Y Combinator Co-founder Paul Graham: In the AI Era, Focus on Technology or Customers
Y Combinator co-founder Paul Graham responded to a computer science undergraduate: to have the greatest impact in the era of artificial intelligence, one must stand at either end—either be close to the technology and truly work on large language models; or be close to the customers and use AI to deliver exactly what they want. If one can stretch both ends and reach them simultaneously, that would be even better.
Elon Musk retweeted and replied "True." Graham left the middle ground blank: a position that neither builds models nor directly serves users is implied to have less impact. The two ends correspond to two sets of scarcity: engineering judgment for training and alignment, and shaping model capabilities into a product that a specific person is willing to pay for.
This aligns with his consistent entrepreneurial writing: create what users truly want, rather than middleware that looks like the future. The undergraduate asked about career leverage, and he provided a choice of distance—close to weight updates or close to payers. Musk's agreement maps the same structure to his own path: building models and computing power, while also creating products for car owners, users, and advertisers.
"Stretching both ends and reaching them simultaneously" is a narrower third path: laboratories can hear customer sentences, and sales can adjust model objectives. Most organizations separate research from customer success, while Graham identifies those who can bridge the gap as sources of exceptional impact.
In market mechanisms, this is a talent pricing signal. Buyers are cutting-edge laboratories and application companies that can turn models into workflows; sellers are those squeezed into prompt packaging, industry reports, and pure middleware roles. Compensation and equity will concentrate on those who can train models and close deals. Beneficiaries will be model teams and industry vertical applications, while those who neither touch loss curves nor customer bills will be under pressure.
On a supplementary level, Graham did not name specific companies or salary figures. Both ends are changing rapidly: the model side requires computing power and data, while the customer side needs licensing, integration, and taste. Those who can reach both ends simultaneously are still very few.
Source: Public Information
ABAB AI Insight
Graham uses distance instead of an industry list. The undergraduate wanted a map, and he provided two magnetic poles. Building models is a supply-side leverage, while delivering to customers is a demand-side leverage; the middle layer will be compressed as model prices drop. Y Combinator has long made "creating something users love" its first principle, and this time simply marks both users and models as reachable poles.
Capital has already raised funds along these two poles. Base models consume training budgets, while vertical applications consume customer budgets; the "AI-enabled platforms" caught in the middle must prove they are not easily replaceable skins by model manufacturers. Musk's "True" response is because Tesla, xAI, and platform businesses are trying to integrate both poles into the same organizational chart, rather than choosing sides.
A similar structure existed in the early internet: building browsers and search engines, or e-commerce and customer-facing pages; the middle portals later became thinner. The mobile era repeated this once: building systems or super apps. The industry is currently in a phase where model capabilities are rising, but scenarios where customers are willing to pay remain scarce. Impact will fall on those who can adjust loss functions and those who can modify deliverables, rather than those who can hold seminars.
Essentially, this is a restructuring of the industrial chain. The mechanism is: once intelligence becomes a tunable commodity, value concentrates at the ends—those who produce it and those who turn it into orders; the middle layer only has rent when both ends cannot reach each other, and once both ends shake hands, the rent disappears.
ABAB News · Cognitive Laws
Leverage is at the two poles, not in the commentary between them
Either change the model or change what the customer receives
Those who can reach both ends simultaneously are more valuable than those standing in the middle