AI models get better at anything you can write a loss function for, and school is mostly loss functions: well-defined problems graded against known answers. Therefore, the valuable work of the next decade is everything that can't be graded within the span of model training.
During my 6 years of work, I've been fortunate to have collaborated with amazing people from companies of all sizes, from my own startup, to Helm AI (15→50 FTE), Scale AI (500→1500 FTE), OpenAI (1500→3000 FTE), and Google (100,000+ FTE). As a founder, I spend a lot of time thinking about the right hires for our company's present and future. Because we are fully agent-native, our needs are very different from any company where I've previously worked.
For motivated, ambitious, early-career individuals, I now have a clearer perspective on what skills are valuable in the coming decade. I've given and received a lot of career advice, and while many famous adages remain true (something something rocketship don't ask which seat), much has changed due to the rise of agentic coding. Here's what has remained true, as well as what is new.
1. Focus on resources that are truly limited
Before joining Scale, I had quant offers with much higher guaranteed cash, but I decided to join Scale because I was excited by the community and exposure to all of Scale's various products and applications. Through Scale, I gained exposure to LLM inference providers, leading to my DeepMind and OpenAI opportunities. I also met many other ambitious colleagues who now form a community of founders from Scale. Today, the unique network and learning opportunities from Scale have contributed far more to my life than the extra cash I would have received from quant.
Access to capital is far easier now than ever before. Access to real time and strong relationships with other humans is still rare. Proven excellence in past, related endeavors remains of the highest signal, so my concrete advice is to spend time doing good work and ensuring it's known to other reputable people who themselves do good work. Relentlessly prioritize your time so that whatever you work on, whether it be school, projects, or internships, you focus on problems that you find meaningful. With vibe-coding, it's easy to find opportunities that turn a quick buck, but the prize is usually much larger when you search for real value.
Time, relationships, and reputation: these are the true limited resources in which to focus attention.
2. Learn to find problems in addition to solving them
To find signal in a sea of candidates, we thought deeply about what skills matter today for engineers working in an agent-native company. Given that no one writes any lines of code manually, traditional Leetcode-style questions and even system design questions feel uncorrelated with actual job performance. Eventually, we arrived at a series of interviews that measure how well someone can quickly understand the environment they're placed into, identify problems worth solving, and then execute on solving those problems under the constraint of the existing environment.
The most important skills will be the ones related to problem selection and resource allocation. Ever-powerful agents are able to take on complex, well-defined problems, so the most impactful people will be the ones best at identifying important problems and then allocating tokens and time to solving them.
I see a trend of students feeling discouraged by the fact that agents can solve all their problem sets. But in my experience giving interviews, candidates still have widely varying performances in terms of how much time and tokens they need to arrive at the solution. Great candidates usually bring high-level intuition and outside context to their collaboration with agents.
Concretely, candidates we have rated highly have immersed themselves into problem-solving environments, either from their own passion projects or from being in high-growth companies where meaningful problems outnumber the people.
3. Work on the most ambitious form of a problem
For the past decade, one of the most useful mental frameworks in research has been the “bitter lesson”: scaling general methods ultimately outperforms task-specific optimizations. This lesson applies to choosing problems and companies as well.
Companies and careers have always had power-law outcomes, but AI has accelerated the rate of progress towards these outcomes. Because building software is now much more accessible, anyone can build simple systems with relative ease. Real, durable value only gets built with extreme focus on truly ambitious problems.
To choose a company, the advice here is simple: evaluate whether the company is working on the most ambitious form of their problem, and then whether they actually have a shot at solving it. To choose a role, think about whether the role will allow you to work directly on the frontier of whatever problem the company is solving.
4. Sprint the last mile
For startups, Alfred Lin has a great article about how the last 10% is both 90% of the work and 90% of the reward. AI has polarized outcomes because the median result is what an agent can produce with a sloppy prompt. Value therefore comes from providing a unique perspective on a slice of problems or attention to detail.
Learning to execute well in the last mile requires both practice and focus. Nothing is perfect on the first try, so the last mile is often about iteration. Because progress with coding agents has been so rapid, it is often better to take learnings from prior iterations and just start from scratch with the next generation of intelligence. Practice this with your own projects. Take the initiative to spend just a little more time on polish, clean architecture, scalability, or creativity. I’ve definitely seen the impact across candidates for those who have done this.
5. Increase both xG and efficiency
In soccer, xG (expected goals) is a metric for how many goals a team is expected to score in a match based on their opportunities, accounting for distance, angles, goalkeeper position, etc. Efficiency is the relative conversion rates on these opportunities.
The xG and efficiency analogy to my own career has been fairly accurate. In 2023, I turned down offers from Anthropic (~50 FTE at the time) and Cursor (2 non-founder FTEs at the time) because I wanted to work on frontier model inference and training at DeepMind. In 2024, I turned both down again to work at OpenAI. Each of these alternative opportunities would have been high xG from a career perspective, but I ended up choosing companies that aligned more with my interests, culture fit, and goals (pun intended).
Careers are long and opportunities come and go. I don’t believe that ASI will replace all humans in knowledge work jobs because humans have differential capabilities in selecting meaningful problems for ASI to solve, and in allocating capital to solve these problems.
Not every opportunity will materialize into a goal, but being in the right position to see the opportunities is the first step to scoring goals. This again comes down to reputation and expertise. The Cursor opportunity came because I had a good reputation among my mutuals with Michael and Aman, and the Anthropic opportunity came because I had been investing both professional and personal time into problems that were interesting to the team there.
At some point, life is about scoring goals, not just seeing the opportunities, so efficiency in front of goal also matters. Looking back on my decisions, I think I made many of the right decisions but would have preferred to have spent more time gathering data to inform my decisions.
At its core, selecting early-stage companies is primarily about the team and the market. Many candidates today anchor onto the existing product, but that almost always evolves into something very different if the team is good. Anthropic’s initial demo was a Slackbot that was worse than ChatGPT for me.
6. You can break into research now
Recently, I’ve gotten a lot of questions from people on how to break into research. My former colleague Vlad is a lead on the Gemini team and has an excellent writeup on his perspectives here.
Modern research is easier to do with more compute, but a great place to start is using the models and distilling your own intuitions into evaluations. Public optimization leaderboards publicized by my former colleague @kellerjordan0 also provide great forums for exploring ideas in a more structured setting.
Many compute providers like Modal provide credits for academics. Use them and explore your ideas now. Most ideas will eventually fail at scale and understanding these failures is the first step to building an understanding of what actually works.
Ultimately, I believe that being a researcher is a mentality, not an occupation. Most work of a researcher in frontier labs is a mix of being curious enough to explore new ideas, fighting against infrastructure to implement the ideas, understanding the full system in extreme detail to debug issues efficiently, and articulating the value of the results to secure more compute. You can do all of this without being in a frontier lab.
Closing thoughts
The world is still full of opportunity. The key to unlocking them is to focus on finding interesting problems and delivering extraordinary results. If this appeals to you, reach out and we’d love to work with you.
