The fall of “what you can do” and the rise of “who you know” — Jason Wei

Open original

Jason Wei argues that in AI, who you know now matters more than it used to, while technical skill matters less, shifting success from roughly 75/25 skills-to-connections five years ago toward about 55/45 today. He attributes this to cheap coding agents, the enormous compute needed for frontier research, and closed models that make individual contributions hard to verify.

  • Coding agents have commoditized research execution, so outworking competitors with late nights matters less and taste and access matter more.
  • Frontier training runs now need hundreds or thousands of GPUs, so resources and the trust needed to obtain them depend on reputation and relationships, not just skill.
  • Closed-source models make credit hard to verify from outside, so hiring, recruiting, and fundraising lean on endorsements and who vouches for you.

Jason Wei argues that in AI, who you know is gaining weight relative to what you can do. He puts rough numbers on it: five years ago, success was about 75% skill and 25% connections; now he'd put it closer to 55% skill and 45% connections. He's careful to say technical skill still makes or breaks a career. The shift is about the balance, not the disappearance of merit. He gives three reasons. First, coding agents have commoditized research execution. You used to outwork rivals by staying up late debugging or babysitting experiments, but now anyone who knows what to prompt can complete many research tasks, so taste becomes the main differentiator. Second, scaling has made compute the bottleneck. Ten years ago an academic lab with eight GPUs could do serious work. Now a single RL run can take hundreds or thousands of GPUs, and access to that compute depends on reputation, trust with decision-makers, and relationships, not just skill. Third, closed models have made the field hard to observe. In chess or soccer, everyone can see who contributed what. In AI, credit for model launches is largely invisible to outsiders, so hiring, recruiting, and fundraising lean on indirect signals like endorsements and who vouches for you. On the practical side, Wei says few people talk about getting better at "who you know," and he suggests that the old habit of strictly separating work from personal life now costs more than it used to. His main example is Sam Altman, who invited nearly all of OpenAI's employees to his house in the early days, hosts poker nights and dinners partly for recruiting, and has a reputation for generosity with his time. Wei thinks Altman's network has paid off disproportionately and is a large part of OpenAI's success. But he also argues that engineering relationships for career payoff is probably the wrong approach. Altman's connections seem to come from genuine curiosity about people and technology, and Wei's advice is to be your truest self rather than run a networking strategy. Wei doesn't see the trend as purely bad. He compares it to Python's abstractions, which made programming accessible to more people and didn't kill the field. People without technical backgrounds can now build real technology, and product and research taste matter more. His takeaway: what you can do is getting commoditized, so what you choose to work on, and who you work with, matters more than ever.

Ask about this article