Deep-ML's Deep-Path: 103 concepts from classical ML to GPU kernels, mapped onto this course (with a placement check)
One ordered route through every problem on Deep-ML, framed as a dive from the sunlit surface to the ocean trench. This module explains it, maps all 14 parts onto the course, and gives you a 25-question self-check to find where to start.
Start here (no maths needed)
A giant practice map for AI skills, drawn like a dive into the ocean. Easy topics are near the surface, and harder ones are deeper down.
Think of it like this
Imagine a video game with levels from the beach down to the deepest trench. Before you start, a quick test checks how far down you can already go, so you don't waste time on levels that are too easy. Then each day the game gives you a few new challenges at your level. It also brings back old ones just before you'd forget them.
How it works
- Take a short test (about 25 questions) to find your level.
- Each day, do a few steps: a bit of maths, some coding problems, sometimes a project.
- The site brings back topics you're about to forget, so they stick.
- Go deeper as you improve, from basic machine learning down to GPU programming.
Why robots care
The map includes reinforcement learning (how robots learn by trial and error), a section on training robots in simulation before the real world, and robot arm maths. That robotics part is still small, though, so this course fills the gap.
Try it
Do the 25-question self-check further down. It tells you which "depth zone" to start in.
Words to know
- Placement test
- A short quiz that finds your level.
- Frontier
- The edge of what you know: just hard enough to learn from.
- Spaced repetition
- Reviewing things right before you'd forget them.
- Concept
- One topic on the map, like "Attention" or "Decision Trees".
Going deeper: the full route, the crosswalk and the self-check.
Track: LLM + Robotics. Deep-Path covers the whole ML stack, with LLMs and inference most heavily. It includes RL with a sim-to-real unit and a small robotics-and-control toolbox. It's the same kind of resource as TensorTonic (Module 02).
1. Summary
TL;DR
Deep-Path puts every problem on Deep-ML into one order: 103 concepts in 14 parts. Each concept teaches the maths it needs, then gives problems, and sometimes a project and a lab. There are two tracks:
- Everything: 103 concepts, 2,126 steps. By the site's totals, that's 329 maths items, 1,659 problems, 87 projects and 51 labs.
- Essentials: 69 concepts, 456 steps.
A placement test of about 25 short questions finds your frontier. After that, the daily "Today" view gives a few steps at that frontier and brings back concepts you're about to forget.
- Deep-ML itself describes itself as a free, open-source ML challenge platform with 1,200+ problems, labs, projects and maths. There's also a paid Premium tier.
- The depth metaphor: Sunlit (classical ML, deep learning) → Twilight (LLMs, diffusion and world models) → Midnight (RL, GNNs) → Abyss (inference engineering, CUDA/Triton, ML compilers, safety) → Trench (toolboxes: SQL, algorithms, robotics and control, maths drills).
- It's deepest where the job market is hot: LLMs (20 concepts) and inference engineering (10 concepts) get a lot of space. Robotics and control has a single concept.
- Some concepts are still filling in ("no project or lab yet"), and a couple of projects are marked as built but held. It's a living route.
2. The route: five zones, 14 parts
| Zone | Parts | Concepts |
|---|---|---|
| Sunlit · 0–200 m | 1 Classical ML · 2 Deep Learning | 21 + 16 |
| Twilight · 200–1,000 m | 3 Large Language Models · 4 Diffusion & World Models | 20 + 2 |
| Midnight · 1,000–4,000 m | 5 Reinforcement Learning (Sutton & Barto) · 6 Graph Neural Networks | 10 + 3 |
| Abyss · 4,000–6,000 m | 7 Inference Engineering · 8 GPU Programming · 9 ML Compilers · 10 AI Safety & Governance | 10 + 4 + 3 + 1 |
| Trench · 6,000 m+ | 11 SQL & pandas · 12 Algorithms & Data Structures · 13 Robotics & Control · 14 Math Drills | 4 + 5 + 1 + 3 |
3. Crosswalk: Deep-Path ↔ this course
Use this table both ways. Read the course page to understand a concept, and do the Deep-Path problems to drill it. Open a part to see every concept.
| Deep-Path part | Course pages that teach it |
|---|---|
| 1 Classical ML | 01 (thresholds, calibration), 02 maths map, 06 Days 4–5 (MDL) |
| 2 Deep Learning | 06 · 30 papers (RNN → Transformer), 08 Stage 01, 12 · AdamW, Muon |
| 3 Large Language Models | 05, 07, 08 (all stages), 10, 11, 12 · LoRA |
| 4 Diffusion & World Models | 02 Stage D (DDPM → diffusion policy), 06 Day 23 |
| 5 Reinforcement Learning | 02 RL ladder, 06 Day 29, Day 30, 12 · GRPO, 03 |
| 6 Graph Neural Networks | 06 Day 21, Day 19, Day 20 |
| 7 Inference Engineering | 09 (all 25 problems), 13, 12 · serving, 04 |
| 8 GPU Programming | 09 · kernels, 08 Stage 01, 02 (CUDA/Triton tracks) |
| 9 ML Compilers | Not covered yet. The nearest is 09 · fusion & CUDA graphs. |
| 10 AI Safety & Governance | 07 reward hacking, 08 Stage 10, 10 safety bounds |
| 11–12 SQL, pandas, algorithms | Not covered. These are interview toolbox skills. |
| 13 Robotics & Control | 02 transforms & 2-link arm, 03 shared autonomy. Kalman filtering and control are course gaps (see below). |
| 14 Math Drills | 02 maths for robots |
Every concept, by part
Part 1 · Classical Machine Learning
- 001 Vectors and Matrices
- 002 Linear Regression
- 003 Data, Splits and Baselines
- 004 Regression Metrics and Diagnosis
- 005 Regularization and Generalization
- 006 Linear Regression as Statistics (inference)
- 007 Hypothesis Testing and Experiments
- 008 Bayesian Inference
- 009 Survival Analysis
- 010 Logistic Regression
- 011 Classification Metrics: thresholds, imbalance and calibration
- 012 Generative Classifiers and Nearest Neighbours
- 013 Decision Trees
- 014 Ensembles and Boosting
- 015 Support Vector Machines and Kernels
- 016 Beyond Linearity: Polynomials, Splines, Local Regression and GAMs
- 017 Clustering and Mixture Models
- 018 PCA and Dimensionality Reduction
- 019 Anomaly Detection
- 020 Text, Retrieval and Ranking (classical)
- 021 Time Series and Financial ML
Part 2 · Deep Learning
- 022 Tensor Puzzles: broadcasting fluency
- 023 Neurons, Activations and Backprop
- 024 Training an MLP: initialization, parameters and convergence
- 025 PyTorch: tensors, autograd and modules
- 026 tinygrad and JAX
- 027 Optimizers and Learning-Rate Schedules
- 028 Normalization Layers
- 029 Regularization and Losses for Deep Nets
- 030 Training Mechanics: precision, checkpointing, data loading
- 031 Convolutional Networks and Computer Vision
- 032 Recurrent Networks and Sequences
- 033 Attention
- 034 The Transformer Block: positions, feed-forward, residuals
- 035 Autoencoders, GANs and Contrastive Learning
- 036 Distillation and Pruning Basics
- 037 Softmax in finite precision: collapse, cancellation and grokking
Part 3 · Large Language Models
- 038 Tokenization
- 039 Token embeddings and next-token data
- 040 First language models: bigram, MLP and perplexity
- 041 Build a GPT
- 042 Decoding and sampling
- 043 Positions and long context
- 044 Efficient attention and hybrid architectures
- 045 State space models: Mamba and the selective scan
- 046 Mixture of experts
- 047 Pretraining at scale: data, scaling laws, distributed training
- 048 Supervised fine-tuning and PEFT
- 049 Preference alignment
- 050 RL for LLMs: PPO, GRPO and beyond
- 051 Evaluating LLMs
- 052 Embedding models, retrieval and RAG
- 053 Interpretability
- 054 Continual learning
- 055 Research architectures: looped, latent and recurrent-depth transformers
- 056 Lattice deduction transformers
- 057 Vision-language and multimodal models
Part 4 · Generative Modeling: Diffusion and World Models
- 058 Diffusion: from DDPM to flow matching
- 059 Video, world models and latent actions
Part 5 · Reinforcement Learning (Sutton and Barto)
- 060 Bandits
- 061 MDPs and dynamic programming
- 062 Monte Carlo and temporal-difference learning
- 063 n-step Methods and Eligibility Traces
- 064 Function Approximation: semi-gradient TD, tile coding, the deadly triad
- 065 Planning and Dyna
- 066 Policy Gradients and Actor-Critic
- 067 Deep RL: DQN, replay, options and beyond
- 068 Games, search and self-play
- 069 RL systems, environments and sim-to-real
Part 6 · Graph Neural Networks
- 070 Graphs as matrices
- 071 Message passing and its limits
- 072 Node embeddings, link prediction and spectral filters
Part 7 · Inference Engineering
- 073 Latency, throughput and the roofline
- 074 KV cache, paging, batching and scheduling
- 075 Speculative decoding
- 076 Quantization
- 077 Decode kernels and fusion
- 078 Parallelism and disaggregation
- 079 Serving operations and MLOps
- 080 Speech and multimodal serving
- 081 Serving at DeepSeek scale: load balancing, the KV store and disaggregation
- 082 Verifiable inference
Part 8 · GPU Programming: CUDA and Triton
- 083 CUDA basics
- 084 CUDA intermediate: memory, warps, scans
- 085 CUDA pro: attention kernels
- 086 Triton
Part 9 · ML Compilers
- 087 Graph IR, rewrites and symbolic index math
- 088 Views, movement ops and the scheduler
- 089 Codegen: loops, tiling and GPU mapping
Part 10 · AI Safety and Governance
- 090 Proxies, privacy, fairness and gates
Part 11 · Toolbox: SQL and pandas
- 091 SQL basics
- 092 SQL intermediate: joins, windows, product metrics
- 093 SQL hard: retention and streaks
- 094 pandas
Part 12 · Toolbox: Algorithms and Data Structures
- 095 Arrays, strings and two pointers
- 096 Hashing, heaps, stacks and queues
- 097 Linked lists, trees and tries
- 098 Graphs, grids and dynamic programming
- 099 Systems-style design drills
Part 13 · Toolbox: Robotics and Control
- 100 Kinematics, rotations and filtering
Part 14 · Toolbox: Math Drills
- 101 Linear algebra drills
- 102 Calculus drills
- 103 Probability and statistics drills
4. Self-placement check
This is not Deep-ML's test; that one needs an account. It's 25 questions, five per zone, written for this course. Answer honestly and skip guessing. It shows a score per zone, and suggests where to start on Deep-Path and which course modules to read first.
Interactive: find your frontier
5. Why placement + spaced repetition works
- Placement skips material you already know and starts you at your "frontier": hard enough to learn, not so hard you stall. That's scaffolding again (Module 03).
- Spaced repetition brings a concept back just before you'd forget it. Combined with retrieval practice (actually solving, not re-reading), it builds durable memory. These are the "desirable difficulties" from Module 03.
- Maths, then problems, then a project mirrors how this course pairs explanation, a demo and a hands-on build.
6. Deep-Path vs TensorTonic vs 30u30
| Deep-Path (Deep-ML) | TensorTonic | 30u30 | |
|---|---|---|---|
| Shape | One ordered route, 103 concepts, with placement and daily review | Problem bank, study plans, maths lessons, paper re-implementations | 30 papers, one per day |
| Strongest on | Breadth: classical ML through inference, CUDA, compilers, SQL | From-scratch ML/DL coding, CUDA/Triton, maths visualisations | Foundational papers with full implementations |
| Robotics | RL incl. sim-to-real; 1 robotics & control concept | 3D transforms, CV, RL ladder | Indirect (PPO, RLHF, CNNs) |
| Best use | Daily drilling at your frontier | Targeted practice by topic | Deep understanding of the classics |
7. Why it matters for robotics
- Part 5's last unit, "RL systems, environments and sim-to-real", is the bridge to training robot policies in simulation.
- Part 13's single concept (kinematics, rotations and filtering; its first problem is forward kinematics of a 2-link arm) shows the gap. Robotics needs far more: rotations and quaternions, Kalman and particle filters, control (PID, LQR, MPC), planning. These are the same gaps this course logged back in Module 02.
- Part 4's "Video, world models and latent actions" connects to learning robot world models from video.
- Parts 7–8 (inference, CUDA) are what put models onto robot hardware within a latency budget (Module 04).
8. Self-check
What's the difference between Deep-Path's Everything and Essentials tracks?
Everything is all 103 concepts (2,126 steps). Essentials is a 69-concept, 456-step subset.
What does the placement test do, and what happens after it?
About 25 short questions find your frontier. The daily view then gives a few steps there and brings back concepts you're about to forget.
Is the "trench" the hardest material?
No. The trench holds toolbox parts (SQL, algorithms, robotics and control, maths drills). They're supporting skills placed at the bottom of the map.
Which Deep-Path part has no matching module in this course yet?
Part 9 (ML compilers), and the SQL/algorithms toolboxes. Robotics and control is only partly covered.