Andrej Karpathy's 10 Jobs That Won't Survive AI (and 5 that will)
Summary
- Andrej Karpathy's AI exposure report reveals that high-paying, computer-based jobs (software engineering, data science, graphic design, transcription) face the highest risk of AI automation, scoring 8 to 10 out of 10.
- Blue-collar and physical labor jobs (plumbing, truck driving, construction, janitorial work) are currently the safest, scoring 0 to 1 out of 10, because robotics development significantly lags behind LLM software capabilities.
- Data from Anthropic confirms that AI already possesses the intelligence to automate significant portions of business, finance, and math roles; the current bottleneck is enterprise diffusion and adoption, not model capability.
- The next frontier of AI is breaking out of the digital box into the physical world via humanoid robots and automated logistics, driven by leaders like Elon Musk (Tesla) and Travis Kalanick (Atoms).
- Jevons Paradox suggests that as AI drastically lowers the cost of intelligence, demand for new types of tech-enabled jobs and overall human productivity will exponentially increase, rather than resulting in permanent mass unemployment.
Trade Ideas
Graphic designers, data scientists... customer service reps... general office clerks. It largely involves computer-based tasks, but ones that are very repetitious and most likely to be automated, kind of like low-hanging fruit. Freelance marketplaces generate the vast majority of their revenue by taking a cut from gig workers performing digital tasks like logo design, basic coding, copywriting, and transcription. Because these specific entry-level digital skills have the highest AI exposure scores, enterprise and retail buyers will increasingly use free or cheap AI agents to do this work instantly, structurally shrinking the Total Addressable Market for human freelancers. SHORT. The core supply-side product of these digital gig platforms is being rapidly commoditized to zero by LLMs. These platforms could successfully pivot into AI agent marketplaces or integrate proprietary AI tools that make their top-tier freelancers exponentially more productive, driving higher project volume that offsets the lower cost per task.
Customer service reps, 2.8 million jobs, nine out of 10... back office work, like moving files around, basic analyst stuff. Business Process Outsourcing (BPO) companies rely on human-in-the-loop models to provide outsourced customer service, call centers, and back-office data entry to large corporations. With conversational AI and voice agents now capable of handling these tasks autonomously 24/7 at a fraction of the cost, enterprise clients will aggressively churn from traditional human BPO contracts in favor of AI software solutions. SHORT. The traditional human-centric outsourced customer service business model is fundamentally broken by the new economics of AI voice and text agents. BPOs might successfully license enterprise AI technology themselves, transitioning from human labor providers to managed AI service providers, thereby protecting their enterprise contracts and profit margins.
These models are so powerful now, and they're so capable, that it's no longer a matter of increased intelligence. It's more a matter of diffusion... getting the AI into these systems and automating them because it exists today. Frontier AI labs (OpenAI, Anthropic) have already solved the intelligence required to automate cognitive labor. The remaining hurdle is distribution. The mega-cap tech companies that own the cloud infrastructure, enterprise software suites, and direct investments in these frontier labs are the ones who will distribute these AI agents to Fortune 500 companies. They will capture the massive financial upside of this diffusion phase. LONG. They are the ultimate toll collectors for the enterprise adoption and integration of AI. Open-source models could commoditize the foundational AI layer, or strict antitrust regulations could prevent these tech giants from bundling new AI agents with their existing enterprise software monopolies.
Once AI breaks out of its box, once there are physical robots kind of moving around... into the Elon-based world... where the AI breaks out of the box. It is becoming physically manifested through these robots. While software jobs are currently highly exposed to AI, physical labor is safe only because robotics is a younger industry. The next multi-trillion dollar opportunity is applying AI "brains" to physical hardware to automate manual labor. Companies with massive manufacturing scale, proprietary real-world vision data, and active humanoid robot programs are uniquely positioned to solve the physical labor shortage. LONG. Tesla's valuation will increasingly be driven by its positioning as an AI and robotics company (via Optimus) capable of automating the physical world, rather than just an automotive manufacturer. Humanoid robotics is incredibly capital intensive and technologically difficult. Timelines for viable commercial deployment may be delayed by decades, or specialized robotics startups could solve the hardware challenges faster.
