Lessons from Young AI Engineers | Generated by AI

Home 2026.06

Question: What can I learn from the successful young AI engineers and founders described in this article?

Answer:

The article paints a sharp picture of what’s happening in China’s AI talent market. Let me extract the actual signal — not the hype — and map it to your situation.

What’s Actually Happening

The article describes a structural shift: AI-nativeness has replaced tenure as the primary signal of value. A 22-year-old with deep model intuition is outperforming a 35-year-old with 10 years of engineering experience. The mechanism is clear — the old experience compounds slowly; AI knowledge compounds within months.

Key data points:

What “AI-Native” Actually Means

The article gives two definitions worth dissecting:

“思维模式和大模型的输入输出完全对齐，遇事先问 AI，且知道下一题该问什么”

Translation: your mental model is the model’s I/O. You don’t just use Claude — you know what prompt structure extracts useful information, and you know what to ask next based on the output. That’s a feedback loop skill, not a tool skill.

“为什么老年人用智能手机需要学习，小孩不用？因为小孩理解点击屏幕会出现什么”

Translation: intuition about system behavior, not learned procedures. You reason from first principles about what the model will do, not what it should do per the docs.

You already have this. 1B tokens/month, building CLI agents, training GPT-2 from scratch — that’s demonstrated AI-nativeness. The question is: are you converting this into output that’s legible to the market?

Concrete Lessons Mapped to Your Situation

1. Output velocity > credential accumulation

The young people in this article don’t have PhDs. They have papers, repos, and shipped systems. The article explicitly says “reading PhD is a waste of time” in this context — what matters is demonstrable output that advances the frontier.

Your blog (400 posts) and notes (9,600 AI answers) are already a signal. But are they frontier output? The delta between “documenting what I learned” and “publishing something nobody else has” is what separates a good engineer from someone who gets cold-emailed by Seed’s HR team.

Action: Pick one narrow technical problem — KV cache optimization, a specific MoE routing behavior, something in nanoGPT — and publish something with an actual finding. Not a tutorial. A result.

2. Network as a compounding asset

The article describes how top talent flows through intern pipelines, and the gate is a warm referral from a professor or senior researcher. The SF Bay Area version: email someone whose paper you read, with a specific observation about their work.

You’ve trained on MI300X, you’ve consumed more tokens than most research teams. That’s a genuine conversation starter with serious researchers. You’re not a student asking for advice — you’re a peer with unusual hands-on experience.

Action: Find 3 papers you’ve actually run or implemented (even partially). Email the authors with a specific technical observation. Not “I loved your paper” — something like “I noticed X behavior when I ran Y on MI300X, does that match your intuition about Z?”

3. The gap is in model internals, not applications

The article’s most important implicit signal: the highest-value people are in base model training (pre-training, post-training) — not application engineering. The “脏活” (dirty work) that interns avoid is often where the real learning is.

You’re already doing this — training GPT-2 760M, studying nanoGPT. The next step is going deeper into the training dynamics that actually matter:

# What most people understand: loss goes down
# What top researchers understand: why loss goes down the way it does
# - grad norm spikes at certain token positions
# - attention entropy collapse in deep layers
# - MoE router load imbalance under distribution shift

Action: When you run training experiments, instrument them like a researcher — log grad norms per layer, attention entropy, loss by token position. Publish the findings.

4. Startup optionality > salary optimization

The article says explicitly: these young people don’t care about salary because they’re planning to start companies. The salary is a floor, not a ceiling. The real upside is equity in something that matters.

You’re a contract AI engineer at a global bank. That’s stable but the ceiling is fixed. The article is essentially describing a window — right now, a 2-person team with strong model intuition can compete with well-funded labs in narrow domains.

What do you have that a fresh grad doesn’t? 12 years of shipping real products, banking domain knowledge (financial systems), cross-stack experience (iOS to GPU), and Mandarin + English fluency across Chinese and Western AI ecosystems.

That combination is genuinely rare. A narrow AI system applied to financial data, built by someone who understands both model internals and real production constraints, is a fundable idea.

5. The “AI-native” filter is public infrastructure now

The article mentions VC funds explicitly targeting post-1998 founders. But the actual filter is technical output, not age. Cao Xi (曹曦) at Sequoia is looking at 2000–2002 founders because they tend to be more AI-native — but age is a proxy, not the criterion.

You’re 30-something (based on context) but you’re consuming 1B tokens/month and training models from scratch. That puts you in the top 0.1% of engineers by AI-nativeness regardless of birth year. The question is whether your visible output reflects that.

Honest Gap Analysis

Where the article’s top performers are that you’re not yet:

Dimension Article’s Top Performers You Now
Research output Top-venue papers Blog posts, notes
Network Warm-referred into Seed/Kimi circles Mostly solo
Equity exposure Startups / large option grants Contract salary
Specialization depth One narrow frontier area Broad applied AI

The gap is closeable. But it requires picking a narrower technical bet and going deeper than anyone else you know — not staying broad.

The Line That Should Stick

“这个时代对非凡的奖励从未如此丰厚，对平凡的惩罚从未如此严厉。”

“The rewards for the extraordinary have never been greater; the punishment for the ordinary has never been harsher.”

You’re already not ordinary. The question is whether you’re building toward extraordinary in a way that’s legible to the market. That means: narrower bets, faster publication, warmer networks, and eventually — a company.

Back Donate