Eight 60-minute, one-on-one classes — personalized to your topic and your level. Each class teaches a stage, reviews your work, and plans the next; you run the experiment between classes with my async feedback.
Not a thesis — one focused question, one clean experiment, written up like a real researcher's short paper.
$500 for the full path · Class 1 is also available standalone for $40.
We pick your track — LLMs, voice, video, or RL — and set up your repo and environment so a baseline is ready to train. The lowest-risk way to start, and it's the standalone $40 Discovery Class.
Choose “LLM efficiency”, clone a research kit, and get a ~43M-param baseline training in ~15 min on a single GPU.
We read the handful of papers and PRs that actually matter for your track, pick the techniques worth testing, and run your first baseline so you have the number to beat.
Skim the relevant speedrun leaderboard entries and lock a 3-seed baseline val loss as your reference point.
We commit to ONE change versus the baseline — the core of your paper — and go deep enough that you understand it and can defend it.
Swap ReLU² → LeakyReLU(0.5)², loop the last layers N times, or multiply the QK product by a gain.
We design the minimal ablation that isolates your change — the controls, the seeds, and the single metric that decides the answer — so the result is trustworthy.
Sweep gain ∈ {1, 2, 5, 10}, 3 seeds each, identical tokens and config — nothing else moves.
We launch the runs, hunt the silent failure modes that quietly void experiments, and confirm the numbers you collect are actually real.
Verify your flag is really live, re-run a seed, and build the full results table across conditions.
We interpret signal versus noise and what the ablation proves. If it's interesting, we design one follow-up that explains WHY it works or fails. Honest null results count.
Add a channel-level vs head-level control to show where a measured gain actually comes from.
We write it the way researchers actually do — results first, then method, abstract, and charts — turning your experiment into a clear, short paper.
A 4–8 page write-up: question, method, a 3-seed results table, and an honest conclusion.
Final edits, a clean GitHub repo and README, and shipping it publicly. You leave able to run the next one on your own.
Push the repo, post the result thread, and you have real proof you can do research.
Classes are scheduled 1-on-1 with you directly — no group cohort, no fixed weekly pace. Everything runs on a single modest GPU; all starter code and charts are provided.
Founding price for the first members. Start with a single $40 class or claim a full spot.