Escape the Loss Function: 6 Career Rules for the AI Era from an OpenAI Vet
A former OpenAI and Scale AI engineer shares six career principles for the AI era, arguing that valuable work lies beyond what models can optimize via loss functions, and emphasizing problem discovery, ambitious scope, execution excellence, and strategic opportunity selection over mere technical problem-solving.
1. Focus on Truly Scarce Resources
The author chose Scale AI over a higher-paying quant role because of the community and exposure to LLM inference providers, which later led to opportunities at DeepMind and OpenAI. The Scale alumni network became a founder community. The author argues that capital is now abundant, but time, deep relationships, and proven excellence remain scarce. Advice: invest time in producing excellent work visible to other high-performing, trustworthy people; be ruthless with time allocation toward meaningful problems.
2. Learn to Discover Problems, Not Just Solve Them
In an agent-native company where code is written by agents, traditional LeetCode and system design interviews are less relevant. The author's company designed interviews to measure how quickly a candidate understands context, identifies worthwhile problems, and executes within constraints. The most important skill is problem selection and resource allocation. Strong candidates bring high-level intuition and external context when collaborating with agents, often gained from passion projects or fast-growing companies where problems outnumber people.
3. Do the Most Ambitious Version of the Problem
Applying the "bitter lesson" (scaling general methods beats task-specific engineering) to career and company choice. Outcomes follow a power law, and AI accelerates the path. Since building simple systems is now easy, lasting value comes from extreme focus on truly ambitious problems. When choosing a company, evaluate whether it tackles the most ambitious version of its problem and has a real shot at solving it. When choosing a role, ask if it puts you at the frontier of that problem.
4. Sprint the Last Mile
Citing Alfred Lin's observation that the last 10% of work takes 90% of effort and yields 90% of returns. AI makes outcomes more bimodal: the median output is what an agent produces from a sloppy prompt. Value comes from unique perspective or obsessive detail. Mastering the last mile requires practice and iteration; with rapidly improving coding agents, it's often better to absorb lessons from early iterations and restart from scratch with the next-generation agent. The author actively looks for candidates who polish quality, clean architecture, improve scalability, or add creativity.
5. Increase Both xG and Efficiency
Borrowing soccer's expected goals (xG) metric: xG measures the quality of chances created; efficiency is the conversion rate. The author declined offers from Anthropic (50 employees) and Cursor (2 non-founders) in 2023 to join DeepMind for frontier model inference/training, and again in 2024 chose OpenAI over those same companies, prioritizing interest, culture fit, and goals. Career is long; humans retain differentiated ability in selecting meaningful problems for ASI and allocating capital. Reputation and expertise generate high-xG opportunities (Cursor offer came from reputation in Michael and Aman's network; Anthropic offer from investing time in problems their team cared about). Eventually, conversion efficiency matters too; the author wishes they had gathered more data to support decisions. Early-stage company evaluation should focus on team and market, not current product—Anthropic's first demo was a Slackbot worse than ChatGPT.
6. You Can Enter Research Now
Modern research is easier with more compute, but a good start is using models and distilling your intuition into evals. Public optimization leaderboards (e.g., by @kellerjordan0) provide structured exploration. Compute providers like Modal offer academic credits. Most ideas fail at scale; understanding those failures builds intuition for what works. Research is a mindset, not a job title: stay curious, wrestle infrastructure, understand the whole system for efficient debugging, and articulate value to secure more compute—all doable outside frontier labs.
The world remains full of opportunities; the key is focusing on discovering interesting problems and delivering extraordinary results.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
