Status: direction, not a plan
Horizon: ~6 months, with weekly check-ins against this note

What this is #

A reference document capturing a long conversation about how to spend the next ~6 months of personal study time. It is a living note, not a fixed plan. It will be revised as the work reveals what actually matters.

Starting identity #

I am becoming someone who understands modern LLMs and transformers across abstraction levels — from the math to the metal. The working title is research engineer with inference engineer capabilities, but the exact craft will emerge from the work. The trunk is transformers and inference. Branches may form later.

Why now #

• I miss the depth I had during my bachelor/master years.
• My recent production work has been broad and integration-heavy.
• I have flexible personal time at MVF and want to use it deliberately.
• I want to build public evidence of depth, not just consume content.

Priority order #

1. Depth — real understanding, from-scratch implementation, derivation.
2. Feeling of understanding — the satisfaction of owning a concept.
3. Artifacts — public code, notes, blog posts, experiments.
4. Career positioning — becoming credible for the roles I want.

This order protects against chasing hot topics or status.

Structure: core + satellites + horizon #

Layer 1 — Core trunk (gets most time) #

These are the starting focus. They are not separate; they connect.

• Transformer internals: attention, MLPs, normalization, positional encoding, training dynamics.
• LLM inference: autoregressive generation, KV-cache, sampling, quantization, batching, serving trade-offs.
• Efficient kernels: memory-aware algorithms, Triton, profiling.

Layer 2 — Satellites (bounded) #

One active at a time. Max ~20% of weekly study hours. Must produce a small artifact. Time-boxed. No implementation unless it graduates to a core slice.

• Agents — studied through the transformer/inference lens; connects to MVF work.
• SOTA paper reading — one paper at a time, one-page note, connects to current core.
• TypeScript literacy — enough to read Pi and similar agent-harness code; no porting projects.
• Public writing / experiments — distribution and documentation of core work.

Layer 3 — Horizon (not now) #

Interesting, but deferred. Write them down when they appear; do not start them.

• Robotics / embodied AI
• RL / world models as a primary pursuit
• Diffusion / flow models
• CUDA C++ / Tile-Lang / Rust deep dives
• Porting Pi to Python
• Standalone network engineering study

Traversal method: thin vertical slices #

Instead of broad courses, take one concept and move it down through abstraction levels:

1. Math / derivation
2. Algorithm / complexity
3. Code from scratch
4. Systems / profiling
5. Metal literacy (enough to understand compiler output and systems code)

A slice is thin if it has one core concept, one layer at each level, and one artifact. It explicitly does not include full models, production deployment, SOTA extensions, or multiple domains.

Example first slice: attention mechanism from scratch.

Time and execution #

Weekly schedule (added to Google Calendar for 8 weeks, 2026-06-29 to 2026-08-23) #

Total protected time: ~12–18 hours per week. Realistic focused work: ~10–14 hours per week.

How to use each block #

Mon/Tue/Fri 3–7pm “Core Study”:

• 3:00–3:15pm: transition, close work, write today’s goal
• 3:15–5:00pm: deep work (derive, implement, debug)
• 5:00–5:30pm: break
• 5:30–7:00pm: lighter work (reading, notes, Tau satellite)

Sat/Sun 10am–3pm “Deep Study”:

• 10:00–10:15am: review where you left off, set goal
• 10:15–12:30pm: deep work
• 12:30–1:00pm: break
• 1:00–3:00pm: deep work or writing

Execution principles #

• Never-zero rule: on the worst week, do at least one small thing related to the core trunk.
• Review cadence: check this note monthly. Adjust based on what is actually happening.
• Rest is part of the plan: Wed/Thu are intentionally light. One full day off per week.
• Public output as accountability: commit to one artifact per month to keep momentum visible.

Public trajectory #

The goal is to become publicly recognizable for a specific kind of work, by doing permissionless, rigorous work in the open. Inspirations: Will Brown, Jino Rohit, Antirez — not to copy their topics, but to copy the pattern of public depth.

Distribution mix (situation-based, not mandatory):

• GitHub — code, reproductions, experiments.
• Blog — milestone writeups, derivations, lessons learned.
• X / LinkedIn — small updates, observations, distribution of longer work.

The public output is an output of depth, not the driver of it.

What was explicitly rejected #

• A 6-month rigid curriculum.
• Daily LeetCode as a primary activity. LeetCode may be a maintenance habit later.
• Porting Pi to Python as a personal project.
• Co-authoring an academic paper as the main goal. Replaced with permissionless public experiments + writeups.
• Chasing every interesting domain. The three-layer model decides what is active and what waits.

MVF relationship #

MVF is the funding source and an application ground. Current work is building agents with the Anthropic SDK. I want to move toward more provider-agnostic, cost-aware, possibly self-hosted inference. That connects naturally to the inference-study trunk. TypeScript literacy is allowed as a small satellite because Pi is well-designed and provider-agnostic, but no porting project.

Open questions to revisit #

• Which part of the transformer/inference stack lights me up most as I go deeper?
• Do I want to stay at MVF long-term, or is this a stepping stone?
• What does my first public artifact look like, and when?
• How much compute do I realistically have, and how does that constrain experiments?

Last updated #

2026-06-29