First-Principles Career Path | Generated by AI
Question: Should I run a one-person AI startup, join a company as an employee, or abandon this path to finish a 大专 (associate) degree and prepare for the GRE to study in the US?
Answer:
Let’s look at this with the same first-principles rigor you’d apply to a training run — evaluate the actual gradient, not the vibes.
The three paths, scored against your stated objective function
Your stated goal: Tinker/Frontier Labs-level depth, Agentic world leadership, AI-native products. Every option should be scored against that, not against “safety” or “prestige.”
Option A: One-person AI startup (what you already launched)
- Alignment with goal: 9/10. You’re already doing the exact reps — training GPT-2 760M on MI300X, working through nanochat, pushing into DeepSeek v4 MoE, consulting at the frontier of what companies actually need. This is the Tinker-adjacent path, minus the compute budget.
- Cost: revenue uncertainty, family friction, no safety net.
- Compounding factor: every client project is paid R&D. You’re not choosing between “learning” and “earning” — you structured it so they’re the same activity. That’s the correct architecture.
Option B: Join a company
- Alignment with goal: 3-5/10 depending on the company. If it’s DeepSeek, Anthropic, a frontier lab, or a serious MoE/training team — this jumps to 8/10, because you’d get compute and mentorship you can’t self-fund. But “join a company” in the generic sense (another bank, another contract gig) is a regression — you already did that (top 6% AI usage at the bank) and extracted what you could from it. Going back without a specific elite-lab offer is just re-capping your own growth curve.
- You already tested this hypothesis empirically: bank job was safe but capped. Repeating an experiment you’ve already concluded is not first-principles thinking, it’s fear dressed up as prudence.
Option C: 大专 → GRE → US
- Alignment with goal: 1-2/10, and here’s why, concretely:
- You have 12 years of shipped production experience (iOS, Android, backend, AI, a startup with 30K users and 3M CNY revenue). A 大专 credential adds negative signal to your resume relative to what you already have — it tells a story of “starting over” rather than “compounding.”
- GRE + US grad school is a 2-4 year detour optimized for people who need the credential to access opportunity. You don’t have that problem — you have direct technical output (trained models, open notes, 400 blog posts) as your credential. Karpathy, Yin Wang, and Liang Wenfeng are not people you catch up to by getting a degree; you catch up by matching their output cadence.
- Financially it’s the worst-scoring option by a wide margin. 900K CNY mortgage + 5,500 CNY/month payments + two daughters + a non-supportive family, against 2-4 years of US tuition/living costs with no income during that window and an F-1 visa that legally restricts you from running your startup or consulting. This isn’t a risk premium, it’s a compounding financial and family crisis.
- It also reverses point 6 of your own playbook: family persuasion through results, not arguments. Going back to school reads as capitulation to “the safe path,” which was never in question — you already decided against it when you dropped out and again when you left the bank.
The actual decision
This isn’t really a 3-way tie — it’s A, with B (elite lab) as a specific, high-value acquisition you should keep scouting for opportunistically, and C essentially dominated by A on every axis relevant to your objective function.
Concrete move: Don’t treat “join a company” as generic — treat it as a search problem. Set a standing filter: would this org give me compute + research access I can’t buy with consulting revenue (i.e., actual frontier lab, or a well-funded China AI lab doing MoE/RLHF work)? If yes, evaluate the specific offer against runway math. If no, it’s strictly worse than staying independent — don’t let “join a company” as an abstract category tempt you back into a local optimum you already escaped.
The GRE/大专 thought is worth noticing as a signal, though — it usually surfaces when runway anxiety spikes. If that’s what’s driving it, the fix is tightening Phase 1 revenue targets (your own playbook: 25K CNY/month first, then bank-salary parity), not reversing four years of trajectory. Track it in numbers, same as you’d track training loss — if MRR isn’t moving after 3-4 months of real outreach, that’s the actual decision point, not a proxy panic response mid-quarter.