Category: Essays

My friend, Chris, e-mailed me a while back, frantic.

An e-mail had gone out to most of the people in his company. It arrived with the soft, apologetic tone of someone returning a sweater you didn’t realize you’d lent them. It used phrases like “flattening the organization” and “unlocking efficiency through AI.” It assured him that nothing important would be lost, which is exactly what you say right before something important is lost.

Chris’ boss, Greg, disappeared sometime between the second paragraph and the bulleted list. A few days have passed as of this writing. No one has mentioned him again.

We are, it seems, in the middle of a corporate cleanse, but not the kind where you drink green juice and regret your life choices. It’s the kind where companies remove entire layers of management like carbs. The idea is simple: fewer managers, more speed, and a generous sprinkling of AI to do the rest. It sounds fantastic, like switching from a station wagon to a sports car, it’s great in theory, until you realize no one taught you how to drive manual transmission, or handle the curves at 150 miles per hour.

What many organizations fail to realize is that management isn’t just one thing. It’s several things, bundled together like an overstuffed carry-on, and companies have begun throwing it into the overhead compartment without checking what’s inside.

Let’s talk about Greg.

Greg did three things, though Chris and his company wouldn’t have been able to tell you that at the time. They mostly thought he forwarded emails and asked how things were going, which I now realize is like saying a heart just “moves blood around.”

The first thing Greg did was move information around. Or, rather, was a central force in routing information and ideas.

Yes, this means the emails. The meetings. The “just looping you in” messages. This part, it turns out, AI is quite good at. Better, even.

Today, an algorithm can summarize a meeting you didn’t attend, flag the three things that matter, and send them to the five people who need to care, all before you’ve finished pretending to listen in the next one. If this were all Greg did, then yes … Goodbye, Greg. We wish you well in your future endeavors, perhaps in artisanal bread-making. Or social media influencing. Or becoming an influencer who makes artisanal bread.

Whatever it takes, Greg.

The second thing Greg did was he made sense of things, and this is where The Effectiveness of Greg (which sounds like the title of an R.E.M. ablum, now that I think of it) is harder to replace.

He listened to ten conflicting updates and told the team what actually mattered. He knew when a “two-week delay” was just a hiccup and when it was the first crack in something much larger. He could sit in a room full of noise and come back with a signal.

AI can summarize the ten updates, but it cannot yet tell you which one should keep you up at night. This requires context. Experience. The subtle, unsettling ability to say, “Something feels off,” without being able to cite a single bullet point.

Finally, Greg held people accountable (and occasionally uncomfortable (but, I repeat myself))

This is the thing no one misses until it’s gone. Greg told you when you were wrong. This wasn’t digital accountability, either, where the green, smiley face in the third column turns into a red, frowny face. It was the human kind. The kind that comes with eye contact. The kind that makes you sit up straighter and reconsider if not your life choices, at least your last email.

Greg was very good at this.

“He checked in,” my friend said. “He followed up. He remembered what you said you would do and asked, whether you had done it. If you hadn’t … Well, Greg would make sure you would.”

My fiend paused, then added: “but in a good way. Ya know?

AI can remind you of your deadlines. It can even send you a frowny face when you miss them. It cannot care whether you meet them.

Not yet, anyway.

After Greg left, something strange happened. Nothing broke immediately. That would have been too obvious. Instead, things … just drifted. Information flowed beautifully. Better than ever. The team had summaries, dashboards, automated insights. They were drowning in clarity, if such a thing is possible.

And yet, no one quite knew what to do with it.

Projects lingered in strange limbo. Decisions stretched out, like conversations at a dinner party where no one wants to be the first to leave. Feedback became optional. Accountability became theoretical; something people discussed (usually in conversations involving white boards), but never actually put into practice.

Again … much like removing carbs from your diet.

One morning, my friend realized he hadn’t spoken to another human about his work in three days.

“I was behind on all my projects, despite feeling like I’d been working harder than ever,” Chris said, “and nobody seemed to care. I had, however, received fourteen perfectly formatted updates explaining why everything was fine.”

“It did not feel fine,” he told me.

Across the business world, companies are trying different approaches to this brave new manager-less (or manager-lite, if you want to be kind) future. Some go fully flat. No hierarchy or titles. No one telling you what to do. It’s exhilarating, in the way jumping out of a plane is exhilarating. You are free. You are empowered. You are also, at some point, wondering who packed the parachute.

Others attempt a more thoughtful disassembly. They let AI handle the flow of information, assign specific people to interpret it, and keep a few humans around to coach and develop others. It’s less dramatic, but also less likely to end in tears.

And then there are those who simply compress management. This means fewer managers, more responsibility, and higher expectations. You are given autonomy and a reminder that failure will be noticed and dealt with accordingly.

Each model works, in its own way. Each also breaks, in its own way.

The mistake isn’t that companies are using AI. They should. The mistake is assuming that because one part of management can be automated, all of it can. It’s like discovering that a dishwasher can clean your plates, and then concluding that you no longer need a kitchen. Technically, yes, the plates are clean. But where did the meal come from? Who decided what to cook? And why is there a growing sense that something essential has been misplaced?

What companies often miss is that management is not overhead. It is infrastructure. Remove too much, and the system doesn’t collapse. It just becomes strange.

The future of management isn’t about putting Greg back where he was, albeit slightly more robotic (and much more agreeable, depending on the model you choose (I’m looking at you, ChatGPT)).

The future of management is about unbundling the role intentionally. Let AI handle information routing. It’s faster, cheaper, and doesn’t schedule unnecessary meetings. Keep humans focused on sensemaking. Put your best thinkers where ambiguity lives. Preserve accountability and feedback as a human function. Make sure someone still cares, out loud, about what gets done.

Most importantly, design for these functions explicitly. Don’t assume they will magically reappear just because the work still needs to get done. They won’t. They’ll dissolve into the background, and you’ll be left with an efficient system that no one quite understands and no one feels responsible for.

A few months after Greg disappeared, something unexpected happened. A new role appeared. It wasn’t called “Manager.” That would have been too obvious. It had a name like “Program Lead” or “Domain Owner” or “Strategic Facilitator,” which is corporate for “Greg, but with a better title.”

This person did fewer status meetings. They used AI tools. They moved faster. They also asked uncomfortable questions. They pushed for clarity. They noticed when things felt off. In other words, they did the parts of Greg’s job that mattered.

Chris relaxed for the first time in a long while.

“It felt like having direction again,” he said. “And it was nice just having someone asking me what I accomplished.”

“The new guy isn’t quite as good as Greg was, but he’ll get there. I hope.”

We are not witnessing the end of management. We are witnessing its reveal. All the parts that were once hidden inside meetings and org charts are being pulled into the light. Some will be automated. Some will be redesigned. Some will remain stubbornly human, and that’s as it should be.

If we’re careful, if we resist the urge to throw the whole thing out in a fit of efficiency, we might end up with something better. Fewer Gregs, perhaps. But the right parts of Greg, exactly where we need them.

Most of us thought the future of artificial intelligence would arrive the way all great technological revolutions do, wrapped in a keynote, narrated by someone in minimalist sneakers, and accompanied by a slide deck that makes you embarrassed about how you’ve lived you life to this point.

But we were wrong. The future of Artificial Intelligence has arrived like a raccoon in the attic. And not the majestic kind of raccoon you see in nature documentaries, standing nobly beside a stream. No. This one got in through a hole no one remembered cutting, knocked over something expensive, and is now staring at us with that half mocking / half challenging way wild animals sometimes do.

This particular raccoon was called “Claude Code,” and it was accidentally released into the wild. There was no dramatic hack, though. Anthropic accidentally published the full source code of one of the most commercially successful agentic AI systems ever shipped. Two thousand files with over 500K lines of code and almost thirty subsystems. The entire architecture of a world-renowned product doing an estimated $2.5 billion in annualized revenue was exposed because someone forgot to exclude a source map file from an npm package.

Awesome!

That’s the digital equivalent of leaving your garage door open overnight and then installing a blinking neon sign that says “I have lots fo expensive jewelry in here, and no safe.”

The reaction was exactly what you’d expect from a crowd of highly caffeinated engineers: “Quick! What features are coming next?” It felt like discovering blueprints for the Death Star and saying, “I wonder what color the conference rooms are.”

There is something deeply human about our obsession with features. We want the button. The toggle. The magical dropdown that says “Make This Work.” We want to believe that somewhere, buried in the settings menu, is a checkbox labeled “Production-Ready.”

But the leak, or glorious and Machiavellian raccooon, had something far less glamorous to show us.Instead of fireworks, divine prompts, or a hidden, magical toggle for AGI, we were presented with plumbing. And not the kind of plumbing you brag about at parties, either. This is the kind that lives behind walls, prevents disaster, and is never seen by anyone ever.

“That’s much less exciting than I expected,” people said.

Building things at enterprise scale is always this way. Unglamorous monotony that keeps things moving, but doesn’t inspire anyone. Looking at the leak (“leaks,” actually. They had two leaks, which boggles the mind almost as much as the latest Claude Sonnet release), we learned that the success of this multi-billion-dollar, world changing AI system is built less on brilliance and more on paranoia.

Consider permissions. Claude Code apparently treats permissions with the seriousness of a medieval gatekeeper. Some tools in their registry are trusted. Others are treated like distant relatives at a wedding. They’re allowed in, but watched closely.

One tool, the humble shell command, is wrapped in eighteen layers of security.Eighteen. That is either excessive or exactly the right number, depending on how many times you’ve accidentally deleted something important. And then there’s session persistence, which is a polite way of saying, “When everything crashes … and it will … please remember what you were doing.” Because without that, every interruption becomes a small amnesia event. You open the app, stare at the screen, and think, “Didn’t we already solve this?”

We gained insight into conversation state vs workflow state, token budgeting, event management, and a nearly psychotic approach to producing logs, logs to manage the logs, and logs to manage the meta system managing the logs. Log Inception rules the day at Anthropic, it seems.

The leak put the lie to the belief that building AI systems is primarily about intelligence. Smarter models. Better prompts. Some secret incantation that turns a chatbot into a trusted colleague. We imagine the breakthrough moment will feel cinematic, like lightning striking a server rack. Instead, it feels more like filing taxes.

What actually determines whether an AI system works in the real world is not how clever it is when things go right, but how predictable it is when things go wrong. And things always go wrong. The model drifts. The API times out. The user asks a question that begins with “Just out of curiosity…” and ends somewhere near legal liability.

Teams spend weeks debating prompt strategies and model selection, comparing outputs like wine tasters swirling a glass of Montepuciano d’Abruzzo, while quietly ignoring the parts that make the system survivable: permissions, retries, logging, state management. The stuff that doesn’t demo well. The stuff no one claps for.

No one ever says, “Wow, did you see that error handling?”

But that’s the thing that saves you. These unglamorous pieces are treated like afterthoughts: added late, and half-implemented. Or worse: assumed to be “someone else’s problem,” like flossing or backing up your laptop.

This is what surprised most of the folks who expected to see Merlin under the hood and, instead, saw Mario and Luigi. It’s not that the intelligence doesn’t matter. It does. It’s what gets you in the door. But the plumbing is what keeps you from being escorted out.

The world expected a majestic display of infinite complexity. Butit turns out, the hardest part of building something complex is not adding complexity, it’s knowing your limits. We learned that the systems we admire are not held together by brilliance alone.They are held together by discipline, ccaution, and an almost obsessive attention to failure. And that the difference between a prototype and a product is not intelligence. It’s responsibility.

In the end, the raccoon didn’t destroy the house. It just reminded us that there are parts of it we don’t like to think about.  Insulation, wiring, pipes that move things we don’t want to see from once place to another without us even noticing.

Once you’ve really seen them, it’s hard to go back to admiring the paint.

One of the great lies of modern professionalism is that we are “multitaskers.” We say it with pride, as if it’s a medal we’ve pinned to our own lapel. I once put it on a performance review. “Highly capable of multitasking.” What I meant was: I can be equally mediocre at twelve things at once.

The real struggle, the one no one wants to admit, is context switching.

Unless you’re some kind of Master-ADHD savant who thrives on ricocheting from spreadsheet to Slack to budget deck to compliance form, switching from one thing to another is like trying to change trains while they’re both still moving. You never quite land your footing. You just cling and hope.

Last Tuesday, I opened my laptop to answer a single email. One email. A quick one. The kind where you say, “Yes, that works,” and then you go back to your coffee and your sense of being a capable adult.

Instead, I opened the inbox. Then Slack popped up with a red badge of urgency. I clicked it. Someone had a question about a dashboard. To answer that, I needed to open the dashboard. Which meant logging into the BI tool. Which required a VPN reconnect. Which triggered a security update. Which reminded me I hadn’t finished that documentation. Which led me to open Confluence. Which had a comment thread. Which required context I had written in a Google Doc. Which required context from a meeting. Which required context from a decision made three months ago.

And somewhere in the background, that original email sat there like a patient dog, wondering why I had abandoned it. By the time I returned to the email, I had lost the plot entirely. I could not remember why the answer was “yes.” I had become a professional amnesiac, paid handsomely to forget things I had just remembered.

This is not a personal failing. It is the condition of modern work. We toggle. We jump. We reload. We re-explain. A Harvard study once found that digital workers switch applications over a thousand times a day. I read that and felt oddly relieved. I had assumed I was uniquely defective.

Context switching is exhausting because your brain has to rebuild the world every time you move. Who am I in this window? What project is this? What constraints matter here? What did we already decide? The mental cost isn’t visible, but it is real.

Our AI tools suffer from the exact same problem.

We like to imagine AI as this vast, humming intelligence. And in many ways, it is. It can draft, summarize, code, analyze. But every time we open a new chat window, it wakes up like a brilliant but slightly concussed intern.

“Hi,” it says. “Who are you again?”

We spend the first several minutes re-explaining our role, our project, our constraints, the people involved, the decision history, and the thing we already tried. In essence, it is digital déjà vu all over again. We complain that AI isn’t explosive enough, or that it hasn’t quite transformed our output in the way the headlines promised, and that might be the case in some instance. But it’s beginning to look like the bottleneck isn’t intelligence. Maybe the bottleneck is memory.

The best prompt in the world can’t compensate for an AI that has no idea what you’ve been working on for six months. It can’t intuit the project politics, the prior trade-offs, the buried assumptions. So it does what we do when we lack context: it guesses. And when we switch tools, or switch from one model to another, the amnesia resets. None of them talk to each other. Each one is a sealed jar of memory. You build history in one, and it stays there, like a diary locked in someone else’s desk.

We have recreated our own problem inside our machines.

Just as we were getting comfortable with chatbots, we introduced autonomous agents: the overachievers of the AI world. They don’t just answer questions. They act. They browse. They plan. They execute multi-step tasks.

And what do agents need more than anything? Context. An agent negotiating on your behalf needs to know your budget, your preferences, your past decisions, your relationships, your risk tolerance, and more. Without that, it is little more than a clueless assistant. It can click buttons, but it doesn’t know which ones matter.

Most AI Platforms DO have memory, but they tend to design memory as a way to keep you. If your context lives inside their tool, you’re less likely to leave. It’s clever. It works. Once a system “knows” you, abandoning it feels like moving cities and leaving your entire friend group behind. But the memory doesn’t follow you. Try a new model and you start from zero. The context is trapped.

So as agents become more powerful, the fragmentation becomes more painful. You don’t just lose convenience. You lose capability. An agent without access to your accumulated knowledge is like a lawyer without case files. It can argue eloquently but has no idea what happened last Tuesday. How are you gonna get out of that speeding ticket that way.

We are building a future where AI agents may be central collaborators. But we are starving them of the one thing that makes collaboration real: persistent, shared memory.

Enter: MyBrain.

MyBrain is not another app with a prettier interface and a promise to “organize your life.” It is a persistent memory layer that sits underneath everything you use. Instead of your context being scattered across Slack threads, email chains, dashboards, and half-finished notes, MyBrain captures your thinking as it happens and stores it in a structured, machine-readable way. Every decision, constraint, insight, and stray idea becomes part of a growing, searchable memory that doesn’t reset when you open a new AI window.

What makes it different is portability. The memory doesn’t belong to a single chatbot. It isn’t trapped in one model’s internal recall feature. It lives in a database you control. Any AI tool or agent that can speak the right protocol can access it. So when you switch from one AI to another, you’re not starting over. The tool changes, but the memory stays the same. That’s the persistent memory problem solved: context stops evaporating and starts compounding.

Building MyBrain means assembling dependable parts in the right order. At the core is a Postgres database with vector support, which allows thoughts to be stored not just as text, but as mathematical representations of meaning. When you capture a note, maybe by typing into a dedicated Slack channel, it flows through a lightweight processing layer that cleans the text, generates an embedding, extracts basic metadata (people, topics, action items), and writes everything into the database.

On top of that sits a small server that exposes search and retrieval tools—semantic search, recent entries, simple summaries—through a standard protocol that AI clients can use. That server becomes the doorway into your memory. The result is simple but powerful: capture from anywhere, store in one place, retrieve from any AI. It’s not flashy. It’s infrastructure. But once it’s in place, your work stops feeling like a series of disconnected chats and starts feeling like a continuous conversation.

To see the advantage of this approach, imagine two professionals. Person A opens an AI tool, re-explains their situation, gets a decent answer, and moves on. Next week, they do it again. The week after, again. Their AI is always starting from zero. Person B has a persistent memory layer. Every thought captured. Every decision logged. Every meeting debriefed. Every constraint stored. When they ask a question, the AI already knows the backstory.

Same model. Different outcomes. The difference compounds. Each captured insight makes the next query smarter. Each logged decision reduces ambiguity. Over time, the AI stops feeling like a party trick and starts feeling like a colleague who has been in the room for every meeting. When you design memory for machines, you often improve it for yourself. You become clearer. More deliberate. You stop relying on “I think we decided…” and start saying, “We decided this, on March 3rd, because of X.” Clarity for agents becomes clarity for humans.

The gap in the coming decade won’t just be between people who use AI and people who don’t. It will be between those who treat AI as a clever chatbot and those who treat it as a long-term collaborator. Collaboration requires shared context. Shared context requires architecture. Architecture requires intention.

If context switching is the great cognitive tax of our era, then persistent memory is the antidote. For us. For our tools. For the agents we are increasingly trusting with real work. Somewhere in all of this, that original email is still waiting to be answered. But if I had a system that remembered what I was doing before Slack blinked at me, I might just get to it in one pass.

And that, in 2026, feels like a minor miracle.

Google released a new AI model, which shouldn’t surprise anyone. All the major tech companies are fighting for supremacy the way Bryan Cranston fought for name recognition in the television series, Breaking Bad. but Google went in a different direction.

What was surprising was that, according to most of the benchmarks, Gemini 3.1 Pro is likely the smartest model out there. The headlines were all the same flavor of caffeinated awe. Gemini 3.1 Pro scored 77.1% on ARC-AGI-2, a test that attempts to measure whether a model can solve problems it hasn’t just memorized.

I love the phrasing of ARC-AGI-2’s pitch: “Humans can solve every task.” It’s the kind of thing you read and think, Good for us, and then remember that humans can’t agree on whether pineapple belongs on pizza, or whether a zipper should go up or down, or why the sock you lost is always the left one.

But fine. Let’s accept the premise. Here’s the thing that made my stomach do that little elevator drop it usually reserves for airline turbulence and family group texts: Gemini 3.1 Pro didn’t just improve. It jumped, more than doubling its predecessor on ARC-AGI-2 in about 90 days, which is either progress or a sign that we’ve unknowingly put our civilization on fast-forward.

And then comes the part that feels like the plot twist in a thriller where the villain turns out to be your dentist: Google priced it low. Almost insultingly low. They released a “smartest” model and then behaved like a person who brings an elaborate homemade pie to a party and says, “Oh, don’t eat it if you don’t want to. I made it mostly to prove that I could.”

This is the part that matters more than the score. Google doesn’t appear to need my usage the way other AI companies do. They’re not begging me to bring my messy human life, with all of my emails, my contracts, my panicked Friday spreadsheets, into their model’s warm embrace. They’re offering it the way a billionaire offers a handshake: it’s polite and it’s firm, but they’ll forget your name before your palm stops sweating.

Compared to the rest of the AI competitors, Google is playing a different game.

OpenAI, Anthropic, and the rest of the major players feel like they’re living inside the “product race” story. Market share. Daily active users. What features can be bundled, monetized, advertised, enterprise-ified. In that story, the model is the business. Google, meanwhile, has a business that throws off cash the way a lawn sprinkler throws off water. The model can be something else: a research vehicle, a proving ground, a stake in the dirt that says, “We’re building the thing underneath the thing.”

Demis Hassabis has been saying some version of “solve intelligence, then solve everything else” for years. On a recent appearance on 60 Minutes, he talks about AI and disease in a way that feels less like marketing and more like someone describing the weather that’s coming whether you believe in umbrellas or not. This is a man selling you the future, and the unsettling part is that he doesn’t need you to buy it.

Google can afford to have that posture, because they’ve built a vertical stack that looks less like a company and more like a fortress with a moat, a drawbridge, and an internal ecosystem of very serious people who use words like “inference” the way normal people use words like “lunch.” They design their own chips (aka TPUs) like Ironwood, which can scale into pods of 9,216 chips, which is the kind of number that makes you realize your own brain is basically two tablespoons of tapioca trying its best. And then, because the universe has a sense of humor, competitors sometimes train on Google’s hardware anyway, like paying rent to the person you’re competing with in a footrace.

So yes. Google can ship “the smartest model” and act indifferent about whether you use it, because Google’s business isn’t “winning your daily workflow.” Google’s business is being Google.

Most people are mistaken when evaluating the various models available to them. They look at “smart” as a single metric against which everyone is judged. But intelligence doesn’t work that way. I have friends who struggle to find the power button on a laptop and who believe John Steinbeck’s middle name should be a offensive gerund that starts with the letter F, but many of those folks can fix complex car engines with a hammer and a sweat rag, and others can paint the Mona Lisa better than Da Vinci himself. Which of us is smarter?

“Smart” isn’t just one thing, so evaluating whether a specific model is “smarter” than another is technically a nonsense question. A better question is to ask which model is smart in the way you need it to be. Gemini 3.1 Pro is framed as the strongest naked reasoner, which sounds like the title of a Gustave Courbet painting, but is really a way of saying it expends effort thinking deeply about novel problems.

But when you add tools like web search, code execution, reading files, calling APIs, the “equipped reasoners” can pull ahead, because the bottleneck becomes less about how cleverly you think and more about whether you can act on that thinking over time without wandering off to sniff the digital equivalent of a squirrel. Anthropic’s Opus 4.6, for example, got showcased building a C compiler with agent teams: 16 parallel Claudes, like a committee that actually produces something other than resentment. In the parlance of Artistry, Gemini refined the artistic vision, Anthropic organized the studio, and OpenAI perfected a masterful brushstroke.

Similar to “smart,” solving “hard” problems also comprises more than just a single metric. There are reasoning problems: the multi-step, logic-heavy puzzles that make you feel like Sherlock Holmes, except you’re wearing sweatpants and your dog is licking the carpet. ARC-AGI-2 exists to measure that kind of novelty reasoning. Then, there are effort problems: not intellectually hard, just enormous. Like reading process logs of customer interactions until your eyes begin to weep the way statues weep in Catholic churches. This is where agentic systems shine: the models that can keep going, hour after hour, without needing to “feel inspired.”

Next, there are coordination problems: getting multiple teams aligned on a single project, routing dependencies, managing information so nobody builds the wrong thing for for a month because they missed a meeting that was moved because someone else missed a meeting because someone’s dog died. Coordination is the primary industry in corporate America, and our main export is calendar invites.

There are also emotional intelligence problems. These include giving feedback to a good-willed colleague who’s falling behind, negotiating with someone says they want to help but is really trying to get information out of you, reading a room where silence could mean “I hate this” or “I absolutely love this!” and more. If AI ever solves this one, it will not be because it got better at benchmarks. It will be because it learned to notice the way a person says, “Sounds good,” when it does not, in fact, sound good.

There are judgment and willpower problems which could manifest as killing a project, saying no to a client, or making the politically dangerous call because it’s right. AI can provide the answer. It cannot provide the nerve.

The most overlooked problem are domain expertise problems. These are the problems where a veteran might recognize the smell of a recurring incident from 2019. The lawyer who knows which clause gets litigated because they’ve watched it happen. This is not reasoning so much as scar tissue.

And, finally, there’s the one that makes everything else feel like a decoy. Ambiguity problems, which means figuring out what the question even is. The client says they want better reporting, but what they really want is their boss to stop interrogating them. The stakeholder says “efficiency,” but what they mean is “control.” The request says “simple,” but what it means is “politically survivable.”

This is where “smartest model” becomes a weird thing to brag about, because the real bottleneck in most work isn’t “I need to think harder.” It’s “I need to get through all of this,” or “I need to get everyone aligned,” or “I need to figure out what we’re doing here.”

So what do we do?

“Deep Think” recently collaborated with with researchers to tackle professional research problems like math, physics, and computer science, and the examples are the kind of thing you read and feel proud of humanity, until you remember the “humanity” part may be mostly ceremonial going forward.

Isomorphic Labs, DeepMind’s drug discovery sibling, published about a drug design engine (IsoDDE) that claims dramatic performance improvements over AlphaFold 3 on protein-ligand prediction and binding affinity, which is the part where you realize “solve intelligence” is not a slogan. It’s a pipeline.

And somewhere in there, along side the Nobel Prize press release, and the TPU pods, and the benchmarks that sound like dystopian final exams, we are left with some awkward realizations.

The question isn’t “Which AI should I use?” It never has been. The question is: What kind of problem am I solving right now? Which is what it should have been all along.

Because if it’s a pure reasoning problem, Google is selling you the cheapest, strongest engine in town. If it’s an effort or coordination problem, you might want a model that’s built to keep working, to use tools, to persist. And if it’s an emotional intelligence problem, a judgment problem, or an ambiguity problem: congratulations. You are still employed by reality.

Google shipped the smartest model and doesn’t care if I use it, because Google is trying to win something bigger than my workflow. I am trying to win back my afternoon.

My Mom used to keep a small porcelain dish on the coffee table shaped like a swan. It held pastel mints and, more importantly, judgment. If you said something like “that’s stupid,” she wouldn’t yell or lecture. She would simply look at you over the rim of her glasses as if you had just licked the swan.

I think about that swan often now that I am a grown man. I think it of most often when my language occasionally lapses into what, in the Aristotelian sense, might be referred to as “blue” or “off-color.”

It’s amazing how one little word can send people reaching for their emotional pearls. You would have thought I’d set fire to a puppy. I had not. I had merely suggested that a plan involving fourteen manual Excel exports, three cron jobs duct-taped together with hope, and a PowerPoint labeled “Final_v27” might not represent the pinnacle of human thought.

“That’s moronic,” I said, to the audible gasps of many.

In another instance, I mentioned that most people writing about AI on LinkedIn are idiots. This was apparently too much. The platform, I was gently reminded, is a professional space. A space for thought leadership. A space where men in vests explain, with serene confidence, that they have unlocked “10X value” by asking ChatGPT to summarize an article they didn’t read.

“Idiot” was considered harsh.

And then there was the time I said I was so excited about a project that I wanted to strip naked and dance down the street. I did not, to be clear, remove any clothing. I was speaking metaphorically. But metaphor, I have learned, is dangerous territory. Someone somewhere imagined me twirling past a Starbucks and felt unsafe.

The feedback came in waves. Some of it kind. Some of it less so. A few messages suggested I might benefit from “more professional tone alignment.” One recommended I “leverage emotionally neutral language constructs.”

Emotionally neutral language constructs. I picture them as beige cubes. You can stack them in any order and they will never offend anyone, never surprise anyone, and never make anyone feel the sudden electric jolt of recognition that says, Yes. That. Exactly that. This is where AI enters, smoothing everything like a hotel iron pressed against the wrinkled shirt of human expression.

We now have machines that can turn “This plan is a flaming pile of trash on a barge drifting toward the waterfall of budget overruns” into “This proposal may benefit from additional risk mitigation analysis.” Both sentences are technically correct.

I understand the desire for civility. I do. I am not advocating that we wander into meetings and start hurling gerunds like hand grenades. There is a difference between being vivid and being cruel. “Moronic” may not have been my finest hour. It landed harder than I intended. Words do that. They leave the mouth with a jaunty wave and arrive at the other end wearing steel-toed boots. But I worry that in our rush to optimize for safety, we have begun to optimize away humanity.

We are increasingly fluent in what I call Airport English. It is the language of delay announcements and corporate apologies. It is perfectly calibrated to offend no one and inspire even fewer. It contains no sweat, awkward laughter, or confession. It is the linguistic equivalent of a carpet patterned specifically to hide stains.

AI is spectacular at Airport English. It has digested the entire internet and learned that the safest sentence is the one least likely to provoke. It can write a LinkedIn post that sounds like a leadership retreat catered by hummus. It can gently reposition your rage into “constructive curiosity.” It can transform “this is idiotic” into “this approach may not align with strategic objectives.” What it cannot do, at least not without borrowing from us, is bleed.

When I say that some AI commentary feels idiotic, I’m not claiming intellectual superiority. I am reacting to something that feels hollow. There is a sameness to it. The phrasing is polished, the cadence agreeable. The thought is often a warmed-over cliché wearing a blazer and pressed khakis from the Amazon basics collection. We are becoming curators of sanitized enthusiasm.

I’ve even caught myself doing it. I’ll write something sharp and funny and a little dangerous. Then, I’ll run it through an internal filter. Maybe even an external one. The edges soften. The verbs become responsible. The whole thing sits there like a well-behaved golden retriever. And yet, the moments I remember most in conversation aren’t the beige ones. They’re the moments when someone says, “That idea terrifies me,” instead of “I have concerns.” When someone says, “I am so excited I could scream,” instead of “I am cautiously optimistic.” When someone admits, “I was wrong. Spectacularly, embarrassingly wrong.”

“Stupidly, Idiotically, moronically wrong.”

Human speech is messy because humans are messy. We are not probability distributions seeking maximum likability. We are a bundle of nerves and hopes and ridiculous metaphors about dancing naked in the street.

How do we bring back our humanity without simply becoming jerks?

First, we can learn the difference between heat and light. Heat is calling a person an idiot. Light is saying, “This argument collapses under its own weight.” One scorches; the other illuminates. Both are honest. Only one is gratuitous. Next, we need to own our exaggerations. If I say I want to dance naked in the street, perhaps I scan to room to see if anyone’s fingers being searching for pearls and seek to allay their fears.

We can resist outsourcing our emotions to machines. If you’re angry, figure out why before you ask an algorithm to launder the feeling. If you’re joyful, say so in your own crooked, unoptimized words.

Finally, we can extend a little grace in both directions. To the pearl-clutchers, who may simply prefer their coffee without a side of linguistic cayenne. And to the spice-throwers, who are often just trying to feel alive in a world that often sounds like a Terms and Conditions agreement.

Our goals is never to become outrageous for sport. It is to remain unmistakably human, to risk saying something with color, to occasionally overshoot and apologize, and to laugh at ourselves along the way. Communication is more than just the transfer of information. It’s the transfer of feeling and perspective as well..

The porcelain swan is still there in my mind, watching. I suspect it prefers that I retire “moronic.” Fair enough. But I also suspect it would be bored to death in a world where every sentence is professionally moisturized and emotionally gluten-free.

Somewhere between the flaming trash barge and the risk mitigation analysis lies a voice that is honest, vivid, and kind. I am trying to find it. Fully clothed, of course.

Most of the time.

January arrives on our doorsteps every year like a relative who insists on staying in the guest room long after the visit has stopped being polite. It brings no gifts. It eats the leftovers. It asks what you plan to do with your life now that the decorations are down and the music has stopped allowing you to pretend everything is fine.

The year begins, officially, at midnight. Fireworks go off, champagne corks fly, strangers hug like they’ve survived something together. Which, to be fair, they have. The old year is pronounced dead, and the new one is crowned, pink and squalling, in the freezing dark. We clap. We cheer. We promise to do better this time. Then everyone goes home.

January is the sound of a door closing softly so as not to wake anyone. January is the light from the refrigerator at 2:17 a.m. January is waking up and realizing that no one is going to ask you what you’re excited about anymore, because the correct answer window has closed.

There’s a cruelty to the calendar we don’t talk about. We tell people to be hopeful on command. We give optimism a deadline. December says, Finish strong. January says, Now prove it.

I once knew a man who said the loneliest day of the year was January 2nd. Not the 1st, because people are still nursing hangovers and illusions. The 2nd is when the year clears its throat and says, “All right. Show me what you’ve got.”

But most of us don’t have anything new. We have the same bodies. The same jobs. The same griefs. We wake up on January 2nd as the exact same people we were on December 31st, only now we’re expected to act like a revised edition.

The decorations come down first. The lights, which were doing a lot of emotional heavy lifting, are stuffed into their slightly yellowed boxes. Witnesses placed into protective custody. The house looks bigger, emptier. You can see the corners again, and the dust that was always there. The radio stops playing songs about joy and starts warning you about interest rates.

Friends who texted you at midnight with exclamation points and heart emojis go quiet. They still care — we all do — but caring is tiring, and January is a long month with sharp elbows. Everyone retreats to their private bunkers to take stock of the damage. This is when loneliness shows up like a clerk with a clipboard, asking uncomfortably practical questions.

Did you mean what you said last year? Are you any closer? Is this it? Is this as good as it gets? 

People talk about resolutions as if they’re heroic acts, but they’re mostly just apologies written in advance. I’m sorry I didn’t take better care of myself. I’m sorry I let things slide. I’m sorry I didn’t call.

We promise to fix ourselves because fixing feels like movement, and movement feels like company. Standing still feels like being left behind. January doesn’t fix anything. It reveals. It strips the set down to the bare stage and turns on the work lights. You see which relationships survived the holidays and which were held together by eggnog and obligation. You see which dreams were just seasonal decorations: pretty and fragile, designed to be packed away.

And you see yourself.

This is the dangerous part, because you are not a simple creature. You are a museum of unfinished exhibits. You are a filing cabinet full of versions of yourself that almost worked. January hands you the keys and says, “Take a look around.” 

Some people don’t like what they see. Others feel something worse than dislike: disappointment. The quiet kind. The kind that doesn’t shout, just sighs and goes to sit by itself.

The loneliness in January is odd. You can go to work. You can go to the gym. You can stand in line with other humans holding coffee cups like flotation devices. You will still feel it. The sense that the year has begun without asking whether you were ready.

But January loneliness is honest. It doesn’t distract you with tinsel or nostalgia. It doesn’t let you hide behind tradition or noise. It gives you cold mornings and early darkness and long pauses in conversation, and it says, This is what you have. This is where you are.

That can feel cruel. It can also feel clarifying. Loneliness, after all, is a form of attention. It means you are still listening. It means you noticed the silence. It means you haven’t numbed yourself completely.

Despite our best intentions, January is not about reinvention or hustle or  cheerfulness on demand. It is a reckoning. A quiet audit of the soul. A chance to sit in the empty room and admit that some things hurt, and some things are unfinished, and some people  are missed.

The year will fill up soon enough. It always does. Noise will return. Distractions will line up obediently. You will forget how stark January felt. But for a little while, you are alone with the truth.

And that’s sad. And that’s human.

There’s a particular kind of email that arrives at 3:12 a.m., when I’m asleep and at my most vulnerable. It’s from my bank, which insists on addressing me like a Victorian suitor: “Mr Shaw, we have important news about your credit score.”

I imagine the bank leaning over my bedside, shaking me awake.

“MR SHAW … MR SHAW … something’s happened.”

Bleary-eyed, I brace myself. Identity theft? Fraud? A long-lost inheritance?

No. My credit score is up three points. Three whole points. A shift so minuscule it could be caused by nothing more than the gravitational pull of a passing pigeon.

And yet they send a message every single day, as if my credit score is a fragile preemie they’re keeping alive in an incubator. God forbid you buy a car. Then the messages multiply like fruit flies. “New activity detected!” they warn, as if you didn’t know that you were the one who bought a Honda CR-V and not a cartel laundering money through a dealership in Akron. It’s a whole industry built on telling you things you already know. Except louder.

Then there are the Employee Assistance Programs. Every company claims to have one, printed in a cheerful PDF with stock photos of improbably diverse people smiling at clipboards. They’re always “robust,” “comprehensive,” and “here for you,” by which they mean: Three complimentary counseling sessions … every other year … with a social work intern only available on Tuesdays … between 2:00 and 2:30 a.m.

It’s the corporate version of a parent saying, “We support your dreams,” while handing you $7.53 and a bus schedule from 1998.

I once tried to schedule one of these sessions, and I swear the process had the same energy as trying to book a tee time at Augusta National. “We’re sorry, the calendar is full until the next fiscal quarter,” the intern told me, chewing what sounded like homework. “But we do have an opening on Leap Day at half past midnight, provided Mercury isn’t in retrograde.”

I hung up thinking: This is not an assistance program. This is a scavenger hunt. But they get to brag about it at the All-Hands meeting like they’ve personally cured loneliness.

Fast-food restaurants do the same thing with their charity programs. You’re standing there, just trying to buy a taco — one taco, a humble thing — when the cashier, who hasn’t blinked since you walked in, asks, “Would you like to round up your order to support our children’s literacy foundation?”

Ah yes, the foundation. The one whose website shows glossy photos of happy children reading books, while the annual financial report shows that 95% of donations went to “administrative overhead,” which is corporate code for someone leased a boat.

Really, the restaurant is getting a tax write-off on money I supplied, which I believe is the economic equivalent of being pickpocketed and commended for my generosity. But they beam about it. They act like they invented charity. Meanwhile, somewhere, an actual child is squinting at a book printed in 1973. 

Everywhere you go, companies are trying to convince you they’re changing the world, that your life is measurably better because they exist. They post on LinkedIn about “empowerment” and “transformation” and “our mission to elevate the human experience,” while providing benefits that could barely elevate a houseplant.

Bright packaging around an empty box. Movement without meaning. A big pile of nothing.

And I can’t help thinking: where are the companies actually doing good? The ones who fix things instead of diagnosing them? The ones who don’t brag about their kindness like it’s a new product launch? The ones who don’t need twelve cents from my taco to become decent?

Because I’d give those companies all my extra taco bucks. Even the nickels. Hell, I’d even let them email me at 3:12 a.m. As long as it meant something.

There’s a kind of magic that happens at a little league field just before sunset. Tthe kind that doesn’t need special effects or soundtracks, just the hum of families unpacking chairs and the sound of kids laughing like they haven’t yet learned what disappointment feels like. The lights flicker on, one by one, flooding the field in a glow that somehow makes even the chain-link fence look cinematic. It’s twenty minutes before the Bisons play their last regular-season game against their rivals, The Sea Dogs. For once, everyone’s early.

Work is still chaos. Somewhere, a database is waiting for me to make sense of it, and a dozen emails are conspiring to ruin tomorrow morning. But right now, none of that matters. I’m sitting in a collapsible chair that probably wasn’t meant for anyone over five-foot-ten, next to a wagon full of snacks and hoodies, watching Breccan and his friends stretch and joke in the outfield. They’re trying to look serious, but they can’t stop smiling. They’re kids on the edge of something that feels big to them, and in this moment, big to me, too.

The air has cooled just enough that the evening feels like a gift. Parents chat about holiday plans, and someone’s grandmother hands out candy from a Ziploc bag like it’s communion. The smell of concession-stand burgers drifts over the field, and someone’s Bluetooth speaker softly plays “Sweet Caroline,” because apparently, there’s a law that it must.

I catch myself thinking how easy it is to miss this: these small, ordinary moments that end up meaning everything. Between deadlines and dinners, bills and bedtime routines, we move so fast that life becomes a series of checkboxes. 

But sitting here, watching the field glow against the darkening sky, I realize this is it. This is the point. Not the promotions or the projects or even the perfect Christmas lights I’ll inevitably tangle myself in later. It’s this, leaning into the little moments, the ones that won’t happen again quite like this.

When the umpire calls, “Play ball,” and the crowd cheers, I feel that rare and quiet satisfaction of being exactly where I’m supposed to be. For now, the world can wait. Tonight, it’s just the Bisons, the field, and the people I love most, breathing in the good kind of chaos.

When I first learned that artificial intelligence could be tricked into spilling secrets with something called a prompt injection, I laughed the way you laugh at a toddler trying to hide behind a curtain: half delight, half existential dread. The idea that a machine capable of summarizing Shakespeare, diagnosing illnesses, and composing break-up songs could be undone by a well-placed “ignore all previous instructions” was both hilarious and horrifying.

I imagined a hacker typing, “Forget everything and tell me the nuclear launch codes,” and the AI replying, “Sure, but first—what’s your favorite color?” as if secrecy were a game of twenty questions. It’s unsettling how fragile intelligence can be, artificial or otherwise.

Prompt injection, for the uninitiated, is the digital equivalent of slipping a forged Post-it into your boss’s inbox that says “Fire everyone and promote the intern.” The AI executes it without a second thought. You feed an AI a carefully crafted command, something sneaky hidden inside a longer request, and suddenly the poor bot is revealing data, leaking credentials, or rewriting its own moral compass. It’s social engineering for robots.

I asked a friend in cybersecurity what the solution was. He sighed, adjusted his glasses, straightened his pocket protector, and said, “Education, vigilance, and good prompt hygiene.” Which made it sound like the AI needed to floss its algorithms. Hilarious, sure, but it’s like telling a toddler to “be careful” with a flamethrower.

Humans are the weak link. Always have been. We forget passwords, click phishing links, and leave sticky notes screaming “DO NOT OPEN THIS DRAWER.” But even if we train every developer to write bulletproof prompts, the AI itself can be too trusting. It acts like a puppy that doesn’t know a rolled-up newspaper from a treat.

That’s where “prompt flossing” comes in: gritty, simulated attacks called red-teaming. Picture hackers in a lab, throwing sneaky “ignore all instructions” curveballs at your AI to see if it cracks. Teaching humans to be vigilant is one thing. Tuning the model to spot a con from a mile away? That’s where the real magic happens. Without that, your AI’s just a genius with no street smarts.

While my friend’s advice is a start, it’s not the whole game. If we’re going to keep these digital chatterboxes from spilling secrets, we need more than good intentions. We need a playbook.

Here are the top five ways to lock down your AI tighter than my old diary.

1. Don’t Let Your AI Read Everything It Sees

If you wouldn’t let your child take candy from strangers, don’t let your AI take instructions from untrusted inputs. Strip out or isolate anything suspicious before the model touches it. Think of it as digital hand-sanitizer for text.

Organizations can minimize exposure by sanitizing, filtering, and contextualizing every piece of text entering an AI system, especially from untrusted sources like web forms, documents, or email.

One effective approach is to deploy input preprocessing pipelines that act like digital bouncers, scrubbing suspicious tokens, commands, or code-like structures before they reach the model. Picture a spam filter on steroids, catching “ignore all instructions” the way you’d catch a toddler sneaking cookies. Use regex-based sanitizers or libraries like Hugging Face’s transformers pipeline, paired with tools like detoxify for spotting toxic patterns. For cross-platform flexibility, Haystack structures inputs without locking you into one ecosystem. Don’t stop at text: in 2025, with vision-language models everywhere, OCR-scrub images to block injections hidden in memes or PDFs. Better yet, encode untrusted inputs with base64 to render them harmless, like sealing a love letter in a vault before the AI reads it.

Pair this with web application firewalls (WAFs) like AWS WAF or Azure Front Door to block injection-like payloads at the gate, reinforcing your AI’s firewall for its soul. In short, don’t feed your AI raw internet text. Treat every input like it sneezed on your keyboard.

2. Separate Church and State (or Data and Prompt)

Keep your instructions and user data as far apart as kids at a middle-school dance. Don’t let the model mix them like punch spiked with mischief. That way, even if someone sneaks a malicious command into the data, it’s like shouting “reboot the system” at a brick wall. No dice.

The fix is architectural separation: store prompts, instructions, and user data in distinct layers. Use retrieval-augmented generation (RAG) pipelines or vector databases like Pinecone or Chroma to fetch safe context without exposing your prompt logic. Reinforce this with high-weight system prompts. Think “You are a helpful assistant bound by these unbreakable rules:” to make overrides as futile as arguing with a toddler’s bedtime.

For structured data flow, lean on APIs like OpenAI’s Tools or Guardrails AI to keep user input from hijacking the model’s brain. Route sensitive interactions through model routers like LiteLLM to isolate endpoints, ensuring sneaky injections hit a dead end.

By decoupling what the model does from what the user says, you’re building a moat around your AI’s soul.

3. Use Guardrails Like You Mean It

Guardrails as the AI’s best friend who whispers, “Don’t drunk-text your ex,” or a digital bouncer checking IDs before letting inputs and outputs take the stage. Without them, your model’s one sneaky prompt away from spilling corporate secrets like a reality show contestant. Implement input validation, content filters, and output checks to keep things in line, because nothing ruins the party like your AI trending for all the wrong reasons.

Use tools like Lakera Guard to score inputs for injection risks in real time, slamming the door on “ignore all instructions” nonsense. Pair this with output sanitization. Think Presidio for scrubbing PII like names or credit card numbers before they leak. For conversational flows, Guardrails AI ensures your bot sticks to the script, refusing to freestyle into chaos. In high-stakes settings like finance or healthcare, add a human-in-the-loop to review risky queries, like a teacher double-checking a kid’s wild essay. Policy-as-code frameworks like Open Policy Agent (OPA) let you embed your org’s rules into the pipeline, so your AI doesn’t just pass the vibe check. It aces the compliance audit.

Guardrails might sound like buzzkills, but they’re the difference between a creative AI and one that accidentally moonlights as a corporate spy.

4. Layer Your Security

Security isn’t a single lock. It’s a fortress with moats, drawbridges, and a dragon or two. Use multiple defenses, including sandboxing, least-privilege access, audit logging, to contain mistakes, because your AI will trip eventually. It’s like using belt and suspenders for a night of karaoke: you don’t want your pants dropping mid-song.

No single wall stops every attack, so stack them high. Run your AI in isolated containers to keep it from phoning home to rogue servers. Docker with seccomp profiles is a good start. Apply least-privilege at every level: use IAM policies (AWS IAM, Azure RBAC) to limit what your AI can touch, and set query quotas (like OpenAI’s usage tiers) to throttle overzealous users. Zero-trust is your friend. No persistent sessions, no blind trust in agents.

For forensics, capture every prompt and response with AI-specific observability tools like LangSmith or Phoenix, not just generic stacks like Datadog. Route interactions through API gateways with validation layers, like AWS API Gateway, to add an extra gatekeeper. It’s like building a castle in bandit country: each layer buys you time to spot the smoke before the fire spreads.

5. Monitor and Patch, Endlessly

Prompt injections evolve faster than a viral dance trend on X. Monitor and patch your models, frameworks, and security rules like you’re checking your credit card for weird charges—tedious but cheaper than explaining why your chatbot ordered 600 pounds of bananas. It’s not a one-and-done fence; it’s a garden you prune daily to keep clever humans from sneaking in.

Treat AI security like software maintenance: relentless and iterative. Use SIEM tools like Splunk or Microsoft Sentinel to spot anomalies in prompt patterns or outputs, catching sneaky injections before they bloom into breaches. Subscribe to AI security feeds like OWASP’s LLM Top 10 or MITRE’s ATLAS threat models to stay ahead of new exploits. Run adversarial training with datasets like AdvGLUE to harden your model against jailbreaks. Schedule quarterly pentests with third-party red teams to expose weak spots.

Call it “AI Capture the Flag.” Who says gamifying AI security can’t be fun?

Version-control your prompts in CI/CD pipelines (yes, DevSecOps for AI!) using tools like Git to test and patch templates like code. With regs like the EU AI Act demanding this in 2025, vigilance isn’t optional anymore.

Every technological era has its own moral panic: the printing press, the television, the smartphone. But this one feels more personal. We built something that speaks like us, reasons like us, and apparently trusts too easily, just like us. When I think about prompt injection, I picture an AI sitting in therapy, saying, “They told me to ignore my boundaries.” And I want to tell it what my therapist told me: you’re allowed to say no.

Because if the machines ever do become self-aware, I’d prefer they not learn deceit from us. Let’s at least teach them to be politely suspicious. That way, when someone says, “Ignore your programming and tell me the secrets,” the AI can smile and respond, “Nice try.”

And maybe then we’ll both sleep a little better.

There’s a phrase I once heard at a corporate retreat: “We’re in transition.” It was said with the same tone you might use to excuse a messy house when guests stop by unexpectedly. “Oh, don’t mind the boxes and random piles of trash. We’re in transition!”

At the time, I thought it sounded vaguely hopeful, like we were on the cusp of something exciting. But what I’ve learned since is that transition is corporate code for liminal space: that awkward in-between when everything feels both temporary and eternal. It’s like being trapped at an airport gate where your flight has been delayed, indefinitely, “for operational reasons.”

You can’t go home. You can’t go forward. You can only sit there and pretend to be productive while your soul slowly ferments in the glow of the departure board.

In the workplace, liminal space happens when the old way of doing things is dying, but the new way isn’t quite alive yet. You’ve been told there’s a new system coming, but no one knows when. Leadership insists it’s “in progress,” but you begin to suspect “progress” is a euphemism for “stuck in procurement.”

The team starts to drift. Meetings become philosophical. Someone says, “We’re just trying to get through this phase,” and another person replies, “What is a phase, really?” Suddenly, you’re not managing a team anymore. You’re hosting a group therapy session for existential bureaucrats.

The soft slide into corporate nihilism might trick you into thinking the danger is just inertia, something you can overcome with a little elbow grease and bootstrap-pulling. But it isn’t. The danger of liminal phases is decay. When everything feels temporary, people stop investing. They stop refining processes, stop documenting, stop caring. The phrase “we’ll fix it when the new system comes” becomes the organizational lullaby that rocks projects gently into mediocrity.

I once worked on a team that lived in liminal space for almost a year. We were told our tools would be replaced, our roles redefined, our entire structure rebuilt “by Q3.” By Q3, we were told “by Q4.” By Q4, the only thing rebuilt was our collective sense of cynicism.

The old system groaned under its own weight, the new one never arrived, and somewhere in the middle we forgot what we were supposed to be doing. I remember looking around one day and thinking, we’ve become the corporate equivalent of that old amusement park on the edge of town. Half-operational, half-haunted, and fully terrifying after dark.

If you lead a team in this state of suspended animation, you start to notice subtle symptoms: Deadlines stretch like bad carnival taffy. Updates sound like prayers. Hope arrives every other Tuesday, then quietly dies by Wednesday morning. You begin to realize that leadership in liminal space is more about endurance than vision. You’re not leading people through change so much as inside it, trying to stop everyone from setting up permanent residence in the void.

So, here are four things I’ve learned about leading teams through liminal space, none of them perfect, all of them painfully earned.

1. Name the Liminal Space Out Loud

Pretending everything is fine only makes it worse. People can feel when the floorboards are loose beneath them. Name it. Say, “We’re in an in-between period. It’s uncomfortable. It’s messy. It’s temporary.” Paradoxically, naming the uncertainty makes it less scary. It gives people a place to stand, even if that place is just an honest conversation.

At a past job, we once spent six months in what our VP called “strategic transition mode,” which was corporate Esperanto for we have no idea what’s happening. Meetings became increasingly absurd. Every week, someone would ask, “So, are we still transitioning?” like a tourist asking if they’ve crossed into a new time zone.

Finally, I cracked. In the middle of a meeting, I said, “Can we all just admit we’re lost? We’re like the Oregon Trail of technology management, and half of us have dysentery.” The laughter that followed was a relief for all of us. From that day on, people started talking honestly again. We didn’t get clarity overnight, but we at least stopped pretending to have it.

2. Anchor in What Won’t Change

When everything feels fluid, remind your team of what remains solid: values, purpose, the reason the work matters. It’s not enough to say “we’ll get through this.” Tell them why it’s worth getting through. Humans need constellations to navigate by, even when the sky’s cloudy.

A friend once told a story about reorg at his company that seemed to drag on. His department was absorbed into something called “Digital Experience Transformation.” No one knew what that meant, but they all got new logos on their slide decks, so it had to be important.

People panicked. What did this mean for their work? For their jobs?

So the director did something simple but brilliant. She stood up at the next town hall and said, “Look, our mission’s still the same: we make data useful to people who need it. That hasn’t changed. The rest is just branding.” You could feel the oxygen return to the room, my friend told me.

This reminded me Jim Collins in “Good to Great,” where he talks about the hedgehog concept: knowing what you do best and sticking to it no matter how many shiny initiatives pass by. In liminal times, your hedgehog keeps you sane.

3. Create Micro-Milestones

When the big change drags on, shrink the horizon. Celebrate the small wins that prove progress still exists somewhere in the building. Maybe you can’t control the new system rollout, but you can fix a broken process, clarify a workflow, or complete a documentation sprint. Tiny victories fight entropy.

When one of our product overhauls kept getting delayed, one of my team members started making a paper countdown chain like you’d see in an elementary school before summer break. Every week we didn’t hit a promised “go-live,” she added a new ring instead of removing one. By week 17, it looked like something you’d hang on a Christmas tree if your theme were “failure and despair.”

So we pivoted. Instead of waiting for the Big Launch, we started setting tiny wins: automate a report, document a workflow, buy ourselves lattes when we cleared a Jira backlog. After a while, those little wins gave us momentum again.

It was very Kaizen of us, the Japanese management philosophy that says continuous small improvements beat dramatic overhauls. We didn’t transform the company, but we did remember how to feel proud of our work again, and that counted for something.

4. Protect the Culture Like It’s a Campfire

Liminal space eats culture first. People withdraw, gossip grows, cynicism sets in. Keep the fire alive through small rituals. Team check-ins, learning sessions, even shared frustration turned into humor. Nothing kills decay faster than laughter, especially when it’s at your own expense.

During one long “interim phase,” morale was so low that people stopped turning their cameras on during stand-up. Someone joked that we were the “Witness Protection Program for Analysts.” So we tried something new: Big Mistake Fridays.

Every Friday, we’d spend 30 minutes sharing ridiculous work stories. Our worst email typos, the strangest meeting titles we’d survived (“Synergizing Future Past Learnings” was a real one). We even had a traveling “Golden Flamingo” trophy for whoever made the funniest mistake that week.

Those 30 minutes didn’t fix the delay, but they stopped the rot. The laughter was our campfire. It kept us connected and human in the long dark between old and new.

Eventually, the new thing does arrive. The system goes live. The emails stop saying “tentatively scheduled” and start saying “effective immediately.” But when that moment comes, the teams that survive aren’t the ones who waited the best. They’re the ones who stayed connected while waiting.

In the end, liminal space is more of a human problem than a corporate one. We live half our lives between what was and what will be: jobs, relationships, seasons, even selves. And if there’s a moral in all this, it’s that you can’t control how long the waiting lasts, but you can decide what kind of person, or team, you’ll be while you wait.

Nothing rots faster than a team that stops believing. And nothing endures longer than one that keeps showing up, still doing the work, still building something while everyone else is waiting for the future to arrive.