Jason Wei argues that connections now account for roughly 45 percent of career success in AI, up from 25 percent five years ago. In an essay published November 11 on his blog, Wei admits the split is arbitrary, a rough number he assigned himself. He still insists the direction is real. Technical skill makes or breaks a career, he writes, but it no longer separates people the way it used to.
Wei gives three causes. The first is that coding agents have flattened research execution. People once pulled ahead by debugging late into the night or checking on experiments more often than rivals did. Now a well-aimed prompt gets many research tasks done, whoever types it. Taste still counts, Wei says, but when execution stops being a differentiator, other advantages gain weight.
The second cause is the price of entry. Wei recalls that a decade back, a university group running just eight GPUs could still produce excellent research. Today a single reinforcement learning run can call for hundreds of GPUs, sometimes thousands, along with data and infrastructure. Two researchers of equal ability can land in very different places depending on what they can reach. Getting that hardware turns partly on reputation, standing with the people who allocate compute, and professional ties.
The third cause is secrecy. Wei points out that once labs stopped publishing and shifted to closed-source models, outsiders lost the ability to see who contributed what. In chess or soccer, performance is visible to everyone. In AI, a model launch rarely tells you which person deserves the credit. Hiring managers, recruiters, and investors fall back on a candidate’s name, recommendations from others, and whoever is willing to vouch.
That last point has a catch Wei does not dwell on. A reference from a trusted contact is cheap to give and hard to audit. If outsiders cannot check contributions, the system rewards whoever has the most credible vouchers, whether or not those vouchers know the work well. For a founder raising money or a lab filling a senior seat, the signal is only as good as the person behind it.
Wei’s chosen example is Sam Altman, whom he describes as someone who mixes work and personal life. According to Wei, when OpenAI was smaller, Altman hosted nearly the whole staff at his home for evening socials, and he still runs poker nights and dinners to recruit. Wei credits Altman’s generosity with time and friends for a disproportionate return, and says much of OpenAI’s rise traces back to those connections. That is Wei’s judgment; the essay offers no data behind it.
Wei does not recommend networking as a tactic. He suspects that trying hard to meet people is probably the weaker strategy, and that Altman built his circle out of real curiosity about people and technology. His advice is to show up as an authentic version of yourself, which he concedes is easier said than done. He does argue that the old rule of walling off work from personal life now costs more than it once did, because sociability pays better than it used to.
Wei does not read the shift as a loss. He compares it to Python, whose abstractions made coding less technical and let more people write software. Newcomers who never learned to program can now build strong technology, he says, with product taste or research taste doing the heavy lifting. His closing line is that what people can do is commoditized, while what they choose to do and who they do it with matter more than ever.
The practical upshot lands on both sides of the table. Candidates who treat their network as a side project are competing against people who treat it as part of the job, and the gap widens as agents erase the execution edge. Anyone hiring should write down, before the interview loop starts, what they can verify about a candidate’s work without relying on a referral.
Published by Jason Wei on his blog (jasonwei.net) on November 11, 2026.
