AKP Anthropic Kernel Path

0/21 read · no cycles yet

12-hour day operating system

You asked for 12-hour days. That is survivable only if it is not 12 hours of papers. Anthropic performance work is a craft. Craft is built by making things, then studying why they were slow.

6 days/week × 12 hours = 72 hours/week. Take the 7th day almost off (walk, light Anki, no new material). If you skip rest you will fail the take-home by being sloppy, which is the same as being slow.

Daily template (local time, pick a 8:00 start)

ClockBlockWhat
8:00–8:20BootSleep, body, one sentence: today's artifact. No Twitter, no HN, no Discord.
8:20–11:20Deep 1 (3h)Hardest making of the day. Kernel, derivation, CSAPP lab, take-home cycles. Phone in another room.
11:20–12:00Walk + retrieveNo headphones-as-podcast-dump. Recite what you just did. Write 5 Anki cards from holes.
12:00–12:40Eat
12:40–15:10Deep 2 (2.5h)Second hard thing, often the other track (if Deep 1 was metal, this is model, or reverse).
15:10–15:30BreakOutside.
15:30–17:30Paper / book (2h)One paper or one textbook chapter, with exercises. Notes are a derivation, not a summary.
17:30–18:00Eat / reset
18:00–19:30Deliberate practice (1.5h)Interview reps: timed CodeSignal-style implementation, or a whiteboard derivation, or a system-design drill out loud.
19:30–20:00WritePublic artifact: git commit + 10–20 line log in log.md. If you cannot write what you did, you did not do it.

That is ~11.5 hours of clock, ~9 hours of real cognitive work. The remaining slack is transitions. Do not "make it 12 hours of Deep" to feel serious. That is how you write buggy kernels.

Weekly rhythm

  • Mon–Tue: new hard material + new code
  • Wed: profile / debug day. No new papers. Make Tuesday's thing correct and measured.
  • Thu: paper-heavy, but still one coding lab
  • Fri: take-home sport (Anthropic sim) or a from-scratch reimplementation
  • Sat: long build (4–6h on one artifact) + review
  • Sun: off-ish. Anki 30 min. Walk. Sleep 8+.

Monthly checkpoint (non-negotiable)

On the last Saturday, you must have one public artifact with:

  • repo link
  • what you built
  • numbers (cycles, TFLOPS, tokens/s, loss curve)
  • what was wrong the first time
  • a 500–1500 word writeup

If the month has no artifact, the month did not happen. Studying is not an artifact.

Spaced repetition

Anki decks you will build yourself (do not download a "transformer deck" and feel prepared):

  1. Numerics and IEEE + GPU arch facts you keep missing
  2. Formula → complexity → IO cost cards (attention, collectives, KV cache)
  3. Paper claims: one card per number you actually need (Chinchilla tokens/param, FlashAttention IO reduction)
  4. Anthropic-specific: CAI procedure steps, RSP levels, take-home machine constraints

Review Anki only on the walk and Sunday. Do not let Anki eat Deep 1.

Burnout rules (you will ignore these once; do not twice)

  • Sleep < 7h: cut Deep 2, do not cut sleep
  • Three days of "I read but wrote nothing": that is a failed week, restart with a smaller artifact
  • Pain in wrists / back: you need this body for 12–24 months. Fix the chair.
  • If you start optimizing your note system, you are avoiding the take-home.

Compute

You will need, in order:

  1. Laptop + CPU is enough for months 1–2 (from-scratch GPT on Shakespeare/TinyStories, CSAPP, take-home sim is CPU)
  2. One consumer GPU (4090 / 5090 / used 3090) or a cloud T4/L4 for Triton/CUDA
  3. Occasional A100/H100 hours (Lambda, RunPod, CoreWeave, academic credits) for FA2-competitive numbers and Nsight
  4. Do not wait for an H100 to start. The take-home is a Python simulator. Most of the skill is independent of the SKU.

Nsight Compute on a laptop GPU still teaches the questions. The absolute TFLOPS change; the habit of asking does not.