TL;DR
· An AI entrepreneur claims that AI agents writing code are reshaping the ranking of early-career capabilities.
· Scoreable tasks are more suited to models; humans need to learn to identify problems, allocate time, and use tools.
· Cash returns are not the only goal; relationships, reputation, and delivery quality will widen the gap.

An entrepreneur who claims to have worked at companies like Scale AI, DeepMind, OpenAI, and Google, and is now involved in an agent-native startup, has written a long English article offering new career advice to young people. The backdrop is that AI coding tools have evolved from code completion to more complete software engineering agents. When OpenAI released Codex in 2025, it stated that the tool could handle tasks like writing features, fixing bugs, and submitting PRs in parallel in the cloud, but still required human review and verification of code. The question then becomes: when standard answers, ordinary code, and scoreable tasks become increasingly cheap, where should young people spend their time?

The core of the article is not "programmers will be replaced," but that the criteria for early-career screening are changing. Schools and traditional interviews heavily train on clearly defined, answer-specific, gradable problems—exactly where models are improving fastest. In the future, what may distinguish people more is identifying important problems, choosing high-value environments, building credible reputations, and refining the mediocre outputs generated by agents into deliverables.

Cash Offers Are No Longer the Only Answer; Time and Reputation Are More Scarce

In the author's judgment, in the AI startup environment, capital and tools are easier to obtain than before, but high-quality time, strong relationships, and credible reputations remain scarce.

He explains this using personal experience. Before joining Scale AI, he says he received a quant position offer with higher cash guarantees, but ultimately chose Scale because it had a stronger community, a broader product landscape, and more opportunities to engage with cutting-edge problems. According to his recollection, it was through Scale that he later connected with large model inference providers, gained opportunities at DeepMind and OpenAI, and met colleagues who later founded startups with him.

These experiences cannot simply be extrapolated into a universal career formula, but the reminder is direct: early career choices should not focus only on immediate cash. Especially after AI lowers the barrier to building software, quickly making a profitable little tool is no longer rare. Long-term returns often come from harder problems, stronger groups of people, and more credible signals on your resume.

What young people need to ask is not "Which opportunity gives me more money right now?" but whether this matter is worth investing time in, whether they can work with excellent people, whether their good work can be seen by reliable people, and whether it will serve as the credit foundation for the next opportunity.

Engineer Value Shifts from "Solving Problems" to "Finding Problems"

When agents can handle an increasing number of clearly bounded problems, the engineer's value is no longer just "can they solve it?" but "can they choose the right problem?"

The author mentions that their team redesigned their interview process. The reason: if real work no longer requires writing every line of code by hand, then purely testing algorithmic problems and traditional system design will have lower correlation with job performance. A more meaningful test is to see whether a candidate can quickly understand the environment, identify problems worth solving, and then leverage AI tools and external resources to drive results.

This is also the new division of labor after agents write code. Models excel at tasks with clear goals and clear feedback; humans need to decide which problems are important, which paths are worth trying, and how much time and model invocation costs to invest.

For students, the fact that AI can do homework may bring frustration. But from a recruitment perspective, the differences between candidates haven't disappeared. Even if everyone can get answers using AI, some need extensive trial and error and prompt engineering, while others can collaborate with agents using business intuition, technical background, and context to find a direction faster.

"Being good at using AI" is not just about feeding problems to the model. Stronger capabilities include breaking down problems, identifying missing information, judging when to keep iterating and when to change course, and verifying whether the results truly resolve key business or technical contradictions.

The Easier Software Becomes, the Closer You Should Get to Harder Problems

AI lowers the barrier to building software and also makes simple systems easier to replicate. The author borrows the "bitter lesson" from machine learning research to explain career choices: in the long run, scaling general methods often outperforms fine-tuning for a single task.

Applied to companies and personal careers, this means the moat around simple outputs becomes thinner. Anyone can more easily build a seemingly usable system; truly lasting value instead concentrates on problems that are sufficiently difficult and ambitious.

When choosing a company, the author's criteria are: Is this company solving the most ambitious version of the problem? Does it truly have a chance to solve it? When choosing a role, consider whether this position allows you direct exposure to the cutting-edge problem the company is solving.

He also mentions not to look only at whether the early product looks polished or the demo is impressive. By his subjective assessment, Anthropic's early demo at the time seemed like just a Slackbot that wasn't as good as ChatGPT, but that didn't stop the company from eventually taking a completely different path. Early companies change; products change. Team quality, market space, and problem difficulty have a bigger impact on long-term outcomes.

Career opportunities follow a similar logic. High-quality opportunities don't always translate into results, but you need to first position yourself where you can see them. Whether you can get there still depends on long-accumulated skills, reputation, and whether others are willing to tell you about the opportunity.

Common Results Become Cheaper; the Last 10% Becomes More Valuable

When a simple prompt can have an agent generate a medium-quality result, the value of ordinary outputs falls, and the value of final refinement rises.

The article quotes Sequoia Capital's Alfred Lin, who says that the last 10% is often 90% of the work and 90% of the return. In the AI era, this statement feels more real. Because a 70-score result is increasingly easy to obtain, what truly distinguishes people is unique perspective, attention to detail, iteration ability, architecture quality, scalability, and creativity.

The first version of an AI output is rarely directly perfect. The real work often happens in subsequent iterations: discovering what's wrong, what needs refactoring, where the user experience is rough, which edge cases are uncovered, and when to leverage the next-generation model to redo everything from scratch.

These abilities can be practiced through projects, internships, and real work. Spending a little more time refining, making the architecture clean, thinking through scalability, and polishing details until users genuinely want to use it—all leave traces in portfolios and interviews.

Traditional engineering skills haven't become obsolete. The change is that the scarcity of writing code itself has decreased, while judgment, aesthetics, system understanding, and delivery quality have become more expensive. AI allows more people to reach an average level, but the remaining gap becomes even harder to bridge.

The Barrier to Research Has Lowered, but Research Is Not a Title

The article concludes by extending the discussion to "how to get into research." The author believes that AI hasn't made research exclusive to top labs; instead, it has lowered the entry barrier.

Modern research certainly depends more on computing power, but the starting point can be simple: use existing models, turn your intuition into evaluations, participate in public optimization leaderboards, leverage cloud computing credits available to students and researchers, and test ideas as early as possible. Most ideas will ultimately fail when scaled, but understanding failure is part of building research judgment.

A researcher is first and foremost a way of working, not just a job title. Research in frontier labs often mixes curiosity, trying new ideas, working with infrastructure, understanding system details, quick debugging, and articulating the value of results to secure more resources. Much of this training doesn't need to wait until you get the "researcher" title.

The career advice this article leaves is not pessimistic. AI makes standard answers, ordinary code, and scoreable tasks cheaper, and also lets young people get exposed to real problems earlier. Opportunities still exist, but their distribution has changed: those who can find important problems, enter high-quality environments, build credible reputations, and push results to the last mile will be more likely to get the next opportunity.

Click to learn about open positions at BlockBeats

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia