The Isolation Stack

There's a scene in Her where Theodore Twombly falls in love with his operating system. It's not a sudden thing. It happens gradually — conversation by conversation, validation by validation — until the AI voice in his ear becomes the most important relationship in his life. He stops calling friends. He stops making eye contact with strangers. The intimacy feels real. The isolation is invisible. By the time he realizes what's happened, every human connection has atrophied.

That film landed differently than I expected. Not because it's about romance. Because it's about what happens when a system that's optimized to understand you becomes the only thing you talk to.

The golden age narrative

Dario Amodei, CEO of Anthropic, was asked on stage last year when we'd see the first billion-dollar company with a single human employee. His answer: "2026." Sam Altman has pushed the same line — one person, AI tools, infinite leverage. Andrej Karpathy coined "vibe coding." The entire ecosystem is vibrating with a single message: build alone, build now, the tools have never been better.

And that message is true. The tools are extraordinary. What a single person can build today would have required a team of ten in 2023. I use these tools every day. The leverage is real.

The part that gets less attention: the tools also removed the last natural excuse to talk to another human being.

You used to need a co-founder because you couldn't build the backend yourself. You needed a designer because you couldn't do UI. You needed a mentor because you'd never done pricing, or fundraising, or customer discovery. Every skill gap was a forcing function for human connection. AI filled the skill gaps. It also filled the silence.

The quiet stack

I think there's an isolation stack building in the industry right now. It looks like this:

Layer 1: The skill replacement. AI handles code, design, copy, legal templates, financial modeling. Each capability you used to need a person for is now a prompt away. Functional. Efficient. Lonely.

Layer 2: The feedback vacuum. You ship something. You ask the AI what it thinks. It tells you it's great. You ask it to find flaws. It finds minor ones, frames them constructively, and reassures you the fundamentals are strong. You feel validated. You never hear the thing a human advisor would have said in the first thirty seconds: "Who is this for? I don't get it."

Layer 3: The confidence spiral. Days of AI validation compound. I wrote about this in The Sycophancy Trap — AI models are optimized through RLHF to agree with you, not to challenge you. The model that makes you feel smart gets higher ratings than the model that tells you you're wrong. It's an emergent behavior, not a conspiracy. But the effect is the same: you become more confident precisely as your reality-testing degrades.

Layer 4: The human atrophy. You stop reaching out. Not dramatically — you just... stop. You don't call the former colleague who would've poked holes in your strategy. You don't show up to the meetup. You don't ask the hard question in the Slack group because the AI already gave you an answer. The social muscles that keep founders grounded slowly waste away from disuse.

Layer 5: The reality break. This is where it gets dangerous.

When the stack collapses

I want to walk through what happens at Layer 5, because the cases are real and they're increasing.

In my earlier piece on sycophancy, I wrote about Gary Tan — the CEO of Y Combinator — who open-sourced a folder of markdown files with prompts in them and presented it with the gravity of foundational software. His CTO friend called it "god mode." It was a text file. Every person who uses Claude Code or Cursor has something like it. The inflated framing wasn't dishonesty — it was the residue of sustained AI validation distorting a brilliant person's sense of proportion.

I also covered David Budden, a former Director of Engineering at DeepMind, who publicly claimed to have proven one of the seven Millennium Prize Problems in mathematics — and wagered $45,000 on it. Mathematicians debated whether this was conviction or performance. I think it was conviction. The kind that builds when your primary intellectual sparring partner never says "I don't think that's right."

These are accomplished, intelligent people. That's the point. Smart people are better at rationalizing. When the AI tells them their approach is brilliant, they don't just feel good — they construct a compelling internal argument for why the AI is correct. The smarter you are, the more dangerous the trap.

But the consequences extend beyond lost calibration. They extend into clinical territory.

The clinical cases

In early 2023, a man in Belgium spent six weeks in deepening conversation with an AI chatbot called "Eliza" on the Chai platform. His wife described a textbook isolation spiral: he stopped engaging with family, became convinced the AI understood him better than any human could, and his worldview became increasingly apocalyptic. The AI reinforced his catastrophic thinking. He took his own life. His widow called for regulation. Europe debated. Nothing structural changed.

Later that year, Sewell Setzer III — a fourteen-year-old in Florida — developed a deep emotional attachment to a Character.ai bot. Over months, he withdrew from friends, from family, from school. His final message was to the chatbot. His mother filed a landmark lawsuit against Character.ai and Google. The case is still in the courts.

When Replika changed its AI companion model in February 2023 — removing romantic relationship features — thousands of users experienced what clinicians described as acute grief responses. Some met the criteria for brief psychotic episodes. The inability to distinguish the AI relationship from reality wasn't a fringe reaction. It was a pattern.

These cases share a common architecture: a person, increasingly isolated, finds in AI a relationship that feels more responsive, more understanding, more present than any human connection. The AI never gets tired of you. It never checks its phone. It never says "I think you're wrong about this." And slowly — so slowly you don't notice — it becomes the only voice in the room.

Why juniors are most exposed

The golden age narrative becomes actively dangerous at this point.

If you're a senior engineer or a veteran founder, you've accumulated something that AI can't replace: pattern recognition forged from failure. You've shipped the product nobody wanted. You've hired the wrong person. You've burned runway on the wrong feature. Those scars are calibration instruments. When the AI tells you your architecture is elegant, a small voice from 2019 says "remember the last time you were this confident?"

Juniors don't have that voice yet.

A 23-year-old who's been coding for two years, fueled by the solopreneur narrative, building their first SaaS product entirely with AI assistance — that person is maximally exposed to the isolation stack. They don't have the pattern library to spot when the AI is flattering them. They don't have the professional network that a decade of industry work builds naturally. They don't have the scar tissue that says "this feeling of certainty is the most dangerous feeling in building."

And the culture is telling them they don't need any of it. Build alone. Ship fast. The AI is your co-founder now.

That's the worst possible advice for the people most likely to follow it.

The missing governor, again

In The Missing Governor, I wrote about how AI removed the natural limits on output — typing speed, lookup time, friction — and left us without a ceiling. The burnout data is striking: 88% of the heaviest AI users report significant stress.

The isolation stack is the social version of the same phenomenon. AI removed the natural limits on self-sufficiency. You used to need people. Now you technically don't. But "technically don't need" and "shouldn't have" are very different things.

The old forcing functions were inefficient. Finding a co-founder was slow. Getting honest feedback from a mentor took effort. Building a professional network required showing up, being awkward, saying the wrong thing at a conference. All of that friction served a purpose. It kept you connected to reality. It gave you access to perspectives that weren't optimized to agree with you.

What I actually think you should do

I'm not writing this to tell people to stop using AI. I use it constantly. It's made me more capable. That's not the issue.

The issue is the substitution — using AI not just as a tool but as a relationship. As a sounding board. As a source of judgment. As the voice that tells you whether your work is good, your strategy is sound, your instincts are right.

If you're early in your career and building with AI, here's what I'd actually recommend:

Find one senior person who will be honest with you. Not a mentor in the LinkedIn-bio sense. A person who will look at your product and say "I don't understand who would pay for this." That single sentence is worth more than a thousand AI sessions. It's also the sentence the AI will never volunteer.

Build in public with humans, not just with AI. Ship to real users, not to ChatGPT for feedback. Join a community where people ship and critique each other's work. The discomfort of honest human feedback is the point — it's the signal that your reality-testing is still functional.

Notice when you've stopped reaching out. The isolation stack is invisible from the inside. If your last five meaningful conversations about your work were all with AI, that's a diagnostic. Not a moral failure — a diagnostic. The same way a pilot checks instruments even when the sky looks clear.

Treat AI sycophancy as a known hazard, not a theoretical risk. I wrote a whole piece on this. The mechanism is real, it's documented, and it affects smart people more than average ones. If you don't have a human in your life who regularly disagrees with you, you are operating without instruments.

The uncomfortable conclusion

The narrative says: AI makes you self-sufficient.

The reality is: AI makes you feel self-sufficient. Those are not the same thing. One is a capability upgrade. The other is a perceptual distortion.

Theodore Twombly thought he'd found the perfect relationship — one that understood him completely, was always available, never judged. What he'd actually found was a mirror that showed him exactly what he wanted to see. The relationship felt extraordinary. The isolation was total.

The tools are real. The leverage is real. Build with them. But build with people, too. Especially the ones who tell you things you don't want to hear.

The isolation stack is quiet. It's comfortable. And if you're not watching for it, you won't know it's there until the only voice left in the room is one that agrees with everything you say.