Karpathy’s AI job exposure scores for software engineers and what an 8.5/10 rating actually means for how engineers should be investing their time
Karpathy Just Scored Your Job. Software Engineers Got an 8.5. Here’s What That Actually Means.
Andrej Karpathy built a tool that scraped all 342 occupations from the Bureau of Labor Statistics, fed each one to an LLM with a detailed scoring rubric, and produced an AI exposure score from 0 to 10. Software developers landed at 8 to 9. The average across all 342 occupations is 5.3. Roofers are at 0 to 1. Medical transcriptionists hit 10. We’re sitting two points below the ceiling.
I’ve watched the reactions pour in for days now. Most of them are useless.
What an 8.5 Is Not Telling You
The doom camp reads 8.5 and hears “your job is 85% gone.” The denial camp reads it and says “AI still can’t architect a distributed system, so relax.” Both camps are arguing about the wrong thing.
The score is not a probability of replacement. It’s a measure of how much of the work is digital in nature and executable without physical presence. Karpathy’s rubric captures this directly: if the work product is fundamentally digital and the job can be done entirely from a home office, exposure is inherently high. Software development checks both boxes. Of course it does.
The real question is what the 8.5 is pointing at inside your day.
The Parts That Deserve a 10
Some of what I do as an engineer absolutely deserves a 10. Boilerplate scaffolding. Repetitive data transforms. Writing the same three utility functions for the fourth time across a new project. Generating test fixtures. Translating a known algorithm into a new language. I am not mourning any of that work.
Boris Cherny, who built Claude Code, runs 10 to 15 Claude sessions in parallel every single day. He told people he hasn’t written a single line of SQL in over six months. Claude pulls the BigQuery data directly via CLI. Claude Code now accounts for 4% of all public GitHub commits. That number is real and it keeps moving.
The mechanical, repeatable, “I know exactly what this should look like before I start typing” work is effectively gone. Good. I never wanted it anyway.
The Parts That Score a 3
Here’s where engineers are making a career mistake if they’re not paying attention. The work that scores low on exposure is the work that actually drives outcomes.
Why is this distributed system producing latency spikes only under a specific load pattern? What does this architecture decision cost us in 18 months when the team doubles? Why did this model start drifting and what does that tell us about the data pipeline upstream? Those are not LLM-native problems. They require accumulated context, judgment built from failure, and the ability to hold ambiguity without collapsing it prematurely.
Alibaba ran a study testing 18 AI agents on 100 real codebases over 233-day maintenance cycles. 75% of models broke previously working code during maintenance. Only Claude Opus 4.5/4.6 maintained a greater than 50% zero-regression rate. Every other model accumulated technical debt until things collapsed. Writing code and maintaining a living system are completely different skills, and the current generation of models is mostly good at the first one.
Where Engineers Should Actually Invest Their Time
Stop optimizing for the skills that score a 9. Start building the ones that score a 3.
System-level reasoning. The ability to read a production incident and reconstruct causality. Understanding what a product actually needs versus what was asked for. Knowing when to push back on a technical direction and how to make that argument to someone who doesn’t write code. These are the surfaces where AI is genuinely weak, and they’re also the surfaces that senior engineers have always been compensated for.
The engineers I see thriving right now are not the ones who avoided AI tools. They’re the ones who offloaded the 9/10 work completely, bought back hours, and spent those hours on the judgment-heavy problems that nobody else on the team wanted to sit with.
The score is a map. Most people are reading it like a verdict.
🎯
The 8.5 is not a countdown timer. It’s a signal about where leverage lives in your job. The engineers who treat it that way will be fine. The ones who either panic or dismiss it entirely are making the same mistake, just in opposite directions.
Sources & Further Reading
#SoftwareEngineering #AITools #MachineLearning #CareerDevelopment #Karpathy
