Future Companies Need Only Two Kinds of People: Goal-Setters and Hard-Core Geeks
Developers will see jobs split into two tracks: those who can set meaningful goals and those who can solve problems AI can't touch. The path forward isn't to out-code AI but to move up the abstraction ladder, define objectives, and architect agent systems that execute them.
Large language models now write code, documentation, and unit tests so fast that they flood the workplace. That leads to a simple question: if an organization's intelligence can be provided by AI, what do humans do? One answer: humans become the goal-setters, because LLMs have no intrinsic needs and therefore no motivation to set goals. This shifts the organizational chart to two roles: "deep-tech weirdos" who wrestle with the hardest problems, and "bosses" who define goals and manage agent teams.
To frame this, the author abstracts a worker's job into harness (codebases, permissions, docs), loop (daily check-in), and graphs (understanding organizational relationships). LLMs already handle the harness and loop; they only lack the goal. So future companies will give AI agents harness, loop, and graphs, plus a human-provided goal, and have humans review the output. Along the way, salaries will drop as AI replaces part of your intelligence, not suddenly, but gradually.
The article also speculates on an omniscient AI, arguing it may be impossible because rigorous logic and creative design are fundamentally opposed, as seen in GPT vs. Claude's strengths. The ultimate vision is a merged human-AI intelligence spreading across the universe.
The harness/loop/graphs framework is a useful decomposition for automation, but it reduces a worker to an input/output system, ignoring the irregular, social problem-solving that doesn't fit a formal schema.
The claim that AI can't set goals because it lacks Maslow needs conflates goal-setting with intrinsic motivation; organizations routinely impose goals on both humans and AI, so the real gap may be judgment and accountability.
The observed split between GPT and Claude strengths is anecdotal evidence of a possible trade-off between logical rigor and creativity, but it doesn't rule out a future model combining both.