Phil Chen on Career Advice in the Age of AI

⬅️ Back to Articles

Phil Chen posted career advice in 2026 that cut through the usual “learn to prompt” noise. Chen has worked across Scale AI, DeepMind, OpenAI, and Google while building his own agent-native startup. His opening line stuck: AI gets good at anything you can write a loss function for, and school is mostly loss functions. Execution is cheaper. The edge is choosing problems worth the tokens, building relationships that outlast any single job, and putting in time where reputation actually compounds.

  1. Capital is abundant; time and trust are not. Chen turned down higher guaranteed cash from quant finance to join Scale at 500 people. The network and exposure there led to DeepMind and OpenAI. His point is blunt: vibe-coding side gigs can turn a quick buck, but proven work known to reputable people is still the highest-signal currency.

  2. Problem selection beats problem solving. At his agent-native company, Leetcode and textbook system design stopped predicting performance. What worked in interviews: drop someone into an unfamiliar environment, see if they spot problems worth solving, then watch how efficiently they allocate time and tokens with agents. Students panicking that agents ace problem sets are looking at the wrong scoreboard.

  3. Great candidates still differ wildly on token efficiency. Agents can finish the assignment. They cannot replace the intuition and outside context strong people bring into the collaboration. Chen rates people higher when they have lived inside high-growth environments or passion projects where meaningful problems outnumbered headcount.

  4. Pick the most ambitious form of the problem. The bitter lesson applies to careers, not just models: general scaling beats local hacks, and AI has widened the power-law gap. Anyone can spin up a simple app now. Durable value needs extreme focus on hard problems at companies that are actually trying to solve the ambitious version, not a demo that will pivot in six months.

  5. The last mile is where humans still win. Chen cites Alfred Lin: the final 10% is 90% of the work and 90% of the reward. A sloppy prompt gets you median agent output. Polish, architecture, and iteration on your own projects are what separate candidates who look interchangeable on paper.

  6. Careers need both xG and finishing. Chen uses the soccer expected-goals frame: get into positions where good opportunities appear (reputation, expertise, the right mutuals), then convert when they arrive. He passed on early Anthropic and Cursor offers twice to chase frontier inference work that fit his interests. Reasonable trade. He also wishes he had gathered more data before some calls.

  7. Research is more accessible than the gatekeeping suggests. Public eval leaderboards, model credits from providers like Modal, and building your own benchmarks from daily model use are viable on-ramps. Chen’s line: being a researcher is a mentality (curiosity, infrastructure pain tolerance, articulating results to get more compute), not a job title locked behind a lab badge.

  8. You might read this and think it only applies to people already orbiting frontier labs. Fair pushback. Chen’s network is unusually dense. But the underlying move generalizes: stop optimizing for gradable tasks agents already ace, and start stacking reputation with people who solve real problems in public.

The takeaway: Pick one problem you find meaningful, not one that grades cleanly. Ship it with agents if you want, but spend the extra hour on polish and write up what you learned where serious people can see it. DM one person whose work you respect with a specific question, not a generic coffee chat. That is how xG turns into goals when execution is cheap.

Related TMFNK Content

Crepi il lupo! 🐺