Deep AI Engineering in 500 Hours | Generated by AI
Question: How should I allocate my ~4 hours/day over the next 3 months to maximize AI engineering skill compounding?
Answer:
Your situation: bank contractor by day, AI engineer by trajectory. 4 hours/day is ~500 hours over 3 months. That’s enough to go deep on one thing or shallow on five. The mistake most people make is spreading it thin. Here’s how I’d structure it for you specifically.
The Core Constraint
You’re not skill-zero. You’ve trained GPT-2 760M, built CLI agents, consumed 2B tokens. The gap between you and Karpathy isn’t breadth — it’s depth on internals and shipped artifacts that prove it. Those 500 hours should close that gap.
Allocation Framework (per week ~28 hrs)
Block 1 — Deep Technical (12 hrs/week, ~2 hrs/day on weekdays)
Goal: One core AI engineering domain per month, studied to implementation depth.
| Month | Focus | Concrete Deliverable |
|---|---|---|
| 1 | Transformer internals → attention, KV cache, sampling | Implement MHA + KV cache from scratch in ~200 lines of Python. Beat your nanoGPT notes. |
| 2 | Training loop mastery → LoRA, DPO, GRPO | Fine-tune a 1B model on your RTX 4070 with LoRA. Log every loss curve. |
| 3 | Inference optimization → quantization, speculative decoding, batching | Run DeepSeek-V4 locally or via API with custom inference harness. Benchmark throughput. |
This is Karpathy-style: one thing, all the way down, working code.
Block 2 — Ship Something (8 hrs/week, weekend-heavy)
Goal: One public artifact per month. Not blog posts — tools with GitHub stars or paying users.
Candidates ranked by leverage for you:
- An opinionated LLM eval harness (CLI-first, wraps OpenRouter) — scratches your own itch, useful to others
- A LoRA fine-tuning CLI (wrappers around transformers/unsloth, one command) — high star potential
- A nanochat-derived chatbot trained on your own blog/notes — living proof of your skill stack
Ship → get feedback → iterate. This is your startup validation runway.
Block 3 — Reading/Synthesis (4 hrs/week, ~30 min/day)
Not passive consumption. Active reading = paper → annotate → implement one key idea.
Prioritized reading list for next 3 months:
- Flash Attention 2 paper → implement naive vs. flash attention, benchmark
- DeepSeek-V4 MoE paper (technical report) → understand expert routing
- Chinchilla laws → understand why you’d train 760M vs 7B given compute budget
Skim your own notes first. If it’s already at lzwjava.github.io/notes-en, go one layer deeper.
Block 4 — Eliminate (0 hrs)
Based on your own playbook: WeChat noise, non-compounding networking, passive video watching, anything that doesn’t feed Block 1 or 2. Track for one week, then cut.
Daily Time Architecture
6:00–8:00 Deep technical (Block 1) — before brain is taxed by work
Evening
19:00–21:00 Ship/build (Block 2 on weekends, light reading Block 3 on weekdays)
Morning is non-negotiable. Bank work drains context. You want transformer internals in your working memory before the daily standup, not after.
The 3-Month Output Test
At the end of 90 days, you should be able to say:
- “I implemented attention + KV cache from scratch and can explain every line”
- “I fine-tuned a model with LoRA and DPO and have loss curves to show it”
- “I shipped a tool that has >100 GitHub stars or earns any CNY”
If you can’t say all three, the allocation drifted. Recalibrate at the 30-day mark.
One Tactical Suggestion
Use your bank job’s AI assistant quota (top 6% globally) as a forcing function. Every task at the bank that touches data, automation, or tooling — extract a pattern, abstract it, add it to your tooling. Don’t waste the access.
The mortgage clock is running. These 500 hours are the compression mechanism.