@philhchen - Career Advice
@philhchen on X Phil Chen, a founder who has worked at Scale AI, OpenAI and DeepMind, argues that AI is good at well-defined, gradable problems, so the valuable skills now are picking the right problems, building reputation, and finishing work well. He backs this with his own career choices and with how his startup now interviews candidates.
- Time, relationships and reputation are the scarce resources; capital is not. He took a lower-paid job at Scale for the network, which later led to DeepMind and OpenAI opportunities.
- Problem selection matters more than problem solving. His company's interviews test how fast candidates understand an environment, spot worthwhile problems, and allocate tokens and time, not LeetCode-style coding.
- Polish is the differentiator, since agents reliably produce a median result. Spending extra effort on the last 10 percent, and restarting from scratch as models improve, separates strong candidates.
The author, Phil Chen, argues that AI models improve at anything you can write a loss function for, and school is mostly that: well-defined problems with known answers. So the work that lasts is whatever can't be graded inside model training. His career advice is built around that idea, and it draws on his path through a startup, Helm AI (15 to 50 people), Scale AI (500 to 1,500), OpenAI (1,500 to 3,000), and Google DeepMind. He says the scarce resources now are time, relationships, and reputation, not capital, which is cheap. He turned down quant offers with much higher guaranteed cash to join Scale, and says the network from that choice did far more for him than the money would have. Proven excellence on real work, and having that work seen by people who do good work themselves, is the strongest signal. He warns against chasing quick wins from vibe-coding, since the bigger payoffs come from searching for real value.
On hiring, he says his company is fully agent-native, so no one writes code by hand, and classic Leetcode and system design interviews stopped predicting performance. His team now tests how fast someone can understand an unfamiliar environment, spot a problem worth solving, and execute within existing constraints. The skills that matter are problem selection and allocating tokens and time. Strong candidates differ widely in how much compute and time they burn to reach a solution, and the best ones bring intuition and outside context to their work with agents. He also argues for going after the most ambitious version of a problem, since AI has made simple software easy to build and only extreme focus produces durable value. When evaluating a company, ask whether it's chasing the most ambitious form of its problem and whether it has a real shot at solving it. Then ask whether your role lets you work at the frontier. Finally, he says the last 10 percent of polish, iteration, and architecture is where most of the reward sits, because the median agent output is sloppy.
His framing for career decisions is borrowed from soccer: expected goals (xG) measures the quality of your opportunities, and efficiency measures how often you convert them. He turned down offers from Anthropic and Cursor twice, first for DeepMind in 2023 and then for OpenAI in 2024, choosing fit and interest over the higher-xG options. He admits he'd have liked more time gathering information before deciding. He doesn't think ASI will replace humans in knowledge work, because people still differ in choosing which problems matter and where to put capital. For early-stage companies, he says the team and the market matter most, since the initial product usually changes a lot if the team is good.