When the Interviewer Is Also an AI Newbie, Hiring Becomes Two Newbies Nodding at Each Other
In an a16z interview, Tyler Cowen said AI makes hiring harder in the short term: old signals stop working, new credentials have not arrived, and everyone is a beginner during the transition. The replies offer a more direct approach: give a task, watch the process.
a16z released a clip from an interview with Tyler Cowen. He said AI makes hiring harder in the short term. The gist is straightforward: it is now hard to tell whether someone you hire is good at AI. The abilities companies used to look for had imperfect signals—like having attended Cornell, getting straight A's, or interviewing well. What needs to be judged now is psychological flexibility—whether someone can learn AI—and hiring teams often do not know how. Good matches are harder, and searches take longer. In the long run, AI itself will solve this: credentials will emerge that assess how strong someone is with AI. During the transition, everyone is a beginner.
Cowen is an economist at George Mason University and has long written Marginal Revolution. The clip's focus is on hiring evaluation: old proxy metrics have no predictive power for 'can this person adapt to AI.' He is talking about the short term. Credentials will emerge in the long run, but in the short term no one has a standard answer.
In the same a16z conversation, Erik Torenberg, Alex Imas, and Tyler Cowen discussed a bigger question: could rapid gains in AI capability translate into runaway economic growth? Both economists lean skeptical. The reasons listed in the conversation include: smarter models cannot automatically bypass economic bottlenecks; slow institutions limit AI's economic impact; competition among AI labs makes strong models cheaper; the benefits consumers get from AI may be far larger than what GDP captures; scientific progress and longer lifespans may be among AI's biggest gains. The full video is on YouTube.
Full video: https://www.youtube.com/watch?v=g_67TmQ8lGY
The replies split the topic into two categories: how to interview people, and how to measure the economy.
On the hiring side, one commenter said a 10-minute screen share is enough: watch what a candidate does after their first answer is wrong. Another suggested giving the candidate a real task, a generative AI tool, and 30 minutes, then reading what they ask. The questions someone asks reveal more than a transcript. Another said résumé signals are becoming useless. Electra's sarcasm is spot-on: if everyone is a beginner, an interview is two beginners nodding at each other. Another question is stranger still: if the person being interviewed is an agent, how do you interview something that only learned to work yesterday?
Jonathan Sandhu's long reply took issue with 'waiting for AI to issue credentials.' His point is that credentials move the measurement problem into another AI system and then declare it solved. His approach is to have the candidate do the work: give them an unfamiliar task, appropriate tools, clear constraints, and measurable acceptance criteria; record the completed output, correctness, latency, cost, error detection, recovery, independent verification, and whether they can recognize when a model is confidently wrong. Then change the requirements and see how they adapt. The approach is not new; engineering organizations have always evaluated systems this way. It's just that now we need to evaluate a person's ability to operate systems.
Video attached to Jonathan Sandhu's reply:
On the economy side, Hassan Nasser's reminder is concrete: smarter models cannot build you a new machine tool overnight; capital equipment does not compile. Anthony Sun said model quality can jump by a lot and still fail to touch the slow parts—regulation, physical factories, and people reaching agreement. Smarter models will not automatically fix coordination. Tobias Sherman put it bluntly: GDP is not a good metric; growth has no purpose in itself. The consumer-surplus camp argues that the benefits AI brings will exceed the GDP numbers. Another long reply from Jonathan Sandhu takes aim at a16z: on one hand it invests in data centers because frontier users say there is not enough compute, and on the other it uses weak macroeconomic productivity evidence to question an AI economic takeoff—that is using two different evidentiary standards. His suggestion is to measure the full workflow: whether collaborating agents can reliably execute state, authorization, verification, recovery, and accountability; what the total cost, latency, throughput, and failure rate are; and whether surrounding institutions can deploy the results.
Taken together, these discussions point to the same change behind hiring difficulty and slow economic impact: the unit of evaluation has changed. Résumés, grades, and interview performance are personal signals; model benchmarks are capability signals; real workflows require reliable results under constraints. In the short term, the more practical move is to bring work samples back into interviews and observe the process, errors, and adaptation. Waiting for an AI certificate can be put off. Certificates will come, but models may change faster than certificates. During the transition, everyone is a beginner; admitting that is more useful than pretending there is a standard answer. Interviewers and candidates are taking the same exam, except the interviewer also has to pretend to be writing the questions.
发布时间: 2026-11-17 11:32