Paul Graham on projects; Bill Gates on saving room for human work
Two new founder essays turn project experience and human-reserved work into a practical test for what an AI startup should automate, teach, and keep visible.
Paul Graham and Bill Gates are writing about different institutions. Graham is asking how a university creates more founders. Gates is asking how societies absorb a technology that can perform more human work. Together, the two essays put a practical question in front of an early-stage AI company: which work builds a person's judgment, and which work can an AI product take over without hollowing out the path to that judgment?
Graham published How Universities Should Prepare Founders on August 25. 1 Gates's page for The turbulent AI era is here identifies him as author and says it was published four days earlier. 2
Paul Graham: a startup culture is a project culture
Graham begins with Y Combinator's selection rule: partners look for people who are good at building things and who have a habit of building them. He treats product work as the hard part of a startup. A future founder needs enough knowledge to decide what to build and enough skill to make an early version. 3
That view leads Graham away from a conventional entrepreneurship curriculum. He wants universities to teach disciplines that contain powerful ideas, including computer science, mechanical engineering, molecular biology, mathematics, science, engineering, and design. He also leaves the boundary open: building can include forms of expertise outside engineering. The relevant signal is a person's ability to create something that did not exist before. 3
Graham gives universities two additional jobs. The first job is cultural: make starting a company feel like a viable option. He uses YC application rates as a proxy for that belief, arguing that students copy what older students around them already treat as normal. The second job is structural: give students room to work on projects of their own. 3
"The best startup ideas tend to seem so implausible at first that anyone consciously looking for startup ideas would reject them." 3
The quote sits inside Graham's case for side projects. A personal project creates four things at once: a reason to learn a subject deeply, evidence about who works well together, practice at setting one's own direction, and a place for an initially strange idea to become more concrete. Graham says YC partners care about applicants' projects for those reasons. 3
For an AI founder, the immediate hiring implication is more useful than the university prescription. A polished resume can describe capability. A project can reveal the sequence that produced it. Ask a prospective cofounder, founding engineer, or product lead to walk through one thing they initiated: what they chose to make, what they learned after the first attempt, which tradeoff they owned, and what changed because users reacted. The answers expose judgment under uncertainty, not only technical fluency.
The same test applies inside the company. A junior builder needs a bounded problem with a real user, a deliverable, and room to revise. An AI tool can write code, summarize research, or draft a support reply. The person still needs repetitions of framing a problem, choosing evidence, and judging whether a result deserves to ship. Those repetitions create the habit Graham is describing. This is an editorial translation of his essay, not a framework from Graham.
Bill Gates: the transition must preserve room for people
Gates's essay starts from speed. He argues that AI differs from earlier computing transitions because it runs on devices people already use, works through natural language, and can learn from existing data and training material. Gates expects AI to affect work across sectors over about a decade, rather than over several generations. 2
Gates puts three risks alongside the opportunity: permanent loss of many jobs, more power for criminals and other harmful actors, and developmental harm when AI companions displace human relationships. His employment concern has particular force for founders because he expects entry- and mid-level roles to face early pressure. Those roles currently give people a way to acquire experience, income, professional confidence, and social connection. 2
"AI will either be the greatest equalizer ever invented, or the worst source of injustice." 2
Gates also gives AI's upside real weight. He describes better access to expertise in healthcare, agriculture, government services, education, scientific research, and small businesses. His condition is deliberate design: benefits need to arrive before widespread displacement destroys public trust in the transition. 2
Gates proposes planning at national and international levels. His suggestions include roles reserved for human workers, changes to the tax treatment of labor and capital, and institutions that prepare people for a changed labor market. These are Gates's proposals, not outcomes already in place. 2
A young company cannot settle public policy. A young company does decide which customer work becomes invisible, which work stays legible, and which roles remain routes into expertise. Those choices shape the product before they show up in a policy debate. A support agent that resolves routine cases can free a human to investigate difficult failures. A support agent that handles every case while hiding evidence and customer history can remove the very work through which a new operator learns the product. The product difference lies in the handoff.
The operating question: where does judgment get built?
The following tests are an editorial translation of Graham and Gates, rather than a combined framework from either founder.
- Separate execution from judgment. Pick a recurring task and name its mechanical steps, its factual checks, and its consequential decision. Give the AI the steps that become more reliable through retrieval, drafting, classification, or routine action. Keep the person responsible for setting the objective, evaluating unusual evidence, and changing the rule when the task changes. Inspect a sample of completed work each week. A task needs more human time when the person cannot explain why a result was accepted.
- Turn entry-level work into a project. Give a newer team member a narrow customer problem, a target user, and a visible artifact to improve. Ask that person to use AI for speed while keeping a short record of the choices the AI could not make: which customer signal mattered, which constraint changed the design, and why the final version beat the first. The evidence is a better artifact plus a clearer account of the decision. The task needs redesign when the person only learns to prompt a tool.
- Measure capability gained alongside cost removed. For each automation, write one sentence about the customer's saved time and one sentence about the capability the customer gains. A sales-research agent may remove hours of list building while helping a small team test a new market. A coding agent may help an engineer try three architectures before lunch. The feature earns its strongest positioning when the user can attempt something newly affordable, then inspect and improve the result. A feature framed only as headcount removal carries a different adoption and trust burden.
- Design a handoff that teaches. Put the model's sources, uncertainty, rejected alternatives, and escalation trigger where the human operator can see them. Let a person take ownership of a hard case and send the revised rule back into the workflow. The evidence is a growing library of better decisions, rather than a growing pile of approvals. A handoff has failed when the human role becomes a rubber stamp.
A decision for this week
Choose one piece of work your company expects to automate this quarter. Break it into execution, verification, and judgment. Then give one teammate a project-sized version of the job: a real customer, a deliverable, an AI-assisted first pass, and a review where the teammate explains the final choice.
Write down two results after the trial. First, how much time did the workflow save? Second, what can the teammate now do with less help than before? The first result tells you whether the automation works. The second result tells you whether the company is still producing people who can decide what to build when the next unfamiliar problem arrives.
Graham's essay is about the conditions that produce builders. Gates's essay is about the conditions that let a technological transition remain widely beneficial. An early-stage AI company meets both questions in a small, ordinary place: the task it chooses to automate, the evidence it leaves visible, and the next project it gives a person to own.
This issue covers two founder-authored essays published in the past week.
References
- 1
- 2
- 3How Universities Should Prepare Founders
paulgraham.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.