TL;DR
· An AI entrepreneur stated that AI coding agents are changing early career skill prioritization.
· Model-suitable tasks are better handled by models, while humans need to learn problem judgment, time allocation, and tooling.
· Cash reward is not the sole objective; relationships, reputation, and delivery quality will make a difference.

An entrepreneur involved in a native AI company, who claims to have worked at companies such as Scale AI, DeepMind, OpenAI, and Google, has reshaped career advice for young people in a lengthy English article. The backdrop is that AI coding tools have evolved from code completion to more comprehensive software engineering AI agents. When OpenAI released Codex in 2025, they stated that it could handle tasks like writing functions, fixing bugs, and creating PRs in the cloud in parallel, but still required human code review and validation. The question then becomes: as standard answers, mundane code, and model-suitable tasks become increasingly inexpensive, where should young people invest their time?

The core of this article is not "programmers will be replaced," but that early career screening criteria are changing. Schools and traditional interviews train heavily on well-defined, clearly answerable, and gradable questions, which happen to be where models excel the most. What may differentiate individuals in the future is the ability to identify important problems, choose high-value environments, build a trusted reputation, and refine AI-generated intermediate results into deliverables.

Cash Offer No Longer the Sole Answer; Time and Reputation Are Scarcer

In the author's opinion, in the AI startup environment, capital and tools are more accessible than ever before, but high-quality time, strong relationships, and a trustworthy reputation remain scarce.

He explains this point through personal experience. Before joining Scale AI, he claims to have received a cash-rich quant role offer, but ultimately chose Scale because of its stronger community, broader product scenarios, and more exposure to cutting-edge problems. According to his recollection, it was through Scale that he was exposed to large-scale model inference providers, gained opportunities at DeepMind and OpenAI, and also met a group of colleagues who later started ventures.

These experiences cannot be simply extrapolated into everyone's career formula, but the reminder given is straightforward: early career choices should not be solely focused on immediate cash. Especially after AI has lowered the software development barrier, rapidly creating a profitable small tool is no longer rare; long-term rewards often come from tackling harder problems, being part of a strong network, and possessing a more credible resume signal.

Young individuals need to ask not "which opportunity offers more money upfront," but whether the endeavor is worth investing time in, if they can work with excellent people, if their good work can be seen by reliable individuals, and if it will serve as the foundation for the next opportunity.

Engineers' Value Shifts from "Problem Solving" to "Problem Finding"

As AI systems become more capable of handling well-defined problems, the value of engineers is no longer just about "whether they can solve it," but rather "whether they can choose the right problem to solve."

The author mentioned that their team has redesigned the interview process. The reason is that if real-world work no longer requires individuals to handwrite every line of code, then simply testing algorithmic problems and traditional system design would be less relevant to on-the-job performance. A more meaningful test is to see if a candidate can quickly understand the environment, identify worthwhile problems to solve, and then leverage AI tools and external resources to drive outcomes.

This is also a new division of labor after AI writes the code. The model excels at handling tasks with clear objectives and feedback, while humans need to discern which problems are important, which paths are worth exploring, how much time and model invocation cost should be invested.

For students, the ability of AI to do homework may bring frustration. However, from a recruitment perspective, the differences among candidates have not disappeared. Even if everyone can obtain answers using AI, some individuals may need a lot of trial and error and hints, while others can leverage business intuition, technical background, and context to collaborate with AI more efficiently and find directions more quickly.

Being "proficient in AI" does not just mean handing problems over to a model. Stronger capabilities include breaking down problems, identifying missing information, determining when to continue iterating, when to pivot, and verifying whether the results truly resolve critical business or technical contradictions.

The Easier It Is to Build Software, the Closer One Should Get to Harder Problems

AI has lowered the threshold for software development and made it easier to replicate simple systems. The author uses the "bitter lesson" from machine learning research to explain career choices: in the long run, expanding general methods often surpasses fine-tuning for a single task.

Applied to companies and individual careers, this means that the moat around simple outputs will become thinner. Anyone can more easily create a seemingly usable system, while enduring value actually lies in tackling sufficiently challenging and ambitious problems.

When choosing a company, the standard the author presents is: Is this company working on the most ambitious version of the problem, and does it genuinely have a chance of solving it? When selecting a role, it is essential to consider whether this position allows one to directly engage with the cutting-edge problems the company is addressing.

He also mentioned that one should not just focus on whether the early product looks good or if the demo is impressive. By his subjective assessment, Anthropic's early demo at the time appeared to be nothing more than a Slackbot inferior to ChatGPT, but this did not hinder the company from later taking a completely different path. Early-stage companies will change, products will change, and factors like team quality, market space, and problem complexity are more likely to influence long-term outcomes.

Career opportunities follow a similar logic. High-quality opportunities do not always translate into outcomes, but one must first position oneself to see the opportunity. Whether one can stand there still depends on accumulated skills, reputation, and whether others are willing to inform you of the opportunity.

Common results are cheaper; the final 10% is more valuable

When a simple hint can enable an AI to generate medium-quality results, the value of common output decreases, while the value of the final polish increases.

The original text quotes Sequoia Capital's Alfred Lin as saying that the final 10% is often 90% of the work and also 90% of the return. In the AI era, this statement holds even more true. As 70-point results become easier to obtain, what truly sets individuals apart is a unique perspective, attention to detail, iterative ability, architectural quality, scalability, and creativity.

The initial output of AI is rarely perfect. The real work often occurs in subsequent iterations: identifying what is incorrect, which areas need refactoring, where the experience is not yet smooth, which edge cases are not covered, and when to start over with the next-generation model.

These abilities can be honed through projects, internships, and real-world work. Spending a little more time on polishing, keeping the architecture clean, planning scalability clearly, and ensuring the details are something users truly want will leave a mark in your work and interviews.

Traditional engineering skills have not become obsolete. The change lies in the diminishing scarcity of coding itself; judgment, aesthetics, system understanding, and delivery quality have become more valuable. AI enables more people to reach a moderate level, making it harder to bridge the remaining gap.

The research barrier has lowered, but research is not a title

The article concludes by extending the discussion to "how to enter research." The author believes that AI has not confined research to top-tier labs; instead, it has lowered the early entry barrier.

Modern research undoubtedly relies more on computing power, but the starting point can be humble: use existing models, translate one's intuition into evaluations, participate in public leaderboards, utilize cloud computing credits offered to students and researchers, and test ideas early. Most ideas will ultimately fail at scale-up, but understanding failure is part of forming research judgments.

Researchers are primarily a way of working, not just a position. Research in cutting-edge labs often involves a blend of curiosity, trying new ideas, infrastructure debugging, understanding system details, rapid debugging, and articulating the value of results to secure more resources. Much training does not have to wait until one receives the title of "researcher" to begin.

The career advice left by this article is not pessimistic. AI has made standard answers, ordinary code, and easily gradable tasks cheaper, while also exposing young people to real-world problems at an earlier stage. Opportunities still exist, but the way they are distributed has changed: those who can identify important problems, enter high-quality environments, build a trusted reputation, and drive results to the last mile are more likely to receive the next round of opportunities.

Welcome to join the official BlockBeats community:

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

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

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