Andrej Karpathy quietly dropped a tool this week that I haven't been able to stop thinking about. No paper. No press tour. Just an interactive treemap of 342 occupations covering 143 million jobs, colored by how much AI is about to reshape each one.

I've been staring at it from both sides of my brain: the engineer side that built software for years, and the VC side that now bets on the people building it. Honestly, the takeaways are uncomfortable for both.

Here's the thing. Buried in Karpathy's methodology is what I'm calling the "Home Office Test," and it's brutally simple:

If your job can be done entirely from a home office on a computer, writing, coding, analyzing, communicating, your AI exposure is inherently 7+ out of 10.

That's it. That's the whole heuristic. Not "does your job involve repetitive tasks" or "is your work creative enough." The question is: does your job live entirely in the digital domain? If yes, AI is coming for it, because that's exactly the domain where AI capabilities are compounding fastest.

This flips the narrative I've been hearing at every LP dinner and founder pitch for the past two years. Knowledge workers told themselves they were safe. AI would automate the factory floor, the warehouse, the call center. Meanwhile the creative class, the designers, the developers, the analysts, would thrive.

Does it actually work that way though?

Karpathy scored each occupation on a 0 to 10 scale. Let me walk you through the highlights, because they read like a roast of my entire professional network:

Software developers? 9 out of 10. Graphic designers, translators, data analysts, paralegals, copywriters? All 8 to 9. Basically the entire laptop class, sitting in the blast radius.

Registered nurses, police officers, veterinarians? A moderate 4 to 5. They blend knowledge work with physical presence and human judgment, so AI can assist but can't replace the core job.

Electricians, plumbers, firefighters? A comfortable 2 to 3. Hands on work in unpredictable environments. AI might help them schedule jobs or look up code specs. It's not climbing under your sink.

Roofers, landscapers, commercial divers? Basically a zero. If your job requires being on a roof in the rain, AI can't touch you.

The irony here is thick enough to spread on toast. The people who built AI tools are among the most exposed to them. The plumber who fixes your sink scores a 2. The developer who built the app to find that plumber scores a 9. I've watched this play out from both sides of the table, and frankly, it tracks.

Karpathy includes a critical caveat, and I want to be fair to it: high exposure doesn't mean the job disappears. He points out that software developers score 9/10, but demand for software could easily grow as each developer becomes more productive.

This is the right framing. I believe it. But "reshaped" is doing a lot of heavy lifting in that sentence.

Here's the thing I keep thinking about from the investing side: when every developer ships 3 to 5x more output with AI assistance, companies don't necessarily hire 3 to 5x more developers. Some will. Most won't. "Reshaped" often means fewer people doing more work, which is great for the people who remain and rough for everyone else.

I've seen this pattern before in tech. It's basically what happened with DevOps. Automation didn't kill ops jobs overnight. It just meant one person could do what five used to, and the other four had to level up or move on. Now multiply that across every knowledge work category scoring 7+.

Yep, pretty much.

Karpathy is upfront that his framework ignores several factors: demand elasticity (will cheaper software create more demand for it?), latent demand (are there jobs AI unlocks that don't exist yet?), regulatory barriers (will governments slow this down?), and social preferences (do people want a human doctor even if AI is technically better?).

These aren't small gaps. They're the entire gap between "high exposure" and "actual job loss." And if you're building a startup right now, they're your opportunity.

Every technological revolution has created categories of work that didn't exist before. The internet didn't just kill travel agents. It created an entire ecosystem of roles nobody anticipated. The same will likely happen with AI. We just can't see those roles yet. Whether that's sufficient to absorb the displacement remains an open question, and anyone who tells you they know the answer is selling something.

Here's what really got me. Karpathy built this entire analysis by having an LLM score every occupation's AI exposure. He wrote custom prompts with calibration anchors, ran them across 342 jobs, and visualized the output. One person. One weekend (I'm guessing). A labor market analysis that would have taken a research team weeks.

He used AI to map AI's impact on work. The tool demonstrated its own thesis in real time.

If you're a founder pitching me on AI productivity tools, this is the slide you should steal. Not a hypothetical. Not a McKinsey projection. One person, shipping a comprehensive labor market visualization, because AI made it possible.

If you're a developer or knowledge worker: stop debating whether AI will affect your job. It will. Start thinking about how fast you can integrate AI into your workflow. I've been doing this myself, sometimes clumsily (my laptop is a laboratory of AI experiments at this point), and the gap between AI augmented and non augmented workers is going to be the defining career differentiator of the next decade.

If you're making career decisions or advising someone who is: the safest jobs aren't the highest paid ones. They're the ones that require physical presence in unpredictable environments. That's not a reason to become a roofer. But it is a reason to value skills that can't be digitized: physical judgment, human connection, navigating real world complexity.

If you're building AI products: the addressable market isn't "everyone equally." It's concentrated in that 8 to 10 exposure band, knowledge workers who live on laptops and are already feeling the pressure. Build for them. That's where the pull is strongest.

Karpathy's visualization is exploratory, not predictive. He says so himself. But the patterns it surfaces are hard to unsee: the more digital your work, the more AI will reshape it. The trades, the hands on professions, the jobs that require showing up in the physical world, those are the ones AI barely touches.

As for me? I'm still processing what it means to invest in companies building tools that score a 9 on the thing they're disrupting. It's like a smoothie, but instead of fruits and veggies, it's existential career questions. Refreshing and unsettling in equal measure.

— SWEdonym