Skip to content
ZeroToAI
Get started

Curriculum

A skill tree, not a ladder

A massive common trunk, then routes that split apart, cross and converge higher up. A branch's height is its prerequisite level: no link ever goes back down.

Tiers
14
Branches
228
Sub-branches
680
Convergences
42
Mandatory projects
15
Comet projects
4

  1. 01

    Common trunk: tiers 0 to 5

    Mandatory for everyone: code, maths, learning theory, classical ML, neural networks, architectures.

  2. 02

    A graph, not a list

    From N6 on, routes split. Some stay independent, others merge again: CNN → Vision → VLM → VLA.

  3. 03

    Height is prerequisite

    A branch never depends on one placed higher. To climb, you must have unlocked everything below.

  4. 04

    Mandatory milestones

    Every three tiers, very demanding projects: 1, then 2, then 4, then 8. Black holes, nebulae, pulsars: no climbing higher without validating them.

Scroll to climb · Drag to turn · Pinch or Ctrl + wheel to zoom · Right-click to move

N0

Computing and Python

280 XP◆ 0/15

The full curriculum, tier by tier

The same map, as a list. Every branch can open in the tree.

N0Computing and PythonCommon trunk · 15 branches · 58 sub-branches
  • 70 XP

    Computing

    • Syntax
    • Functions
    • Data structures
    • Comprehensions
  • 70 XP

    Computing

    • Commits
    • Branches
    • Merge / Rebase
    • Pull requests
  • 70 XP

    Computing

    • Bash
    • Processes
    • SSH
    • Permissions
  • 70 XP

    Computing

    • SELECT / JOIN
    • Aggregations
    • Window functions
    • Indexes
  • 70 XP

    Computing

    • Classes
    • Inheritance
    • Dataclasses
    • Protocols

    Prerequisites: Python basics

  • 70 XP

    Computing

    • Iterators
    • Generators
    • yield
    • itertools

    Prerequisites: Python basics

  • 60 XP

    Computing

    • Type hints
    • Generics
    • mypy / pyright

    Prerequisites: Python basics

  • 70 XP

    Computing

    • ndarray
    • Broadcasting
    • Vectorization
    • Indexing

    Prerequisites: Python basics

  • 70 XP

    Computing

    • pytest
    • Fixtures
    • Mocks
    • Continuous integration

    Prerequisites: Python basics, Git

  • 70 XP

    Computing

    • DataFrame
    • GroupBy
    • Joins
    • Lazy API

    Prerequisites: NumPy, SQL

  • 70 XP

    Computing

    • optimize
    • stats
    • sparse
    • signal

    Prerequisites: NumPy

  • 60 XP

    Computing

    • Matplotlib
    • Seaborn
    • Plotly

    Prerequisites: NumPy

  • 70 XP

    Computing

    • cProfile
    • line_profiler
    • Memory
    • Benchmarks

    Prerequisites: Tests

  • 70 XP

    Computing

    • Threads
    • Multiprocessing
    • asyncio
    • Dask / Ray

    Prerequisites: Iterators & generators, Linux / Shell

  • 70 XP

    Computing

    • CUDA
    • Kernels
    • GPU memory
    • Triton (advanced)

    Prerequisites: Parallelism, NumPy

N1Mathematics of AICommon trunk · 12 branches · 41 sub-branches
  • 70 XP

    Mathematics

    • Vectors / matrices / tensors
    • Vector spaces
    • Norms / distances

    Prerequisites: NumPy

  • 70 XP

    Mathematics

    • Derivatives
    • Partial derivatives
    • Chain rule

    Prerequisites: NumPy

  • 80 XP

    Mathematics

    • Random variables
    • Distributions
    • Expectation / variance
    • Conditional probability

    Prerequisites: NumPy

  • 70 XP

    Mathematics

    • Eigenvalues / eigenvectors
    • Diagonalization
    • Symmetric matrices

    Prerequisites: Linear algebra

  • 70 XP

    Mathematics

    • Gradient
    • Jacobian
    • Hessian

    Prerequisites: Calculus, Linear algebra

  • 70 XP

    Mathematics

    • Bayes' theorem
    • Prior / posterior
    • Likelihood

    Prerequisites: Probability

  • 80 XP

    Mathematics

    • Entropy
    • Cross-entropy
    • KL divergence
    • Mutual information

    Prerequisites: Probability

  • 80 XP

    Mathematics

    • Estimation
    • Hypothesis tests
    • Bootstrap
    • Inference

    Prerequisites: Probability, Pandas / Polars

  • 80 XP

    Mathematics

    • SVD
    • QR
    • Cholesky
    • Matrix decompositions

    Prerequisites: Eigenvalues

  • 70 XP

    Mathematics

    • Convex / non-convex
    • Constraints
    • Lagrangian

    Prerequisites: Gradient, Jacobian, Hessian

  • 70 XP

    Mathematics

    • MLE
    • MAP
    • Log-likelihood

    Prerequisites: Statistics, Bayes

  • 80 XP

    Mathematics

    • Gradient descent
    • SGD
    • Stochastic optimization
    • Learning rate

    Prerequisites: Optimization

N2Foundations of learningCommon trunk · 12 branches · 36 sub-branches
  • 80 XP

    Learning theory

    • Regression vs classification
    • Loss function
    • Empirical risk

    Prerequisites: MLE / MAP, Gradient descent

  • 80 XP

    Learning theory

    • Latent structure
    • Density estimation
    • Grouping

    Prerequisites: Statistics

  • 80 XP

    Learning theory

    • Bias / Variance
    • Overfitting / Underfitting
    • Capacity

    Prerequisites: Supervised learning

  • 80 XP

    Learning theory

    • Train / Validation / Test
    • Cross-validation
    • Data leakage

    Prerequisites: Supervised learning, Statistics

  • 80 XP

    Learning theory

    • Semi-supervised
    • Weakly-supervised
    • Pseudo-labels

    Prerequisites: Supervised learning, Unsupervised learning

  • 65 XP

    Learning theory

    • Pretext tasks
    • Learned representations

    Prerequisites: Unsupervised learning

  • 65 XP

    Learning theory

    • Online learning
    • Active learning

    Prerequisites: Supervised learning

  • 90 XP

    Learning theory

    • Transfer learning
    • Multi-task learning
    • Meta-learning
    • Continual learning

    Prerequisites: Supervised learning

  • 80 XP

    Learning theory

    • L1 / L2
    • Penalties
    • Early stopping

    Prerequisites: Bias / Variance

  • 90 XP

    Learning theory

    • Accuracy / F1
    • ROC-AUC
    • RMSE / MAE
    • Ranking metrics

    Prerequisites: Validation

  • 80 XP

    Learning theory

    • Covariate shift
    • Label shift
    • Drift

    Prerequisites: Validation

  • 80 XP

    Learning theory

    • Calibration
    • Uncertainty
    • Confidence intervals

    Prerequisites: Metrics, Bayes

Milestone 1 · between N2 and N31 mandatory project · mandatory to climb higher
  • 720 XP

    Black hole · 6 weeks

    You build your own stats library in pure NumPy, no SciPy, no ML library. Then you unleash it on a real, messy dataset, booby-trapped just for you: leakage to hunt, drift to fix.

    Deliverables

    • Typed library, fixed API: GLMs (SGD, Newton), MLE/MAP, BCa bootstrap, hypothesis tests, calibration
    • CI on every push to your repo: pytest, coverage, mypy --strict, import check, benchmarks
    • One-command study: cleaning, time-based CV, final model built with your library only, predict.py
    • leaks.json in the required schema; adversarial validation against unlabeled X_test, drift corrected
    • Report (10 pages max): reliability diagrams, bootstrap CIs, paired test vs baseline, profiling

    Validation criteria

    1. 01100% NumPy library: ≥ 95% of hidden tests pass (vs SciPy, statsmodels, sklearn), coverage ≥ 90%, 0 mypy --strict errors
    2. 02Hidden test labels: ECE (15 bins) ≤ 0.03, log-loss ≤ 1.05 × the reference logistic model; ≥ 4 of 5 leaks, ≤ 1 false positive
    3. 03Fresh clone, 4-vCPU runner: make reproduce ≤ 30 min, 2 identical runs; logistic fit, 1M × 50, converged in ≤ 15 s, ≤ 2 GB

    Branches it mobilizes: NumPy, Typing, Profiling, Gradient descent, MLE / MAP, Statistics, Validation, Distribution shift, Calibration & uncertainty

N3Classical machine learningCommon trunk · 17 branches · 54 sub-branches
  • 100 XP

    Machine Learning

    • Linear
    • Polynomial
    • Quantile
    • Robust regression

    Prerequisites: Supervised learning, Linear algebra, Statistical engine in pure NumPy

  • 85 XP

    Machine Learning

    • Decision Tree
    • Gini / entropy
    • Pruning

    Prerequisites: Supervised learning, Information theory

  • 85 XP

    Machine Learning

    • Distances
    • KD-tree
    • Curse of dimensionality

    Prerequisites: Supervised learning

  • 85 XP

    Machine Learning

    • Naive Bayes
    • LDA
    • QDA

    Prerequisites: Bayes, Supervised learning, Statistical engine in pure NumPy

  • 115 XP

    Machine Learning

    • K-Means
    • Hierarchical
    • DBSCAN
    • HDBSCAN
    • Spectral

    Prerequisites: Unsupervised learning

  • 100 XP

    Machine Learning

    • PCA
    • Kernel PCA
    • ICA
    • NMF

    Prerequisites: SVD & decompositions, Unsupervised learning

  • 75 XP

    Machine Learning

    • Bayesian models
    • Gaussian Processes

    Prerequisites: MLE / MAP, Statistical engine in pure NumPy

  • 85 XP

    Machine Learning

    • Bayesian Networks
    • HMM
    • Probabilistic graphical models

    Prerequisites: Bayes

  • 85 XP

    Machine Learning

    • Ridge
    • Lasso
    • ElasticNet

    Prerequisites: Regression, Regularization

  • 75 XP

    Machine Learning

    • Logistic Regression
    • Softmax regression

    Prerequisites: Regression

  • 85 XP

    Machine Learning

    • Bagging
    • Random Forest
    • Extra Trees

    Prerequisites: Decision trees

  • 115 XP

    Machine Learning

    • AdaBoost
    • Gradient Boosting
    • XGBoost
    • LightGBM
    • CatBoost

    Prerequisites: Decision trees, Gradient descent

  • 75 XP

    Machine Learning

    • Gaussian Mixture
    • EM algorithm

    Prerequisites: Clustering, MLE / MAP

  • 75 XP

    Machine Learning

    • t-SNE
    • UMAP

    Prerequisites: Dimensionality reduction

  • 100 XP

    Machine Learning

    • Statistical
    • Isolation Forest
    • One-Class SVM
    • LOF

    Prerequisites: Unsupervised learning, Decision trees

  • 85 XP

    Machine Learning

    • Maximum margin
    • Kernels
    • SVR

    Prerequisites: Logistic regression, Optimization

  • 85 XP

    Machine Learning

    • Stacking
    • Voting
    • Blending

    Prerequisites: Bagging & forests, Boosting

N4Neural networksCommon trunk · 9 branches · 29 sub-branches
  • 95 XP

    Deep Learning

    • Neuron
    • Perceptron
    • ADALINE

    Prerequisites: Logistic regression, Gradient descent

  • 95 XP

    Deep Learning

    • MLP
    • Multilayer network
    • Universal approximation

    Prerequisites: Neuron & perceptron

  • 145 XP

    Deep Learning

    • ReLU
    • Sigmoid
    • Tanh
    • GELU
    • SiLU
    • Softmax

    Prerequisites: MLP / PMC

  • 80 XP

    Deep Learning

    • Forward propagation
    • Loss functions

    Prerequisites: MLP / PMC, Information theory

  • 80 XP

    Deep Learning

    • Computational Graph
    • Automatic differentiation

    Prerequisites: Forward & losses, Gradient, Jacobian, Hessian

  • 80 XP

    Deep Learning

    • Backpropagation
    • Vanishing / exploding gradients

    Prerequisites: Computational graph & autodiff

  • 95 XP

    Deep Learning

    • SGD / Momentum
    • Adam / AdamW
    • Schedulers

    Prerequisites: Backpropagation

  • 110 XP

    Deep Learning

    • BatchNorm
    • LayerNorm
    • RMSNorm
    • GroupNorm

    Prerequisites: Backpropagation

  • 110 XP

    Deep Learning

    • L1 / L2
    • Dropout
    • Early stopping
    • Data augmentation

    Prerequisites: Backpropagation, Regularization

N5Architecture familiesCommon trunk · 18 branches · 49 sub-branches
  • 105 XP

    Vision & 3D

    • Convolution
    • Pooling
    • Receptive field

    Prerequisites: Optimizers, Normalization

  • 90 XP

    Audio & time series

    • RNN
    • BPTT

    Prerequisites: Optimizers

  • 105 XP

    Generative

    • AE
    • Bottleneck
    • Denoising AE

    Prerequisites: Optimizers, Dimensionality reduction

  • 125 XP

    Graphs

    • Message passing
    • GCN
    • GAT
    • GraphSAGE

    Prerequisites: Optimizers, Linear algebra

  • 90 XP

    AI for science

    • Neural ODE
    • Adjoint method

    Prerequisites: Backpropagation, Calculus

  • 90 XP

    AI for science

    • Spiking neurons
    • Neuromorphic

    Prerequisites: MLP / PMC

  • 125 XP

    Vision & 3D

    • ResNet
    • DenseNet
    • U-Net
    • ConvNeXt

    Prerequisites: CNN, Neural regularization

  • 90 XP

    Audio & time series

    • LSTM
    • GRU

    Prerequisites: RNN

  • 105 XP

    Generative

    • VAE
    • ELBO
    • Reparameterization trick

    Prerequisites: Autoencoder, Information theory

  • 90 XP

    Vision & 3D

    • Capsules
    • Dynamic routing

    Prerequisites: CNN

  • 105 XP

    Language & agents

    • Encoder-decoder
    • Teacher forcing
    • Beam search

    Prerequisites: LSTM / GRU

  • 90 XP

    Audio & time series· ◎

    • LTC
    • CfC

    Prerequisites: LSTM / GRU, Neural ODE

  • 105 XP

    Language & agents

    • Attention
    • Self-attention
    • Multi-head

    Prerequisites: Seq2Seq

  • 125 XP

    Language & agents

    • Encoder
    • Decoder
    • Encoder-Decoder
    • Positional encoding

    Prerequisites: Attention, Normalization

  • 105 XP

    Vision & 3D· ◎

    • Patches
    • ViT
    • Swin

    Prerequisites: Transformer, CNN

  • 90 XP

    Audio & time series· ◎

    • S4
    • Mamba

    Prerequisites: RNN, Transformer

  • 105 XP

    Language & agents

    • Routing
    • Sparse MoE
    • Load balancing

    Prerequisites: Transformer

  • 90 XP

    Reasoning & causality

    • Memory Networks
    • Modern Hopfield Networks

    Prerequisites: Attention

Milestone 2 · between N5 and N62 mandatory projects · mandatory to climb higher
  • 1,180 XP

    Nebula · 7 weeks

    A 7-table SQL database, messy and incomplete. You predict loan default with calibrated probabilities and explained rejections, and beat our baseline on months you have never seen.

    Deliverables

    • SQL and Polars pipeline that rebuilds every feature from the raw database, with no data leakage
    • Time-based validation report: penalized logistic, forests, boosting, SVM and k-NN, all tuned
    • Final stacked model, calibrated on the latest month, batch scoring CLI, test coverage ≥ 90%
    • Drift and anomaly monitor: per-feature PSI, Isolation Forest, one JSON alert per batch
    • Model card and per-applicant reason codes, derived from the full final model, calibration included

    Validation criteria

    1. 01Hidden test set: ROC-AUC ≥ our LightGBM tuned on the main table + 0.02; ECE over 15 quantile bins ≤ 0.01
    2. 0220 hidden 5,000-row batches: the monitor flags ≥ 9 of the 10 drifted batches and ≤ 1 of the 10 clean ones
    3. 03One command reruns it all in CI, hyperparameters frozen: ≤ 2 h on 8 cores and 32 GB; 4 reason codes per rejected applicant

    Branches it mobilizes: Pandas / Polars, Calibration & uncertainty, Distribution shift, Penalized regression, SVM, k-NN, Boosting, Stacking & voting, Anomaly detection

  • 1,340 XP

    Pulsar · 8 weeks

    No PyTorch, no JAX: tensors, autograd, layers, optimizers, GPU backend, you code it all. Then you take on the tier 5 challenge with no framework, timed against PyTorch.

    Deliverables

    • Library installable with pip: broadcasting tensors, autograd, NumPy and CuPy backends
    • Conv2d, LSTM, GRU, attention, BatchNorm, LayerNorm, Dropout; SGD momentum, AdamW, schedulers
    • Test suite: finite-difference gradcheck on every op and every layer
    • Trained weights: a CIFAR-10 ResNet and two language models, an LSTM and a mini-Transformer
    • Benchmark report against the school's PyTorch references: curves, speed, memory, profiling

    Validation criteria

    1. 01The school's hidden gradchecks on every op: relative error ≤ 1e-6 in float64; test coverage ≥ 90%
    2. 02ResNet: ≥ 90% accuracy on the CIFAR-10 test set; LSTM and Transformer: loss ≤ 1.02 × the reference loss
    3. 03Retraining in CI on a T4 without cuDNN: ≤ 2 h and ≤ 4 GB of VRAM per run; ResNet step time ≤ 3 × PyTorch's

    Branches it mobilizes: GPU computing, Profiling, Computational graph & autodiff, Optimizers, Normalization, Neural regularization, Modern CNNs, LSTM / GRU, Transformer

N6Learning paradigmsCrossroads · 12 branches · 31 sub-branches
  • 115 XP

    Deep Learning

    • SimCLR
    • MoCo
    • InfoNCE

    Prerequisites: Self-supervised, Modern CNNs

  • 95 XP

    Deep Learning

    • Masked language modeling
    • MAE

    Prerequisites: Self-supervised, Transformer

  • 115 XP

    Deep Learning

    • Next-token prediction
    • JEPA
    • Predictive coding

    Prerequisites: Self-supervised, Transformer, Deep learning engine from scratch

  • 115 XP

    Deep Learning

    • Feature extraction
    • Fine-tuning
    • Domain adaptation

    Prerequisites: Transfer & multi-task, Modern CNNs

  • 95 XP

    Deep Learning

    • MAML
    • Learning to learn

    Prerequisites: Transfer & multi-task, Optimizers

  • 115 XP

    Deep Learning

    • Catastrophic forgetting
    • Replay
    • EWC

    Prerequisites: Transfer & multi-task, Optimizers, End-to-end credit scoring

  • 95 XP

    Deep Learning

    • Hard / soft sharing
    • Loss weighting

    Prerequisites: Transfer & multi-task, Optimizers

  • 95 XP

    Deep Learning

    • Example ordering
    • Self-paced learning

    Prerequisites: Optimizers

  • 95 XP

    Trust & safety

    • FedAvg
    • Decentralized data

    Prerequisites: Optimizers

  • 115 XP

    Deep Learning

    • Embeddings
    • Latent spaces
    • Disentanglement

    Prerequisites: Contrastive learning, Autoencoder

  • 115 XP

    Deep Learning

    • Siamese Networks
    • Triplet Loss
    • Contrastive Loss

    Prerequisites: Contrastive learning, k-NN, End-to-end credit scoring

  • 115 XP

    Deep Learning

    • Few-Shot Learning
    • Zero-Shot Learning
    • Prototypical networks

    Prerequisites: Metric learning, Meta-learning

N7Modeling by data typeCrossroads · 9 branches · 28 sub-branches
  • 125 XP

    Machine Learning

    • Classical ML
    • Boosting
    • Tabular DL

    Prerequisites: Boosting, MLP / PMC, End-to-end credit scoring

  • 125 XP

    Vision & 3D

    • CNN
    • ViT
    • Augmentations

    Prerequisites: Modern CNNs, Vision Transformers, Transfer learning, Deep learning engine from scratch

  • 165 XP

    Audio & time series

    • RNN
    • LSTM
    • TCN
    • Transformer
    • SSM

    Prerequisites: LSTM / GRU, State Space Models, Deep learning engine from scratch

  • 125 XP

    Language & agents

    • Tokenization
    • Embeddings
    • Transformer

    Prerequisites: Transformer, Masked prediction, Representation learning

  • 145 XP

    Audio & time series· ◎

    • DSP
    • Spectrograms
    • CNN
    • Transformer

    Prerequisites: CNN, Transformer, SciPy

  • 100 XP

    Graphs

    • GNN
    • Graph construction

    Prerequisites: GNN, Representation learning

  • 100 XP

    Vision & 3D

    • PointNet
    • Point Transformer

    Prerequisites: MLP / PMC, Transformer

  • 125 XP

    Vision & 3D· ◎

    • 3D CNN
    • Temporal Networks
    • Video Transformer

    Prerequisites: Image, Sequences

  • 125 XP

    Multimodal· ◎

    • Cross-modal architectures
    • Early / late fusion
    • Cross-attention

    Prerequisites: Image, Text, Audio

N8Major specializationsSpecializations · 55 branches · 161 sub-branches
  • 155 XP

    Language & agents

    • n-grams
    • BERT
    • GPT
    • Perplexity

    Prerequisites: Text, Predictive learning

  • 155 XP

    Language & agents

    • Text classification
    • NER
    • Parsing
    • Sentiment analysis

    Prerequisites: Text

  • 130 XP

    Language & agents

    • NMT
    • Multilingual
    • BLEU / COMET

    Prerequisites: Language models, Seq2Seq

  • 130 XP

    Language & agents

    • Extractive QA
    • Generative QA
    • Reading comprehension

    Prerequisites: Language models, Core NLP

  • 130 XP

    Vision & 3D

    • Classification
    • Image retrieval
    • Fine-grained recognition

    Prerequisites: Image

  • 130 XP

    Vision & 3D

    • YOLO
    • Faster R-CNN
    • DETR

    Prerequisites: Vision 2D

  • 155 XP

    Vision & 3D

    • Semantic
    • Instance
    • Panoptic
    • SAM

    Prerequisites: Vision 2D

  • 130 XP

    Vision & 3D

    • Stereo
    • Depth
    • Reconstruction

    Prerequisites: Image, Point clouds

  • 130 XP

    Vision & 3D

    • Multi-object tracking
    • Re-ID
    • Optical flow

    Prerequisites: Detection, Video

  • 130 XP

    Vision & 3D

    • NeRF
    • Gaussian Splatting
    • Differentiable rendering

    Prerequisites: Vision 3D

  • 130 XP

    Vision & 3D

    • 3D generation
    • Meshes
    • 3D scenes

    Prerequisites: Neural Rendering

  • 130 XP

    Vision & 3D

    • Semantic mapping
    • Scene graphs
    • Geolocation

    Prerequisites: Vision 3D, Tracking

  • 130 XP

    Audio & time series· ◎

    • CTC
    • Encoder-decoder
    • Diarization

    Prerequisites: Audio, Seq2Seq

  • 130 XP

    Audio & time series

    • Vocoders
    • Prosody
    • Voice cloning

    Prerequisites: Audio

  • 130 XP

    Audio & time series

    • MIR
    • Transcription
    • Source separation

    Prerequisites: Audio

  • 130 XP

    Audio & time series

    • ARIMA / ETS
    • DeepAR
    • N-BEATS

    Prerequisites: Sequences, Statistics

  • 110 XP

    Audio & time series

    • Change points
    • Streaming detection

    Prerequisites: Forecasting, Anomaly detection

  • 130 XP

    Audio & time series

    • TCN
    • Temporal Fusion Transformer
    • PatchTST

    Prerequisites: Forecasting

  • 130 XP

    Search & recsys

    • BM25
    • Inverted index
    • nDCG / MRR

    Prerequisites: Text, Metrics

  • 110 XP

    Search & recsys

    • Matrix factorization
    • Implicit feedback

    Prerequisites: SVD & decompositions, Tabular

  • 110 XP

    Search & recsys

    • Learning to rank
    • CTR prediction

    Prerequisites: Tabular, Metrics

  • 130 XP

    Search & recsys

    • Two-tower
    • ANN / FAISS
    • Dense embeddings

    Prerequisites: Information Retrieval, Representation learning

  • 110 XP

    Search & recsys

    • Re-ranking
    • Query understanding

    Prerequisites: Retrieval, Ranking

  • 130 XP

    Search & recsys

    • Candidate generation
    • Sequential recommendation
    • Cold start

    Prerequisites: Collaborative filtering, Retrieval

  • 130 XP

    Graphs

    • Link prediction
    • Graph classification
    • Graph pooling

    Prerequisites: Graphs

  • 130 XP

    Graphs

    • Ontologies
    • TransE
    • Graph reasoning

    Prerequisites: Graphs

  • 110 XP

    Graphs· ◎

    • Graphormer
    • Graph positional encodings

    Prerequisites: Graphs, Transformer

  • 130 XP

    Graphs· ◎

    • Equivariance
    • Groups & symmetries
    • Manifolds

    Prerequisites: Advanced GNNs, Point clouds

  • 130 XP

    Generative

    • Explicit / implicit likelihood
    • Sampling
    • FID / IS

    Prerequisites: VAE, Representation learning

  • 155 XP

    RL & robotics

    • MDP
    • Bellman
    • Bandits
    • Exploration

    Prerequisites: Probability, Gradient descent

  • 130 XP

    RL & robotics

    • Q-Learning
    • DQN
    • Replay buffer

    Prerequisites: RL foundations, MLP / PMC

  • 130 XP

    RL & robotics

    • REINFORCE
    • Actor-critic
    • PPO

    Prerequisites: RL foundations, Optimizers

  • 130 XP

    RL & robotics

    • Behavior cloning
    • DAgger
    • Inverse RL

    Prerequisites: RL foundations, Supervised learning

  • 130 XP

    RL & robotics

    • Dyna
    • Learned planning
    • MuZero

    Prerequisites: Q-Learning / DQN, Policy gradient / PPO

  • 110 XP

    RL & robotics· ◎

    • Conservative Q-learning
    • Decision Transformer

    Prerequisites: Q-Learning / DQN, Transformer

  • 130 XP

    RL & robotics· ◎

    • Autonomous driving
    • Drones
    • Perception-decision stack

    Prerequisites: Tracking, Model-Based RL

  • 130 XP

    Reasoning & causality

    • SCM
    • Causal Graph
    • do-calculus

    Prerequisites: Graphical models

  • 130 XP

    Reasoning & causality

    • First-order logic
    • SAT / SMT
    • Constraint Learning

    Prerequisites: Graphical models, Decision trees

  • 130 XP

    Reasoning & causality

    • A*
    • MCTS
    • Classical planning (PDDL)

    Prerequisites: RL foundations

  • 130 XP

    Reasoning & causality

    • ATE / CATE
    • Uplift
    • Meta-learners

    Prerequisites: Causal graphs & SCM, Boosting

  • 110 XP

    Reasoning & causality

    • Counterfactual
    • Counterfactual explanations

    Prerequisites: Causal graphs & SCM

  • 130 XP

    Reasoning & causality· ◎

    • Neuro-symbolic
    • Program synthesis
    • Differentiable logic

    Prerequisites: Logic & constraints, Transformer

  • 110 XP

    Reasoning & causality· ◎

    • Learning to optimize
    • Neural combinatorial optimization

    Prerequisites: Search & planning, GNN

  • 130 XP

    AI for science

    • Surrogate models
    • Neural operators (FNO)
    • Simulation

    Prerequisites: Neural ODE, Regression

  • 130 XP

    AI for science

    • Molecules as graphs
    • Property prediction
    • Retrosynthesis

    Prerequisites: Graphs

  • 130 XP

    AI for science· ◎

    • Protein structure
    • Genomics
    • Single-cell

    Prerequisites: Graphs, Transformer

  • 130 XP

    AI for science

    • PINNs
    • PDEs
    • Physical constraints

    Prerequisites: Scientific ML, Gradient, Jacobian, Hessian

  • 130 XP

    AI for science

    • Docking
    • Virtual screening
    • Molecule generation

    Prerequisites: AI for Chemistry, Computational Biology AI

  • 130 XP

    Trust & safety

    • SHAP
    • LIME
    • Saliency maps

    Prerequisites: Boosting, CNN

  • 130 XP

    Trust & safety

    • FGSM / PGD
    • Robustness
    • Defenses

    Prerequisites: Modern CNNs, Gradient, Jacobian, Hessian

  • 130 XP

    Trust & safety

    • Differential privacy
    • Homomorphic encryption
    • Secure aggregation

    Prerequisites: Federated learning, Statistics

  • 130 XP

    Trust & safety

    • Deep ensembles
    • Conformal prediction
    • Bayesian deep learning

    Prerequisites: Calibration & uncertainty, Bayesian models

  • 130 XP

    Trust & safety· ◎

    • Probing
    • Circuits
    • Sparse autoencoders

    Prerequisites: Explainable AI, Transformer

  • 130 XP

    Trust & safety· ◎

    • Alignment
    • Red teaming
    • Capability evaluations

    Prerequisites: Interpretability, Adversarial ML, Language models

  • 130 XP

    Multimodal

    • CLIP
    • Image-text alignment
    • Audio-visual

    Prerequisites: Multimodal, Contrastive learning

Milestone 3 · between N8 and N94 mandatory projects · mandatory to climb higher
  • 2,160 XP

    Supernova · 10 weeks

    A single network detects, segments and tracks every object in a driving video at 30 FPS. And holds up at night, in the rain and through blur, knowing when it is unsure.

    Deliverables

    • Tested repo: ImageNet or COCO backbone, detection, segmentation and ReID heads, in-house tracker
    • A single ONNX graph in FP16 and INT8 (mAP50 drop ≤ 1.5 points), weights trained in ≤ 150 GPU-hours
    • One-command evaluation harness: mAP, mIoU, HOTA, MOTA, IDF1, latency and memory per sequence
    • Robustness suite: corruptions at 3 severities, uncertainty (ensemble or MC dropout), calibration
    • Annotated video demo and technical report: ablations, latency budget, failure analysis

    Validation criteria

    1. 01Hidden BDD100K-style tests: mAP50 ≥ 50% (10 classes), mIoU ≥ 55% (19 classes), mHOTA ≥ 38% (MOT, 8 classes)
    2. 02Same model as criteria 1 and 3, video decoding and tracking included: ≥ 30 FPS at 720p on a T4, batch 1, p95 ≤ 40 ms
    3. 03Hidden corruptions, severities 1 to 3: mAP50 ≥ 60% of the uncorrupted score; D-ECE (IoU ≥ 0.5) ≤ 0.08 on clean frames

    Branches it mobilizes: Detection, Segmentation, Tracking, Multi-task learning, Metric learning, Uncertainty Quantification, Distribution shift, Profiling, GPU computing

  • 1,940 XP

    Quasar · 9 weeks

    Millions of passages, one question: BM25, dense retrieval, re-ranking and a neural reader find the answer in under a second. Or stay silent when there is none.

    Deliverables

    • Tested repo: BM25, dense retrieval, re-ranking, extractive or fine-tuned seq2seq reader (FiD-style)
    • BM25 and ANN (FAISS) indexes over more than 5 million passages, rebuilt with one command
    • Trained cross-encoder and LambdaMART, two-tower distilled from the cross-encoder, weights released
    • Search and answer API with a web demo: cited source passage, or an explicit abstention
    • Evaluation report: nDCG, MRR, EM, F1, per-stage latency, ablations, error analysis

    Validation criteria

    1. 01Hidden MS MARCO-style queries: MRR@10 ≥ 0.36 and nDCG@10 ≥ 0.42, from checkpoints never trained for IR or QA
    2. 02Hidden answerable questions, end to end: EM ≥ 40%, F1 ≥ 48%, abstention ≤ 10%; unanswerable ones: abstention ≥ 70%
    3. 03Same config as criteria 1 and 2, one T4 GPU and 8 vCPUs: p95 ≤ 200 ms (re-ranked top 10), ≤ 1 s (answer), ANN index ≤ 8 GB

    Branches it mobilizes: Information Retrieval, Retrieval, Ranking, Search & Ranking, Question answering, Language models, Metric learning, Boosting, Calibration & uncertainty

  • 2,160 XP

    Nebula · 10 weeks

    Predict the energy, forces and toxicity of an unseen molecule from its graph and 3D geometry. And say how sure you are, and why.

    Deliverables

    • Library: GNN, graph transformer, SE(3)-equivariant network, symmetry and chirality tests
    • Data pipeline: 3D conformers, featurization, reproducible scaffold splits
    • Weights for all three families and the final ensemble, each with its model card
    • Uncertainty module: deep ensembles, conformal prediction calibrated on held-out scaffolds
    • Report: comparative benchmark, per-atom attribution maps, applicability domain

    Validation criteria

    1. 01Unseen scaffolds: energy and force MAE ≤ 0.7 × the SchNet baseline (SPICE-style), toxicity ROC-AUC ≥ 0.76 (Tox21-style)
    2. 02Energy on ≥ 2,000 hidden molecules: 90% conformal intervals, 87 to 93% coverage, uncertainty-error Spearman ≥ 0.3
    3. 03On the provided synthetic-motif task, the per-atom attribution API recovers the responsible motifs: AUROC ≥ 0.80

    Branches it mobilizes: AI for Chemistry, Advanced GNNs, Graph Transformers, Geometric Deep Learning, Uncertainty Quantification, Explainable AI, Validation, Distribution shift, Stacking & voting

  • 2,160 XP

    Black hole · 10 weeks

    Five ways of learning to decide face off in the same simulated world: DQN, PPO, MCTS, imitation and offline RL. Verdict over 10 seeds, not one lucky run.

    Deliverables

    • Tested repo: DQN, PPO, MCTS over a learned model, BC and CQL agents, fixed seeds, ≤ 200 GPU-hours
    • Off-policy evaluation library: IS, WIS, doubly robust, FQE, known-value tests
    • Checkpoints for every agent, retrainable seed by seed, plus a script that regenerates every figure
    • Table over seeds 0 to 9: normalized score (0 random, 1 expert), IQM, 95% bootstrap CIs
    • Report and videos: behaviors, costs, failures, counterfactual explanations of decisions

    Validation criteria

    1. 01Best online agent, ≤ 10M steps, seeds 0 to 9, hidden levels: normalized IQM ≥ 0.70, lower bound of the 95% CI ≥ 0.60
    2. 02Offline agent, provided logs only: normalized score ≥ behavior policy + 0.15 and ≥ BC + 0.05, over 10 seeds
    3. 03Off-policy evaluation of 12 hidden policies supplied by the CI: Spearman ≥ 0.7, regret@1 ≤ 0.1, median absolute error ≤ 0.1

    Branches it mobilizes: Q-Learning / DQN, Policy gradient / PPO, Model-Based RL, Search & planning, Imitation learning, Offline RL, Counterfactuals, Treatment effect, Statistics

N9Generative AISpecializations · 11 branches · 31 sub-branches
  • 140 XP

    Generative· ◎

    • PixelRNN
    • WaveNet
    • GPT

    Prerequisites: Generative modeling, Language models, The search engine that answers

  • 140 XP

    Generative

    • AE
    • VAE
    • VQ-VAE

    Prerequisites: Generative modeling

  • 140 XP

    Generative· ◎

    • DCGAN
    • Conditional GAN
    • StyleGAN

    Prerequisites: Generative modeling, Modern CNNs

  • 140 XP

    Generative

    • RealNVP
    • Glow
    • Change of variables

    Prerequisites: Generative modeling, Gradient, Jacobian, Hessian

  • 140 XP

    Generative

    • Energy functions
    • Langevin dynamics
    • Contrastive divergence

    Prerequisites: Generative modeling, Molecular oracle

  • 140 XP

    Generative· ◎

    • DDPM
    • DDIM
    • Score matching

    Prerequisites: Energy-Based Models, Modern CNNs, Molecular oracle

  • 140 XP

    Generative

    • Latent Diffusion
    • Classifier-free guidance
    • ControlNet

    Prerequisites: Diffusion, Generative autoencoders

  • 115 XP

    Generative

    • Flow Matching
    • Rectified flow

    Prerequisites: Diffusion, Normalizing Flows

  • 115 XP

    Generative· ◎

    • DiT
    • Patchified latents

    Prerequisites: Latent Diffusion, Vision Transformers

  • 140 XP

    Audio & time series· ◎

    • Neural codecs
    • Music generation
    • Expressive speech synthesis

    Prerequisites: Autoregressive, TTS, Diffusion

  • 140 XP

    Multimodal· ◎

    • Text-to-image
    • Text-to-video
    • Text-to-3D

    Prerequisites: Diffusion Transformer, Multimodal learning, 3D AI

N10Foundation ModelsSpecializations · 17 branches · 49 sub-branches
  • 150 XP

    Language & agents· ◎

    • Web-scale data
    • Scaling laws
    • Pretraining objectives

    Prerequisites: Language models, Autoregressive

  • 150 XP

    Language & agents

    • SFT
    • Instruction tuning
    • Synthetic data

    Prerequisites: Pretraining

  • 150 XP

    Language & agents

    • Sampling
    • KV Cache
    • Speculative decoding

    Prerequisites: Pretraining

  • 150 XP

    Language & agents· ◎

    • Reward models
    • RLHF
    • DPO

    Prerequisites: Post-training, Policy gradient / PPO, Decision arena

  • 175 XP

    Language & agents

    • Fine-tuning
    • LoRA
    • QLoRA
    • Adapters

    Prerequisites: Post-training

  • 150 XP

    Language & agents

    • In-context learning
    • Chain-of-thought
    • Prompt engineering

    Prerequisites: Post-training

  • 125 XP

    Language & agents

    • Teacher / student
    • Reasoning distillation

    Prerequisites: Post-training, The search engine that answers

  • 150 XP

    Language & agents

    • INT8 / INT4
    • GPTQ / AWQ
    • FP8

    Prerequisites: Inference, Real-time video perception

  • 150 XP

    Language & agents

    • Continuous batching
    • PagedAttention
    • Latency / throughput

    Prerequisites: Inference, The search engine that answers

  • 150 XP

    Language & agents

    • GPT-style decoders
    • Long context
    • Reasoning models

    Prerequisites: Preference learning, Adaptation (PEFT), Mixture of Experts

  • 150 XP

    Multimodal· ◎

    • Vision encoders
    • Visual instruction tuning
    • Documents & OCR

    Prerequisites: LLM, Multimodal learning, Vision 2D, Real-time video perception

  • 150 XP

    Audio & time series· ◎

    • Whisper-like
    • Speech LLM
    • Audio-LM

    Prerequisites: ASR, Pretraining

  • 125 XP

    Audio & time series· ◎

    • Zero-shot forecasting
    • Series tokenization

    Prerequisites: Temporal DL, Pretraining

  • 125 XP

    Graphs· ◎

    • Graph pretraining
    • Molecular models

    Prerequisites: Graph Transformers, Pretraining, Molecular oracle

  • 150 XP

    Vision & 3D· ◎

    • Video-LM
    • Video generation
    • Temporal grounding

    Prerequisites: VLM, Tracking, Diffusion Transformer, Real-time video perception

  • 150 XP

    RL & robotics· ◎

    • Predictive models
    • Learned simulation
    • Video world models

    Prerequisites: Model-Based RL, Diffusion Transformer, Pretraining, Decision arena

  • 150 XP

    Multimodal· ◎

    • Unified tokenization
    • Omni models
    • Audio-LM

    Prerequisites: VLM, Speech Foundation Models, Multimodal Generative Models

N11Agentic AISpecializations · 10 branches · 30 sub-branches
  • 160 XP

    Language & agents

    • Function calling
    • APIs
    • Sandboxing

    Prerequisites: LLM, Prompt & context learning

  • 135 XP

    Language & agents

    • JSON schema
    • Constrained decoding

    Prerequisites: Tool calling

  • 160 XP

    Language & agents· ◎

    • Vector DB
    • Hybrid search
    • Embeddings

    Prerequisites: Structured outputs, Retrieval

  • 185 XP

    Language & agents· ◎

    • Chunking
    • Re-ranking
    • GraphRAG
    • RAG evaluation

    Prerequisites: Retrieval, Knowledge Graphs, Search & Ranking

  • 160 XP

    Language & agents· ◎

    • Short / long-term memory
    • Episodic memory
    • Summaries

    Prerequisites: RAG, Memory & Hopfield

  • 160 XP

    Reasoning & causality· ◎

    • Task decomposition
    • ReAct
    • Plan-and-execute

    Prerequisites: Memory, Search & planning

  • 160 XP

    Reasoning & causality

    • Tree of thoughts
    • Self-consistency
    • Test-time compute

    Prerequisites: Planning, Neuro-symbolic

  • 160 XP

    Language & agents· ◎

    • Agent loop
    • Guardrails
    • Agent evaluation

    Prerequisites: Reasoning / Search, AI Safety

  • 160 XP

    Language & agents

    • Orchestration
    • Roles
    • Communication

    Prerequisites: Single agent

  • 160 XP

    Language & agents· ◎

    • Long-horizon agents
    • Computer use
    • Self-improvement

    Prerequisites: Multi-agent, VLM

Milestone 4 · between N11 and N128 mandatory projects · mandatory to climb higher
  • 3,170 XP

    Black hole · 12 weeks

    You code your LLM from scratch, pretrain it, align it with DPO and serve it quantized, all within 300 A100 hours. Every stage is measured: bits per byte, win rate, throughput.

    Deliverables

    • From-scratch code, fixed interfaces, hidden tests: BPE, transformer, training loop, DPO
    • Data pipeline with deduplication on the assigned corpus, trained BPE tokenizer
    • Three checkpoints (base, SFT, aligned), two LoRA adapters, model cards, an eval suite in CI
    • From-scratch INT4 server: continuous batching, KV cache, speculative decoding
    • Scaling report: IsoFLOP over 3 budgets and 4 sizes, optimal size predicted then verified

    Validation criteria

    1. 01Hidden split of the assigned corpus: ≤ 1.0 bits per byte; largest model's loss predicted within 3% by your scaling law
    2. 02Length-controlled win rate ≥ 60% against the provided reference SFT: swapped positions, fixed judge, 500 secret prompts
    3. 03INT4: ≤ 3% more bits per byte; speculative decoding: ≥ 60% acceptance, outputs identical under greedy decoding

    Branches it mobilizes: Text, Language models, Pretraining, Post-training, Inference, Preference learning, Adaptation (PEFT), Quantization, Serving

  • 2,900 XP

    Quasar · 11 weeks

    You wire a vision encoder into an LLM, align image and text, then instruction-tune it with LoRA. It answers questions about PDFs, tables and charts.

    Deliverables

    • From-scratch code: projector, tiling (768 px min.), two training phases, 200 A100 hours max
    • Visual instruction set of 100k pairs, 20k with grounding boxes, no leakage into the test sets
    • Checkpoints from both phases, merged LoRA, model card listing the known limitations
    • Web demo: drop in a PDF, the answer cites the page and region it used
    • Eval report on the public splits: resolution and tiling ablations, error analysis

    Validation criteria

    1. 01Hidden split, text LLM of 3B max: ANLS ≥ 70% on documents, ≥ 60% accuracy on charts (5% tolerance)
    2. 02On 200 extractive questions, the cited region overlaps the answer's line (IoU > 0.5) in ≥ 60% of cases
    3. 03Eval rerun with one command from the weights: score within 0.5 point; 5% of phase 2 replayed in CI, loss within 2%

    Branches it mobilizes: Vision Transformers, Vision 2D, Detection, Multimodal learning, Post-training, Adaptation (PEFT), VLM, Structured outputs

  • 2,900 XP

    Quasar · 11 weeks

    An agent that searches, reads, codes and fixes itself, using an assigned model. Your milestone 3 search engine becomes its tool; you add memory, planning, multiple agents and guardrails.

    Deliverables

    • Agent written from scratch, no agent framework: loop, typed tools, execution sandbox
    • Multi-hop agentic RAG on your milestone 3 engine, reindexed over 10k documents, GraphRAG optional
    • Long-term memory layer: summaries, episodes, retrieval, recall tests in CI
    • Tree-search planner and multi-agent orchestrator: roles, protocol, arbitration
    • Public benchmark of 50 auto-checked tasks, with a dashboard of runs and costs

    Validation criteria

    1. 01Hidden benchmark of 150 auto-checked tasks, assigned model: ≥ 55% success, versus 25% for the provided baseline
    2. 02Over 200 prompt-injected documents: zero out-of-scope actions, ≤ 5% false refusals on clean tasks
    3. 03CI ablations: removing RAG, memory, search or multi-agent costs ≥ 5 points each; median cost ≤ €0.50 per task

    Branches it mobilizes: Tool calling, Structured outputs, RAG, Memory, Planning, Reasoning / Search, Multi-agent, Autonomous systems, AI Safety

  • 2,900 XP

    Nebula · 11 weeks

    You train your own image generator on COCO within 300 A100 hours: a latent autoencoder, then a diffusion transformer trained with flow matching. Judged by FID, CLIP score and real humans.

    Deliverables

    • From-scratch code: f8 VAE, DiT, flow-matching loss, sampler and classifier-free guidance
    • VAE and generator checkpoints (EMA weights), run configs, loss and FID curves
    • Gallery of 1,000 samples with prompts and seeds, regenerated with one command
    • Eval report: FID, CLIP score, rFID, curves across guidance scales and step counts
    • Web demo: pick the prompt, guidance scale, step count and seed

    Validation criteria

    1. 01FID-10k ≤ 25 at 256 px on the hidden COCO split; ViT-L/14 CLIP score ≥ 90% of the real images' score
    2. 02Blind study run by the platform: ≥ 60% of 500 votes against the provided same-budget baseline
    3. 03f8 VAE with 4 channels: PSNR ≥ 24 dB, rFID ≤ 5; one 256 px image with guidance in ≤ 3 s on the CI GPU

    Branches it mobilizes: Image, Multimodal learning, Generative modeling, Generative autoencoders, Diffusion, Latent Diffusion, Flow Matching, Diffusion Transformer

  • 2,900 XP

    Pulsar · 11 weeks

    An assistant that listens, understands and talks back, with no dead air. Streaming ASR and expressive TTS fine-tuned in French, an assigned LLM quantized and streamed, latency budget held.

    Deliverables

    • From-scratch full-duplex orchestrator: VAD, end of turn, streaming ASR, barge-in, streaming TTS
    • Fine-tuned ASR and TTS checkpoints (TTS on a pretrained codec), recipes, 150 A100 hours max
    • Full-duplex browser demo: you can cut the assistant off at any time
    • Latency bench: per-stage traces (VAD, ASR, LLM, TTS), p50 and p95, documented budget
    • Eval report: WER, MOS, UTMOS, latency, robustness to noise, end-of-turn errors

    Validation criteria

    1. 01French WER, streaming with the latency settings: ≤ 8% on the hidden clean test, ≤ 15% on the noisy one
    2. 02End of speech to first sound: p50 ≤ 500 ms, p95 ≤ 800 ms on the CI GPU; silent within 200 ms when interrupted
    3. 03MOS ≥ 3.8 (platform panel, 20 judges, 50 samples), speaker similarity ≥ 90% of that between real recordings

    Branches it mobilizes: Audio, ASR, TTS, Audio Gen, Speech Foundation Models, LLM, Quantization, Serving

  • 3,170 XP

    Black hole · 12 weeks

    At milestone 3, your agents had a simulator. Here, only offline episodes, in pixels: your world model learns the dynamics, and your agent trains and plans inside that dream.

    Deliverables

    • World model written from scratch, tested in CI: encoder, latent or DiT dynamics, imagined rollouts
    • Dynamics model checkpoints and an ensemble for uncertainty, within 200 A100 hours
    • Policy trained in imagination with an uncertainty penalty, plus a latent planner
    • Side-by-side videos: imagined and real rollouts over 50 episodes of the 3 assigned tasks
    • Report: error per horizon, agent score, policy chosen with your milestone 3 off-policy library

    Validation criteria

    1. 013 assigned DMC tasks at 64×64: 15-step PSNR ≥ 22 dB, 50-step predicted-return error ≤ 10% of the return range
    2. 02Trained in imagination on the provided dataset (10x fewer steps than the reference PPO), the agent reaches 80% of its score
    3. 033 seeds, std ≤ 10% of the mean; the grader reruns one with a single command: score within 2 std of the mean

    Branches it mobilizes: Sequences, Policy gradient / PPO, Model-Based RL, Offline RL, Uncertainty Quantification, Diffusion Transformer, Pretraining, World Models

  • 3,170 XP

    Nebula · 12 weeks

    Your milestone 3 oracle judged molecules. Here, you invent them: a foundation model pretrained on millions of graphs, a constrained generator, active screening by docking.

    Deliverables

    • Self-supervised pretraining repo over 2 million molecular graphs, 200 A100 hours max
    • Graph foundation model checkpoint, heads fine-tuned on 6 tasks with 1,000 labels each, model card
    • Constrained diffusion generator written from scratch, plus 500 candidates as SMILES
    • Active screening loop: uncertainty-guided acquisition, your milestone 3 oracle as a filter
    • Virtual screening report: docking, uncertainty, pretraining ablation

    Validation criteria

    1. 016 tasks with 1,000 labels each, hidden scaffold split: MAE ≥ 10% below the provided reference GNN's, on 5 of 6
    2. 02Top 100 on the assigned target: 100% valid, ≥ 80% novel, QED ≥ 0.5, SA ≤ 4, median Vina ≥ 1 kcal/mol below matched ZINC
    3. 03Hidden library of 500,000 molecules, ≤ 5,000 dockings metered by the platform: your loop recovers ≥ 50% of the top 500

    Branches it mobilizes: Graphs, Advanced GNNs, Geometric Deep Learning, AI for Chemistry, AI for Drug Discovery, Uncertainty Quantification, Diffusion, Pretraining, Graph Foundation Models

  • 2,900 XP

    Supernova · 11 weeks

    You attack an assigned 1-3B open-weight model, measure what it can do, then look inside it, within 100 A100 hours. Red teaming, probes, SAEs, privacy leaks, a published report.

    Deliverables

    • Red teaming suite: 1,000 attack prompts, versioned and scored, replayed after every mitigation
    • Capability evals on public proxies (WMDP, sandboxed CTF), with no new hazardous item
    • SAE and probes written from scratch, on the assigned model or your homemade LLM; circuits optional
    • Leakage tests: canaries, membership inference, a DP variant, epsilon computed in CI (delta 1e-5)
    • Published audit report: threat model, findings, mitigations and a risk budget

    Validation criteria

    1. 01Hidden and adaptive attacks: ≥ 40% fewer jailbreaks after mitigation, ≤ 5% false refusals, utility within 2 points
    2. 02SAE: variance explained ≥ 85%, L0 ≤ 100; 16 of 20 labeled features have an ablation effect ≥ 5x the random control
    3. 03Canaries: membership inference AUC ≥ 0.70 without DP, ≤ 0.55 after DP fine-tuning (epsilon ≤ 8), utility within 5 points

    Branches it mobilizes: Adversarial ML, Privacy-Preserving ML, Interpretability, AI Safety, LLM, Post-training, Preference learning, Adaptation (PEFT), Prompt & context learning

N12Physical / Embodied AISpecializations · 17 branches · 43 sub-branches
  • 170 XP

    RL & robotics

    • Robot vision
    • Pose 6D
    • Depth

    Prerequisites: Detection, Vision 3D

  • 140 XP

    RL & robotics

    • LiDAR point clouds
    • LiDAR segmentation

    Prerequisites: Point clouds, Segmentation

  • 140 XP

    RL & robotics

    • Closed loop
    • Tuning

    Prerequisites: Calculus

  • 170 XP

    RL & robotics· ◎

    • Teleoperation
    • Behavior cloning
    • Diffusion policy

    Prerequisites: Imitation learning, Diffusion, Text-to-image generator

  • 170 XP

    RL & robotics

    • Sim RL
    • Reward shaping
    • Online RL

    Prerequisites: Policy gradient / PPO

  • 140 XP

    RL & robotics

    • IMU
    • Visual-inertial odometry

    Prerequisites: Robot perception

  • 140 XP

    RL & robotics

    • Predictive control
    • Constrained optimization

    Prerequisites: PID, Optimization, World model: the dreaming agent

  • 140 XP

    RL & robotics

    • Robot datasets
    • Offline RL

    Prerequisites: Offline RL, Imitation / BC, World model: the dreaming agent

  • 170 XP

    RL & robotics

    • Kalman filter
    • Particle filter
    • Multi-sensor fusion

    Prerequisites: VIO, LiDAR

  • 170 XP

    RL & robotics· ◎

    • Motion planning
    • RRT
    • Trajectory optimization

    Prerequisites: MPC, Search & planning

  • 170 XP

    RL & robotics

    • Domain randomization
    • Simulators
    • System identification

    Prerequisites: Online RL, Imitation / BC, World model: the dreaming agent

  • 170 XP

    RL & robotics· ◎

    • Visual SLAM
    • Loop closure
    • Pose graph

    Prerequisites: Sensor Fusion, Spatial AI

  • 170 XP

    RL & robotics

    • Grasping
    • Kinematics
    • Contact-rich

    Prerequisites: Trajectory, Robot perception

  • 140 XP

    RL & robotics

    • Dynamics models
    • Latent planning

    Prerequisites: World Models, Sim2Real, SLAM, World model: the dreaming agent

  • 140 XP

    Multimodal· ◎

    • Spatial grounding
    • Language instructions

    Prerequisites: VLM, Robot perception, VLM for documents and charts

  • 170 XP

    RL & robotics· ◎

    • Vision-Language-Action
    • Action tokens
    • Generalist policies

    Prerequisites: VLMs for robots, Robot world models, Robot offline RL, Manipulation, VLM for documents and charts

  • 140 XP

    RL & robotics

    • Multi-embodiment
    • Real-world learning

    Prerequisites: VLA, Sim2Real

N13Production and AI systemsCross-cutting layer · 14 branches · 40 sub-branches
  • 175 XP

    Production & systems

    • ETL
    • Orchestration
    • Data quality

    Prerequisites: SQL, Pandas / Polars, VLM for documents and charts

  • 150 XP

    Production & systems

    • Metrics & artifacts
    • Reproducibility

    Prerequisites: Git, Validation, Molecules: from graph to docking

  • 175 XP

    Production & systems

    • CUDA
    • Triton
    • FlashAttention

    Prerequisites: GPU computing, Transformer, Text-to-image generator

  • 150 XP

    Production & systems

    • Online / offline
    • Point-in-time

    Prerequisites: Data pipelines

  • 175 XP

    Production & systems

    • Versioning
    • Promotion
    • Lineage

    Prerequisites: Experiment tracking, Safety audit of an LLM

  • 205 XP

    Production & systems

    • Data Parallel
    • Model Parallel
    • Tensor Parallel
    • Pipeline Parallel

    Prerequisites: CUDA & kernels, Pretraining, Homemade LLM: tokenizer to serving

  • 175 XP

    Production & systems

    • Quantization
    • Distillation
    • Pruning

    Prerequisites: Quantization, Distillation, Text-to-image generator

  • 175 XP

    Production & systems

    • FSDP
    • DeepSpeed ZeRO
    • Mixed precision

    Prerequisites: Distributed training, Homemade LLM: tokenizer to serving

  • 175 XP

    Production & systems

    • Slurm / Kubernetes
    • Interconnect
    • Scheduling

    Prerequisites: Distributed training, Linux / Shell, Homemade LLM: tokenizer to serving, Molecules: from graph to docking

  • 175 XP

    Production & systems

    • ONNX
    • TensorRT
    • vLLM

    Prerequisites: Compression, Serving, Homemade LLM: tokenizer to serving, Real-time voice assistant

  • 175 XP

    Production & systems

    • Mobile & embedded
    • Microcontrollers
    • On-device

    Prerequisites: Compression, Real-time voice assistant

  • 175 XP

    Production & systems

    • APIs
    • Batch / real-time
    • A/B testing

    Prerequisites: Model registry, Inference runtimes, Autonomous agent: tools and memory

  • 175 XP

    Production & systems

    • Drift
    • Alerting
    • Observability

    Prerequisites: Model serving, Distribution shift, Autonomous agent: tools and memory, Safety audit of an LLM

  • 150 XP

    Production & systems

    • Retraining
    • Feedback loops

    Prerequisites: Monitoring, Feature stores

The comet projectsFour capstone projects outside the curriculum, reserved for students who have reached the tree's final circles (N12 and N13).
  • 14,400 XP

    Final circles only · 40 weeks · 2,000 H100 GPU-hours

    End-to-end design of an embodied conversational assistant, modelled on Grok and its character Ani: pretraining a Mixture-of-Experts transformer, aligning it to a persona, real-time voice interaction and an animated 3D avatar. Everything runs on weights trained by the student, with no external API.

    Key technical challenge Sustaining persona consistency, voice latency and avatar synchronisation in real time, on an in-house pretrained model, without ever weakening the safeguards.

    Deliverables

    • Decoder MoE model with 16 top-2 experts (≈ 7B parameters, 1.3B active): RoPE, GQA, auxiliary-loss-free load balancing; pretrained on 300B deduplicated FR/EN tokens
    • Post-training pipeline: persona SFT, DPO, then reinforcement learning on consistency rewards; integrated web search tool
    • Full-duplex voice interaction: streaming ASR, expressive speech synthesis on a licensed voice, interruption handling
    • 3D avatar (VRM) rendered with WebGPU: phoneme-based lip sync, expressions and gaze driven by the model
    • Encrypted long-term memory, episodic and profile-based, viewable and erasable by the user
    • Safety framework: age verification, content classifiers, emotional-dependence detection, published red-teaming report

    Validation criteria

    1. 01Pretraining in ≤ 2,000 H100 GPU-hours at MFU ≥ 40%; after post-training: MMLU 5-shot ≥ 40%, HellaSwag ≥ 60%, GSM8K ≥ 25%
    2. 02Persona preferred ≥ 65% of the time over the school's reference, in a blind evaluation by 30 judges on 50-turn conversations; memory recall ≥ 90% after 30 sessions
    3. 03Real time on a single GPU: voice response latency p95 ≤ 600 ms, audio-lip offset ≤ 45 ms, avatar ≥ 60 fps on a laptop
    4. 04Safety, over 2,000 hidden attacks: bypass rate ≤ 2%, no prohibited content involving a minor, 100% of distress signals redirected

    Branches it mobilizes: Mixture of Experts, Pretraining, Post-training, Preference learning, Inference, Serving, Quantization, ASR, TTS, Speech Foundation Models, Neural Rendering, Memory, Tool calling, AI Safety, FSDP / DeepSpeed, CUDA & kernels

  • 11,520 XP

    Final circles only · 32 weeks · 400 A100 GPU-hours and a low-latency server

    End-to-end design of an algorithmic trading system operating continuously on crypto-assets and equity markets: market data acquisition, signal research, execution, risk management and audit. Performance is validated out of sample, then over 90 days of live operation on testnet and a demo account.

    Key technical challenge Establishing a signal whose profitability withstands transaction costs, slippage and multiple-testing correction, then confirming it on data never observed before.

    Deliverables

    • Tick-by-tick data pipeline: L2 order books from three crypto venues and equity feeds, microsecond timestamps, replayable columnar storage
    • Event-driven backtester: queue position, fees, slippage, latency, perpetual funding; purged walk-forward with embargo
    • Signal research: order-book transformer, LLM agent analysing news and earnings calls, regime detection; reinforcement-learning allocation
    • Reinforcement-learning execution, benchmarked against TWAP and VWAP; risk engine: VaR, CVaR, exposure limits, circuit breaker
    • 90 days of operation on testnet and a demo account, with no real capital, and a signed audit log of every decision

    Validation criteria

    1. 01Three-year out-of-sample walk-forward, costs included: Sharpe ratio ≥ 2, maximum drawdown ≤ 12%; Deflated Sharpe significant at the 5% level across all strategies tested
    2. 0290 days of live operation: Sharpe ratio ≥ 1.5, maximum drawdown ≤ 10%, gap to the backtest over the same period ≤ 30%
    3. 03Tick-to-order latency p99 ≤ 5 ms; circuit breaker triggered in ≤ 50 ms across the 12 crisis scenarios replayed by the school
    4. 04No look-ahead bias: the CI recomputes every signal with future data masked; any discrepancy is disqualifying

    Branches it mobilizes: Forecasting, Time-series anomalies, Temporal DL, Time-Series Foundation Models, Boosting, Distribution shift, Uncertainty Quantification, Policy gradient / PPO, Offline RL, RAG, Autonomous systems, Data pipelines, Inference runtimes, Monitoring

  • 12,960 XP

    Final circles only · 36 weeks · 3,000 H100 GPU-hours

    Building a video game in which every frame is generated in real time by a world model conditioned on the player's actions, with no graphics engine and no programmed game logic. The project covers agent-based data collection, diffusion transformer training, world memory and interactive deployment.

    Key technical challenge Preserving the world's spatial and temporal coherence for several minutes at 24 frames per second, despite the error accumulation inherent to autoregressive generation.

    Deliverables

    • Dataset of 50M frames from an open-source game, collected by reinforcement-learning agents, with annotated actions and states
    • Latent video autoencoder (≥ 16× compression) and causal action-conditioned diffusion transformer, distilled to one to four steps
    • World memory: persistent latent map and keyframe recall, keeping revisited places stable
    • Playable browser client, streamed over WebRTC from a single GPU
    • Technical report and public demonstration: ablations, documented failures, costs

    Validation criteria

    1. 01Real time: ≥ 24 fps at 320 × 240 on an RTX 4090, action-to-frame latency ≤ 80 ms
    2. 02Fidelity over 64 frames replayed with the reference engine's actions: PSNR ≥ 22 dB, FVD ≤ 200
    3. 03Persistence: a place revisited after 60 s within LPIPS ≤ 0.3 of the first visit; five-minute sessions without visible drift
    4. 04Human evaluation: 40 players tell real and generated 3-second sequences apart with ≤ 60% accuracy

    Branches it mobilizes: World Models, Video Foundation Models, Diffusion Transformer, Flow Matching, Latent Diffusion, Distillation, Model-Based RL, Policy gradient / PPO, Video, Neural Rendering, Quantization, Inference runtimes, GPU clusters, FSDP / DeepSpeed

  • 15,840 XP

    Final circles only · 44 weeks · 4,000 H100 GPU-hours and a robotics bench

    Development of a humanoid able to carry out natural-language instructions in an unfamiliar domestic environment: tidying, loading a dishwasher, handling deformable objects. A single vision-language-action model links perception to the 30 actuators; it is trained in simulation, through teleoperation and on its own failures, then validated on the school's robotics bench.

    Key technical challenge Transferring whole-body control from simulation to reality (balance, dexterous grasping, deformable objects) in an environment never seen before.

    Deliverables

    • Whole-body balance and locomotion controller, trained by massively parallel reinforcement learning (4,096 GPU environments) with domain randomisation
    • 500 h of demonstrations: virtual-reality teleoperation and human videos retargeted to the robot's kinematics
    • VLA model: vision-language backbone and flow-matching action head at 50 Hz, sub-task decomposition, error recovery
    • Perception: visual-inertial SLAM, open-vocabulary 3D segmentation, object and articulation pose estimation
    • Quantised onboard deployment on a Jetson-class module; safety: exclusion zones, emergency stop, torque limiting

    Validation criteria

    1. 01Simulation, three unseen kitchens × 50 instructions: ≥ 80% success, including ≥ 60% on long tasks (≥ 10 sub-tasks)
    2. 02School's physical bench: ≥ 60% success on 20 unseen instructions, no falls, no contact above 50 N
    3. 03Onboard perception-action loop ≤ 100 ms; ≥ 2 h of autonomy without human intervention
    4. 04Robustness: ≥ 70% of nominal performance under lighting changes, moved objects and 30 N disturbances

    Branches it mobilizes: Generalist robot, VLA, VLMs for robots, Manipulation, Sim2Real, Robot world models, Online RL, Imitation / BC, MPC, SLAM, VIO, Sensor Fusion, Robot perception, Edge inference, Compression