how I think about it. When we write a new law or regulation, the instinct is almost always to be as specific and comprehensive as possible. Anticipate every scenario. Close every loophole. Craft the perfect set of incentives so people behave exactly the way we want them to. On paper, it looks smart. You’ve thought of everything.
But in practice, you get the U.S. tax code.
The tax code is maybe the best example of what happens when you keep optimizing without regularization. Every provision made sense to somebody at some point. Every deduction, credit, exemption, and phase-out was added to solve a specific problem or create a specific incentive. Each individual addition was clever. And the result is a system so complex that it’s essentially incomprehensible — not just to ordinary people, but to the professionals who work in it every day.
Nobody understands the whole thing. Nobody can predict with confidence how a given change will ripple through the system. The complexity hasn’t made the tax code better at achieving its goals. It’s made it better at being gamed by people with expensive advisors and worse at being understood by everyone else.
That’s the trap. Each increment of complexity feels justified on its own terms. But the cumulative effect is a system that’s too clever by half — one that’s been so optimized for the specific scenarios its authors imagined that it fails badly in the real world, where things are messy and unpredictable.
I want to be careful here, because “just make it simpler” is easy to say and often gets used as a lazy argument against any regulation at all. That’s not what I’m arguing. Regularization doesn’t mean building a stupid model. It means building the simplest model that still captures the important patterns. There’s a big difference.
In policy terms, that means starting with a clear goal and asking: what’s the simplest set of rules that would actually achieve this? Not the most comprehensive. Not the most airtight. The simplest that works.
Take carbon pricing. You could write thousands of pages of sector-specific emissions regulations — different rules for power plants, manufacturing, transportation, agriculture, each with their own standards, exemptions, and enforcement mechanisms. Or you could put a price on carbon and let the market figure out where the reductions come from. The first approach tries to be clever about every scenario. The second establishes a simple principle and lets it propagate.
I’m not saying carbon pricing is easy to implement or that there aren’t real complications. But the principle is simple enough that a normal person can understand it: if you put carbon into the atmosphere, you pay for it. That clarity is worth something. When people understand the rule, they can plan around it, comply with it, and hold their representatives accountable for how it’s designed. When they can’t understand it, you’ve already lost most of the benefit.
If simplicity is so powerful, why do we keep choosing complexity? I think there are a few reasons, and they map pretty well to why ML practitioners overtrain models.
The first is that it feels like progress. Adding a new provision, addressing a new edge case — it feels like you’re making things better. You’re solving a visible problem. The cost of the added complexity is diffuse and delayed, while the benefit of addressing the specific case is immediate and concrete. It’s the same dynamic in ML: adding parameters improves your training metrics, and you have to be disciplined enough to care about what you can’t measure yet.
The second is institutional. Complex systems create their own constituencies. Tax preparers, compliance consultants, lobbyists, regulatory specialists — all of these roles exist because the system is complex. I’m not attributing malice here. But the people best positioned to explain why a rule is necessary are often the same people whose jobs depend on the complexity continuing. In ML, we don’t let the model vote on its own architecture. In policy, we kind of do.
The third — and maybe the most insidious — is overconfidence. The authors of complex policies believe they can anticipate how the world will respond. They think they can design the right incentive structure, predict the behavioral responses, and engineer the outcome they want. But the real world is messier than any model. People respond in unexpected ways. Markets shift. Technology changes the equation. The law of unintended consequences is always lurking, and the more specific your policy, the more brittle it is when reality doesn’t match your assumptions.
In machine learning, this is literally the problem regularization solves. You’re telling the model: don’t be so sure you’ve figured everything out. Leave room for what you haven’t seen yet.
I’ll be honest — I’m not a policy expert. I’m a tech person who learned a concept in machine learning and can’t stop seeing it everywhere. So take this for what it is: one person’s framework, not a prescription.
But I do think we’d be better off if the people writing laws asked themselves the question that ML engineers ask constantly: is this model more complex than it needs to be? Are we adding provisions because they genuinely improve outcomes, or because we’re trying to be clever? Would a simpler approach — one that’s a little less precise, a little less tailored to every edge case — actually perform better in the real world, where things are unpredictable?
The U.S. tax code could be dramatically simpler. Healthcare regulation could be more straightforward. Financial regulation could establish clearer principles instead of cataloging every prohibited behavior. In each case, we’d lose some theoretical precision. But we’d gain something potentially more valuable: rules that people can understand, systems that can adapt to change, and outcomes that are more predictable and more fair.
Regularization isn’t about being lazy or imprecise. It’s about having the discipline to resist unnecessary complexity — to build the simplest thing that actually works, and to trust that simplicity will generalize better than cleverness.
I’ve found that’s true in machine learning. I suspect it’s true in a lot of other places, too.
]]>Doomers believe AI will be amazingly impactful and will ultimately destroy human civilization à la Skynet from Terminator. Maybe they envision a kinder version like Wall-E. But ultimately, this is AI as a massively negative force for humanity.
Zoomers believe AI will revolutionize human experience and lead to some version of Utopia, solving most, if not all, human problems. Think Star Trek — a world of abundance where technology has freed us from scarcity. (Most sci-fi leans dystopian, so the zoomers have fewer good references to point to. Maybe that tells us something about human nature — or maybe just about what makes for good storytelling.)
Skeptics believe that AI is overhyped, overblown, and will ultimately fail to deliver on its promises. Perhaps they view AI as being something like VR — a technology that was always “just around the corner” and never quite arrived.
Realists believe AI is a genuinely transformative technology, but one that will take a long time to produce massive societal impacts and represents a manageable change for humanity. This view is perhaps best represented by Arvind Narayanan and Sayash Kapoor’s essay “AI as Normal Technology” — even revolutionary, general-purpose technologies are still, at the end of the day, technologies.
You could plot each of these positions on a 2x2 grid. The x-axis is “Scale of Impact” — from negligible to transformative. The y-axis is “Direction of Impact” — from catastrophic to utopian. The Doomers live in the bottom right. The Zoomers in the top right. The Skeptics on the far left. And the Realists occupy a big zone in the middle — high impact, but manageable. Positive on balance, but not utopian.
Despite all the commentary going into AI, the overwhelming majority of the public conversation falls into one of the first three categories. The Doomers and Zoomers get the most airtime because their arguments are dramatic and emotionally vivid. The Skeptics get attention because contrarianism always draws a crowd. But the Realists — the most likely camp to be right — are chronically under-discussed.
And I get why. “This will be really important but also manageable” is a boring headline. It doesn’t generate clicks, doesn’t fill conference keynotes, and doesn’t make for compelling late-night dorm room debates. But I think it’s where we’ll end up, and I’d put the probability well above 85%.
Here’s the thing, though. Saying “we’ll probably end up in the Realist zone” is not the same as saying everything will be fine. That box in the middle is enormous. It encompasses a wide range of outcomes — from AI that modestly improves productivity in a few industries to AI that fundamentally reshapes healthcare, education, scientific discovery, and economic opportunity. From a world where we’ve managed the transition reasonably well to one where we’ve captured only a fraction of the potential upside while fumbling through unnecessary disruption.
The difference between landing in the top-right corner of that box versus the bottom-left is — in practical, human terms — really significant. We’re talking about whether millions of people get access to better healthcare, whether scientific breakthroughs happen a decade sooner, whether the economic gains flow broadly or concentrate narrowly.
So the question isn’t really “will AI be transformative?” I think it will. The question is: within the zone of probable impact, where do we end up? And that’s not predetermined. It’s a function of choices being made right now by three groups: developers, deployers, and regulators.
AI developers — the labs building foundation models and the companies building applications on top of them — will shape where we land more than anyone. And the biggest risk I see isn’t that they’ll build something dangerous. It’s that they’ll build for the easiest use cases and leave the hardest, most impactful ones underserved.
It’s natural to focus on where the money is — enterprise productivity tools, coding assistants, marketing copy generators. These are real, valuable applications. But the transformative potential of AI lies disproportionately in harder problems: drug discovery, materials science, climate modeling, education in underserved communities. The areas where the market signal is weaker but the human impact is highest.
I’m not suggesting developers ignore commercial viability — that’s what funds everything else. But I think the best AI companies will be the ones that deliberately invest in high-impact, harder-to-monetize applications alongside their core business. AlphaFold didn’t come from a startup chasing revenue. It came from a lab that believed solving protein folding was worth doing even if the business model wasn’t obvious.
The developers who think beyond the next quarter’s revenue will do more to push us toward the top-right corner of that box than anyone.
By deployers, I mean the organizations — companies, hospitals, schools, governments — that actually put AI into practice. And the biggest issue I see here isn’t recklessness. It’s paralysis.
Too many organizations are sitting on the sidelines, waiting for AI to “mature” or for someone else to go first. I understand the instinct. Nobody wants to be the cautionary tale. But the cost of inaction isn’t zero — it’s just invisible. Every month an organization delays adopting AI in its workflows is a month of lost productivity, worse outcomes, and falling behind competitors who are learning by doing.
I’ve written before that the real danger isn’t making a wrong call — it’s standing still. That applies here. The organizations that will get the most value from AI aren’t the ones that waited for a perfect solution. They’re the ones that started early, iterated, and built the institutional knowledge to deploy AI effectively.
This is especially true in sectors like healthcare and education, where the potential upside is enormous but the institutional inertia is strong. Every hospital system that delays implementing AI-assisted diagnostics, every school district that waits another year to explore personalized learning — those delays have real costs measured in real human outcomes.
Deploy thoughtfully. Deploy carefully. But deploy.
I’ve written about this before, but it bears repeating: my biggest worry about AI isn’t the technology. It’s the regulation.
I keep coming back to nuclear power. Here was a technology with the potential to transform our energy infrastructure and make real progress on climate change. And we effectively regulated it into irrelevance — not because it didn’t work, but because the fear of what it could do overwhelmed the appreciation of what it could deliver. Decades later, we’re desperately trying to reverse course as we realize how much that decision cost us.
The parallel to AI is uncomfortably close. The loudest voices in the conversation are warning about catastrophic risks. Regulation is the natural response to fear. And if we’re not careful, we’ll build a regulatory framework that successfully mitigates the downside risks while also forfeiting the upside — the medical breakthroughs, the scientific acceleration, the expansion of what’s possible.
Good regulation is important. We need thoughtful rules around data privacy, algorithmic transparency, and accountability for AI-driven decisions. But as I’ve argued elsewhere, there’s a difference between regulation that creates guardrails and regulation that creates roadblocks. The former helps us navigate toward the top-right of the Realist box. The latter pushes us toward the bottom-left — or worse, keeps us from entering the box at all.
The regulators who get this right will be the ones who resist the urge to regulate based on what AI might become and instead focus on what it actually does today — who’s harmed, how, and what specific safeguards would help. That’s harder than writing broad prohibitions, but it’s the only approach that protects people without killing the potential.
I’m a Realist. I think AI will be genuinely transformative, broadly positive, and slower to reshape society than either the Doomers or Zoomers expect. But “broadly positive” isn’t guaranteed — it’s an outcome we have to work toward.
The developers building these tools, the organizations deploying them, and the regulators writing the rules all have a role to play. And right now, we’re spending too much of our collective attention on the dramatic but unlikely scenarios — the Doomers and Zoomers arguing about Skynet vs. Utopia — while under-investing in the practical, unglamorous work of steering toward the best realistic outcome.
That’s what I’d like to see change. Less debate about whether AI will save or destroy us. More focus on how we make sure the most likely outcome — the one in the big box in the middle — is as good as it can be.
That’s the zone of probable impact. Let’s make the most of it.
]]>I’m not a developer. I have a CS degree and wrote code professionally — in college, more than twenty years ago. I’ve spent my career since then in product, partnerships, and business roles. I’ve been thinking about products my entire career, but thinking through a product and actually shipping one are separated by a wall of implementation that used to require a team to get over.
That wall is coming down. And I think the implications are bigger than most people realize.
I built Whose Turn using Claude Code, Anthropic’s AI coding tool. And I want to be specific about what that means, because my own process evolved in a way I didn’t expect.
When I first started using Claude to write code, I did what most technical people probably do — I’d write some code, then use Claude for assistance, and I’d carefully read the code it wrote. I was still thinking of myself as the developer, with AI as a helper.
The big unlock was when I stopped looking at the code entirely.
I don’t mean I got lazy. I mean I realized that my value wasn’t in reviewing Swift syntax — it was in knowing what the app should do, how it should feel, and what tradeoffs to make. So I started treating Claude as my entire development team. I describe what I want. I test what it builds. I give feedback on what’s wrong. I iterate. But I have never once read the actual source code of Whose Turn. I couldn’t tell you what it says.
That felt uncomfortable at first. It felt like cheating, or like I was giving up control. But the product is better for it — because instead of spending my time parsing code I haven’t written professionally in two decades, I spent it on the things I’m actually good at: product decisions, UX tradeoffs, and knowing when something doesn’t feel right.
I should be honest about the rough edges. Building a mobile app is still hard — not the code, but everything around it. Getting an Apple Developer account, configuring signing certificates, navigating App Store review, dealing with provisioning profiles. That part is still genuinely technical and frustrating, and AI doesn’t fully smooth it over.
If I were advising someone who wanted to build their first thing, I’d actually steer them away from mobile. The simplest path right now is a web app — something like Supabase for your database and auth, deployed on Vercel. That stack is remarkably accessible, and AI tools can handle almost all of the implementation. You can go from idea to live product in a weekend. Mobile adds real friction that the tools haven’t fully eliminated yet.
For decades, building software has been organized around three distinct disciplines: product management (what should we build and why), design (how should it look and feel), and engineering (how do we make it work). Companies hire separate people for each. There are entire career tracks, job titles, and org charts built around this division.
That division existed for a practical reason — each discipline required deep, specialized skill that took years to develop. You couldn’t just pick up iOS development on a weekend. You couldn’t design a good interface without training your eye over hundreds of iterations. The specialization was a response to complexity.
But what happens when the implementation complexity drops dramatically? When the cost of turning an idea into working software falls by 10x or 100x?
The disciplines don’t disappear — but they collapse into fewer people. Instead of needing a PM, a designer, and two engineers to ship a simple app, you need one person who can think across all three and use AI to execute. The specialist roles don’t vanish for complex systems, but the threshold for what one person can build alone has shifted enormously.
I’ve written before about how the “augmenting vs. replacing” framing for AI misses the point — the real story is usually about how work reorganizes around the technology. This is a vivid example. AI isn’t replacing engineers. It’s changing the shape of who can build what, and how many people it takes.
This is where it connects to something I believe deeply about careers: the most valuable people in any organization are the ones who can think across disciplines. Not the “knows a little about a lot” kind of generalist — the kind who knows enough about product, design, engineering, business, and operations to see how changes in one area ripple through the others.
AI tools like Claude Code are an enormous lever for exactly this kind of person. If you understand what users need, can reason about design tradeoffs, and can think systematically about how software should work — you can now build the thing yourself. The bottleneck used to be implementation. For a huge class of problems, it isn’t anymore.
This is a genuine shift in who gets to build. Not just developers using AI to code faster — though that’s happening too — but product thinkers, designers, and domain experts who can now bring their ideas to life without waiting for engineering bandwidth.
I’ve always believed that asking the right questions matters more than having the right technical skills. That’s even more true now — because the technical skills are increasingly something you can delegate.
I don’t think this replaces professional software engineers. Complex systems, infrastructure, performance-critical code, large-scale architecture — all of that still requires deep expertise. What changes is the floor. The minimum viable team to ship a useful product just got a lot smaller, and the range of people who can participate in building just got a lot wider.
If you’re someone who’s always had ideas for things you wanted to build but couldn’t — the tools are here. The barrier isn’t gone, but it’s dramatically lower. And if you’re the kind of person who’s spent your career learning across disciplines rather than going deep in one — that breadth is about to become a lot more valuable.
]]>Show it what good looks like.
Here’s the thing about describing what you want: you don’t always know. Or more precisely — you know it when you see it, but you can’t fully articulate it. This is especially true for anything stylistic or nuanced.
I ran into this head-on when I started building a new podcast with my friend Cyrus. We wanted the AI-generated scripts to sound like us — not generic podcast-host-voice, but the way Cyrus and I actually talk to each other. The way we interrupt, riff, push back, land jokes.
So I tried telling it. I wrote descriptions of how each of us speaks. I explained our dynamic — who tends to set up points, who tends to land them, how we use humor. I spent a lot of time on this. And the scripts came back… fine. Competent. But they didn’t sound like us.
Then I had a different idea. Instead of describing our voices, I just gave it transcripts of our actual conversations. Real ones, unedited. And the difference was immediate. The AI picked up on patterns I hadn’t even thought to mention — little verbal tics, the way we build on each other’s points, the rhythm of how we go back and forth. Things I couldn’t have described because I wasn’t consciously aware of them.
That’s the core insight: you can’t tell an AI about things you don’t know you know. But you can show it examples that contain those things, and let it figure them out.
I saw this play out even more clearly with my other podcast, What the AI?!, where I co-host with Annie. We have a pretty dialed-in workflow — AI helps generate the script, we record, and then I feed the transcript back in so the system can learn from it.
One lesson it picked up on its own was particularly sharp. We’d recorded an episode where we ran long — too many stories, not enough time — and ended up skipping the last story entirely. The AI noticed this. On its own, it added a check to its script-writing process: make sure the final story in the rundown is skippable. Keep the most important stories earlier in the show so that if we have to cut, we’re not losing something critical.
I never would have thought to write that as an instruction. It’s the kind of operational wisdom that only emerges from watching real work happen — from seeing where the plan met reality and broke down. But because I showed the AI the gap between the script and what we actually recorded, it found the lesson itself.
There’s a useful analogy here to how people learn. If you’re training a new hire, you can hand them a style guide and a list of dos and don’ts. That helps. But they’ll learn far more from sitting in on a few meetings, reading a few real examples of great work, and seeing how the team actually operates.
AI is similar. Instructions set a baseline, but examples create understanding. And the richest examples are messy, real-world ones — not polished samples you curated to illustrate a point, but the actual artifacts of your work. Transcripts, drafts, email threads, before-and-after edits. The stuff that captures all the things you know implicitly but would never think to write down.
Next time you’re struggling to get AI to produce something that feels right, resist the urge to write a longer, more detailed prompt. Instead, ask yourself: do I have examples of what good looks like?
Feed it past work you’re proud of. Show it the real conversations, not your description of them. Give it the before and after so it can see what changed. Let it find the patterns — including the ones you didn’t know were there.
You’ll be surprised how much it picks up that you never thought to mention.
]]>That puts me in something of a minority. Polls show Americans are increasingly pessimistic about AI. The loudest voices in the conversation tend to be warning about job losses, existential threats, or societal collapse. And I get the appeal of those arguments — they’re vivid and specific in a way that optimism often isn’t. It’s easy to picture the jobs that disappear. It’s much harder to picture the ones that don’t exist yet.
But I think the pessimists are mostly wrong — and the optimists have a stronger case than they usually make. Here’s mine.
I’d break the AI pessimists into three broad camps: economic doomers, existential doomers, and societal doomers. Each has a version of the argument that sounds compelling. But I think they all share a common mistake — overestimating how “this time is different” and underestimating what history actually teaches us about how these transitions play out.
The economic doomers believe AI will cause mass unemployment — that it’ll replace huge swaths of white-collar work quickly, concentrating wealth among the few who own the technology. I understand the fear. But I think it dramatically overestimates the speed at which these changes actually work through the economy.
Anyone who studies forecasting knows a useful principle: we tend to underestimate the base case and overestimate how “this time is different.” I think that’s exactly what’s happening in the AI conversation. The base case — what history actually teaches us about technology adoption — is that these transitions are slow. Not because the technology isn’t powerful, but because organizations are slow. Change management is hard. Retraining takes time. Entire industries don’t restructure overnight.
Think about accountants. Spreadsheets didn’t eliminate accounting — they transformed it. There are more accountants today than before spreadsheets existed, and they do more sophisticated work. More people have access to accounting, too. The profession expanded because the tool made it more accessible and more valuable, not less.
Or think about this: twenty-five years ago, “mobile app developer” wasn’t a job. It wasn’t even a concept. Today it’s a massive job category supporting entire businesses and industries. It was essentially impossible to picture that role before the iPhone existed. And that’s the pattern — it’s easy to see the jobs that get disrupted. It’s hard to imagine the new ones that emerge. That asymmetry is a big part of why the negative argument feels more intuitive than the positive one. But the positive one is what history actually delivers.
Will AI change work? Absolutely — and massively. But not as rapidly as the doomers suggest. And will the endpoint be net positive? I believe it will — because that’s what every major technological revolution in history has produced, once we’ve had time to adapt.
The existential doomers worry about something much bigger — that AI will become a superintelligence, a digital god that either destroys or subjugates humanity. The Skynet scenario. I take this concern seriously, and I can’t prove it’s impossible. Nobody can prove a negative.
But I don’t see any evidence that we’re close to that kind of capability. What we have today — and what we’re likely to have for the foreseeable future — are very powerful tools. Not sentient beings. Not digital gods. Very, very good tools that can process information and identify patterns at scales humans can’t match. That’s world-changing, but it’s a fundamentally different thing than the sci-fi scenario.
Arvind Narayanan and Sayash Kapoor make this case well in their essay “AI as Normal Technology” — even transformative, general-purpose technologies like electricity and the internet are still “normal” technologies. AI may be the most important technology of our lifetimes, but it’s still a technology. Not a new species. I think treating it that way leads to much better decisions than treating it as an existential threat.
The societal doomers worry that AI will rip apart the fabric of society — through misinformation, manipulation, deepfakes, erosion of trust. And I’ll be honest: I think this camp has the most legitimate concerns of the three. These are real risks.
But here’s what I keep coming back to: we’re already experiencing most of these problems. The internet, social media, algorithmic feeds, smartphones — these technologies have already strained our information ecosystem, polarized our politics, and created real challenges for mental health and social trust. These aren’t hypothetical concerns. They’re the world we live in right now.
I don’t think AI represents a step-function change in those problems. It’s a continuation of trends that started well before large language models existed. And if anything, AI might actually give us better tools to address some of these challenges — from detecting misinformation to personalizing education to making complex systems more navigable.
I’m not dismissing the risks. I’m saying they’re not new, and I’m more optimistic about our ability to manage them than most.
Here’s where I think the conversation gets most interesting — and where the doomers most badly miss the mark. When most people think about AI, they picture chatbots, image generators, maybe a coding assistant. I’m excited about those things and the impact they’re already having. But they’re just a small part of the story.
The thing that excites me most is AI’s ability to accelerate discovery — particularly in fields where the bottleneck isn’t creativity or insight but the sheer scale of possibilities to explore.
Take medicine and biology. AlphaFold solved the protein folding problem — a challenge that had stumped researchers for decades — by doing something humans simply can’t: systematically exploring an enormous possibility space and identifying the structures that work. That’s not “artificial genius.” It’s a fundamentally different kind of tool — one that can sift through millions of potential drug compounds, protein structures, or genetic combinations and surface the most promising candidates for human researchers to investigate.
This pattern applies well beyond biology. In material science, we’re using AI to evaluate thousands of potential battery chemistries to find the ones worth investing in. In energy, AI is helping optimize everything from grid management to the search for better solar cell materials. In climate science, it’s accelerating the modeling of complex systems that are too large and interconnected for humans to analyze alone.
Dario Amodei, the CEO of Anthropic, wrote a long essay laying out many of these potential upsides in detail. His framing is different from mine — I don’t love the anthropomorphizing of AI as “geniuses in a data center” — but the underlying point resonates. We’re building tools that can explore possibility spaces at a scale that was previously unimaginable. The question isn’t whether that’s powerful. It’s whether we’ll let ourselves use it.
And that’s the pattern that gives me the most confidence. This isn’t speculation about some far-off future. AlphaFold exists today. AI-assisted drug discovery is happening now. The tools that will help us address climate change, develop better energy storage, discover new materials — those are in progress, not in the realm of science fiction.
When I look at the history of technology, this is what the big revolutions actually do. They don’t just automate existing work — they open up entirely new possibilities that we couldn’t have imagined before. The printing press didn’t just make scribes faster. The internet didn’t just make mail faster. And AI won’t just make knowledge workers faster. It will let us attempt things we couldn’t have attempted at all.
So if I’m this optimistic, what keeps me up at night? Regulation.
Not the existence of regulation — some regulation is necessary and important. What worries me is that we’ll over-regulate AI out of fear, locking it away before we get the chance to realize its benefits. That we’ll see the risks — which are real — and respond by putting this technology in a box.
I keep thinking about nuclear power. Here was a technology with the potential to fundamentally transform our energy infrastructure and help address climate change. And we effectively regulated it into irrelevance. Not because the technology didn’t work, but because the fear of what it could do outweighed the appreciation of what it could deliver. Decades later, we’re desperately trying to reverse course as we realize how much that decision cost us.
I worry we’re on that same path with AI. The doomer narratives are loud. Regulation is the natural response to fear. And if we’re not careful, we’ll end up in a world where we’ve mitigated the downsides but also forfeited the upsides — the medical breakthroughs, the scientific acceleration, the expansion of what’s possible.
For what it’s worth, I’m pretty optimistic about the private market figuring out education and corporate adoption. Companies will adapt because they have to — the competitive pressure is too strong. But regulation is the one area where well-intentioned decisions, driven by fear, could hold us back.
I’ll close with the observation that gives me the most comfort. We’ve been here before — and we’ve gotten through it.
There was a time when the vast majority of humans worked in agriculture. That’s not true anymore. We didn’t end up with mass permanent unemployment. We found new kinds of work, built new industries, adapted our institutions. It wasn’t always smooth or fast or painless. But we adapted.
I don’t pretend to know exactly what the AI-enabled future looks like. Nobody does. But I think the historical base rate is overwhelmingly on the side of optimism — not naive optimism, but the earned kind. The kind that says: this will be hard, there will be real challenges, and we’ll figure it out. We always have.
The biggest risk isn’t that AI changes too much. It’s that we don’t let it change enough.
]]>Why? Mostly because models have turned out to be somewhat undifferentiated. Every time a new model comes out with remarkable new capabilities, another matches it within a few weeks — often a more efficient model. Moreover, most models are now more than capable enough for the vast majority of every day use cases. User may prefer one model over another — but likely not enough to pay a premium for it.
So, who will be able to capture the value. I see a few potential winners.
So, where does that leave the players in the space?
OpenAI is winning on 3 right now. It’s trying to tackle 2 through integrations. This framework would suggest they should lean into that hard and move fast.
Anthropic isn’t winning any of these right now either. This framework indicates they are not in a great position.
Microsoft has real advantages in 2 and to some extent 1. That may be enough to win large deals in the enterprise space. I don’t see much of a path for them on the consumer side (though they will try to leverage Windows for 2).
Grok isn’t winning on any of these at the moment. Neither is Meta.
There is a case to make that Google is doing well on all 3. GCP is a solid contender in enterprise data and workflows. Gmail / Docs has a lot of consumer data. And more people likely interact with Gemini through Google search than use ChatGPT. It does seem like Google has the most paths to success at this point — including their own model.
]]>If you hate cookie banners (and who doesn’t), this seems like a clear win. But while the proposed solution (allowing users to specify their preference one time in their browser instead of on every site) will certainly improve everyone’s online browsing experience, it still doubles down on the horrible idea of asking regulators to play product manager.
It turns out that product management is hard. Many companies are quite bad at it. But regulators are usually worse. We would all be better off if regulators gave broad guidance and then let product companies compete and innovate to deliver the best experiences.
]]>