Stop Overthinking, Just Leave | Generated by AI
You’re overthinking the exit and underrating yourself, so let me drop the hedging.
Leave. Start actively looking now. Sixteen months is plenty. You’ve extracted everything the bank has to teach you — the notes-per-month novelty curve has flattened, you said it yourself. Staying longer isn’t risk management, it’s decay. Every month you spend writing Java import/validation/export features is a month you’re not compounding the only skill that matters for where you’re going. The bank is now pure tax, almost zero learning. That’s the signal to move.
Drop the “only give me hard tasks” plan entirely — not because it’s too aggressive, but because it’s the wrong aggression. It’s trying to win a game you’ve already decided to quit. You don’t negotiate scope at a job you’re leaving. You do the minimum to keep the contract clean and the reference intact, and you pour every spare watt into the exit. Coast deliberately and silently. Don’t announce it, don’t argue scope, don’t pick fights with the team. Just redirect the energy.
Here’s the bold version of the truth: you don’t actually have an inference/training profile yet, and no amount of boldness changes that — but you can build one faster than you think, because you already have MI300X access, ROCm familiarity, and you’ve trained real models. Most people chasing these roles have read papers. You’ve melted GPUs. The gap is narrow and closeable in weeks, not years.
So the bold play isn’t “tell the team off.” It’s:
-
Pick your exit target now — inference optimization or training infra. Don’t straddle. Inference is the faster on-ramp (more roles, clearer benchmarks, your ROCm/FlashAttention work transfers directly).
-
Ship one undeniable artifact in 6-8 weeks. vLLM or SGLang serving on the MI300X, benchmarked throughput/latency wins, written up with real numbers and the kernel/batching reasoning. Public. That single thing outweighs your entire bank tenure on an inference resume.
-
Start interviewing while you build it. Don’t wait until it’s “ready.” The interviews tell you exactly which gap to close next. You learn the market by being in it.
-
Use the Principal Engineer as your reference and warm intro — the DuckDB/Python project is your one bank asset that points the right direction. Stay close to him until you leave. A principal vouching “this person ships and goes deep” opens doors no portfolio does.
The only thing keeping you cautious is the mortgage and family, and that’s legitimate — so the boldness is in the speed of the transition, not in quitting raw into nothing. Line up the next thing, then leave clean. Don’t drift out over six more months of resentment-coasting. Compress it: artifact + interviews + exit in a quarter.
You’ve consumed 2B tokens, trained GPT-2 from scratch, and rank top 6% in AI usage at a global bank. Stop asking permission to do the work you’re already better-positioned for than most people who get paid to do it. The bank was the launchpad. You’ve already cleared the tower.
Want the 8-week inference artifact spec? That’s the thing that turns this from ambition into a dated plan.