Section II
I studied Computer Science and Technology at Shenzhen University, and have kept studying since — filling the gaps the work exposes, and following whatever else interests me.
There is no fixed syllabus. A course usually starts from something at work: a system I do not fully understand sends me to the course that teaches how it works, and I take it from the first lecture to the last, labs included.
Since graduating, most of my studying has been open courseware. Computer science and AI are the day job; maths and physics shore up the foundations; economics, psychology, linguistics and philosophy are for interest. I am currently learning French at the Alliance Française.
2020 — 2024
B.Eng., Computer Science and Technology
A four-year engineering degree covering computer systems, software development and the mathematics underneath them, with programming assignments and lab work throughout. The main courses are listed below.
Coursework
Best universities but free, what can I say?
LecturesA close reading of the classic papers — MapReduce, GFS, Raft, ZooKeeper, Spanner — each weighed for its trade-offs in replication, consistency and fault tolerance.
Labs · GoMapReduce, then Raft (leader election, log replication, persistence, snapshots), a fault-tolerant key-value service on top of it and finally a sharded store, all passing randomised tests of partitions, crashes and reordered messages.
LecturesA relational database taken apart from the bottom up: page layout and the buffer pool, B+ tree and hash indexes, sort and join algorithms, query optimisation, concurrency control (2PL, OCC, MVCC) and ARIES recovery.
Labs · C++On the BusTub teaching database: a buffer pool manager with LRU-K replacement, an index, query executors and transactional concurrency control.
LecturesThe hardware–software boundary from the programmer's side, following CS:APP: bit-level data, x86-64 machine code, caches, linking, processes and signals, virtual memory, allocation, I/O and concurrency.
Labs · CData, Bomb and Attack Labs (bit manipulation, reverse-engineering assembly, buffer overflows), then Cache, Shell, Malloc and Proxy Labs.
Lecturesxv6, a Unix V6 descendant running on RISC-V, read line by line: system calls, page tables, traps and interrupts, scheduling, locking, file systems and crash recovery.
Labs · CExtending xv6 with new system calls, copy-on-write fork, an E1000 network driver, a finer-grained memory allocator, large files, symbolic links and mmap.
Taught by Keith Winstein, it starts from layering and the end-to-end principle and covers the core mechanisms of the internet: packet switching, reliable delivery, flow and congestion control, IP addressing and routing, ARP, NAT and DNS. The labs, in C++, build a working TCP/IP stack from the bottom up (Sponge/Minnow): a byte stream and an out-of-order reassembler, a TCP receiver that handles sequence-number wraparound, a TCP sender with a retransmission timer and sliding window, an ARP-backed network interface and a longest-prefix-match IP router, turning textbook protocols into code that talks to real hosts.
Taught by Erik Demaine, Jason Ku and Justin Solomon around the split between interface and implementation: from models of computation and asymptotic analysis, it covers data structures (hash tables, AVL trees, binary heaps), merge and radix sort, graph algorithms (BFS, DFS, topological sort, Bellman–Ford, Dijkstra, Johnson's all-pairs shortest paths) and dynamic programming through the SRTBOT framework, closing with reductions and NP-completeness. Problem sets pair proofs of correctness and complexity analysis with Python implementations, training the habit of modelling a real problem as a graph or a recurrence and then giving a provably efficient solution.
Organised around the modelling–inference–learning paradigm, it climbs from reflex models through state-based and variable-based models to logic, covering search (UCS, A*), Markov decision processes and reinforcement learning, games (Minimax, Expectimax), constraint satisfaction and Bayesian networks. Python assignments such as a Pac-Man adversarial agent and car tracking with particle filters train the habit of formalising real problems as computable models.
Stanford's classic machine learning course, built on rigorous derivation: starting from linear regression, logistic regression and generalised linear models (GLMs), it covers generative learning algorithms (GDA, Naive Bayes), kernel methods and SVMs, the bias–variance trade-off and learning theory, and unsupervised methods such as k-means, EM and PCA, before extending into reinforcement learning. Problem sets pair derivations with NumPy implementations, turning machine learning from a library call back into mathematics you can derive and prove.
Run as a flipped classroom alongside the deeplearning.ai modules, it starts by hand-writing forward and backward propagation in NumPy, then covers initialisation, regularisation (Dropout, BatchNorm), optimisers (Momentum, Adam), CNNs and RNNs/LSTMs, with dedicated teaching on engineering method: error analysis, train/dev/test splits and hyperparameter tuning. A full deep learning project closes the course, building a systematic path from prototype to deployment.
Using image classification as the way in, it hand-writes the forward and backward passes of kNN, Softmax, fully connected and convolutional networks in Python/NumPy, then moves to PyTorch for BatchNorm, Dropout, image captioning (RNN/Transformer), GANs and self-supervised learning, extending to object detection, semantic segmentation and Vision Transformers. It connects the design and training of deep vision models end to end, from pixels to semantics.
From word2vec embeddings and dependency parsing, through RNNs/LSTMs and sequence-to-sequence translation with attention, to Transformers, pretraining (BERT/GPT) and fine-tuning, it traces how modern NLP evolved. PyTorch assignments derive word vectors and implement a neural dependency parser, neural machine translation and Transformer pretraining, bringing the core mechanisms behind large language models down to runnable code.
Built on a from-scratch principle: hand-write a BPE tokeniser, a Transformer and the AdamW optimiser in PyTorch, then work through five assignments — the base model, systems optimisation (FlashAttention in Triton, distributed data parallelism), fitting scaling laws, filtering and deduplicating web-scale data, and alignment with SFT and reinforcement learning (GRPO). Together they span the full engineering of a large language model, from data through training to alignment.
MIT's more abstract, more rigorous track of real analysis, following Rudin's Principles of Mathematical Analysis: from the least-upper-bound property and the construction of the reals, it develops open and closed sets, compactness and connectedness in the setting of metric spaces, then sequences and series, continuity and uniform continuity, differentiation, the Riemann integral and uniform convergence of sequences of functions (the Weierstrass approximation theorem among them). Proof-only problem sets build the discipline of reasoning rigorously from definitions.
Linear algebra and analysis carried into infinite dimensions: from normed and Banach spaces, through the construction of the Lebesgue integral and L² spaces, to Hilbert spaces, orthonormal bases and the Riesz representation theorem, alongside the pillars of functional analysis (the Baire category theorem, uniform boundedness, the open mapping and closed graph theorems, Hahn–Banach). It closes with the spectral theorem for compact self-adjoint operators and Fourier series, laying a rigorous foundation for the Hilbert-space and operator language of quantum mechanics.
Taught by John Tsitsiklis alongside the classic Bertsekas and Tsitsiklis text: from the axioms of probability and conditioning (Bayes' rule), it covers discrete and continuous random variables, joint and derived distributions, conditional expectation, Bayesian inference (MAP, LMS) and classical statistics (maximum likelihood, confidence intervals), then limit theorems (the law of large numbers, the central limit theorem), Bernoulli and Poisson processes and Markov chains, building the probabilistic intuition and computational fluency that machine learning and reasoning under uncertainty rest on.
Taught by Philippe Rigollet: starting from statistical modelling and parametric estimation, it develops maximum likelihood and the method of moments, M-estimation, Fisher information and asymptotic normality (the delta method), then hypothesis testing (Wald, likelihood-ratio, chi-squared goodness-of-fit and Kolmogorov–Smirnov tests) and Bayesian statistics (conjugate and Jeffreys priors), extending to linear regression, generalised linear models and principal component analysis. Rigorous derivations answer how to draw trustworthy conclusions from data, closing the loop from probability models to inference.
Taught by Michael Sipser from his own Introduction to the Theory of Computation: from finite automata, regular languages and the pumping lemma, through context-free grammars and pushdown automata, to Turing machines, the Church–Turing thesis, undecidability and reductions. The second half turns to complexity: P versus NP, the Cook–Levin theorem and NP-completeness, space complexity (Savitch's theorem, PSPACE-completeness), the hierarchy theorems and probabilistic computation, answering the field's founding question of what can be computed, and at what cost.
Discrete mathematics as a tool for real problems: from counting and generating functions, probability and Chernoff bounds, it covers linear programming and duality, number theory and RSA, entropy and data compression (Huffman coding) and error-correcting codes (Hamming codes among them), showing how discrete structures underpin core techniques in algorithms, communication and cryptography.
Taught by Barton Zwiebach, it starts from the experiments that broke classical physics (the photoelectric effect, Compton scattering, electron diffraction), introduces wave–particle duality and the de Broglie relations, and builds the probabilistic interpretation of the wavefunction, Fourier analysis and the uncertainty principle. It then solves the Schrödinger equation in earnest: stationary states and one-dimensional potentials (infinite and finite wells, the delta function, barriers and tunnelling), the harmonic oscillator through ladder-operator algebra, and finally central potentials in three dimensions, angular momentum and the hydrogen atom, laying down both the computational backbone and the physical intuition of quantum mechanics.
The second part of Barton Zwiebach's quantum trilogy, rebuilding quantum mechanics on linear algebra: from the Stern–Gerlach experiment and spin one-half, it develops vector spaces, linear operators, inner products and the spectral theorem, introducing Dirac notation and the generalised uncertainty principle. It then applies them to coherent and squeezed states of the harmonic oscillator, two-state systems (nuclear magnetic resonance, the ammonia maser), time evolution in the Schrödinger and Heisenberg pictures, and finally the theory of angular momentum, central potentials and the addition of angular momentum (the fine structure of hydrogen), lifting wave mechanics into an abstract framework built on Hilbert space.
The last part of the trilogy turns to realistic problems with no exact solution: from charged particles in electromagnetic fields and Landau levels, it develops non-degenerate and degenerate perturbation theory and applies it to the fine structure of hydrogen and the Zeeman and Stark effects, then covers the WKB approximation (tunnelling and connection formulas), time-dependent perturbation theory and Fermi's golden rule, and the adiabatic approximation and Berry's phase, closing with scattering theory (partial waves, the Born approximation), training the ability to extract physical conclusions from complicated systems by approximation.
MIT's first graduate semester of quantum field theory, on how quantum mechanics and special relativity fit together: from classical field theory, Lagrangians and Noether's theorem, it covers the Lorentz group and its representations, then canonically quantises the free scalar, Dirac and gauge fields, explaining antiparticles and the spin–statistics connection. Interactions follow, with Wick's theorem and perturbative expansion leading to the Feynman rules, cross sections and decay rates, closing with tree-level QED processes (e⁺e⁻ → μ⁺μ⁻, Compton scattering), laying the theoretical groundwork for the Standard Model of particle physics.
Taught by Scott Hughes with Carroll's text as reference: from special relativity restated in tensor language, it builds the tools of differential geometry (manifolds, the metric, covariant derivatives and parallel transport, geodesics, the Riemann curvature tensor) and derives the Einstein field equations, then covers the weak-field limit and linearised gravity, the generation and propagation of gravitational waves, the Schwarzschild solution and black holes, and FRW cosmology, turning gravity as the curvature of spacetime from a slogan into a geometric theory you can compute with.
Taught by Alan Guth, who proposed cosmic inflation: from Hubble expansion and the cosmological principle, it derives the Friedmann equations from Newtonian mechanics and covers matter- and radiation-dominated expansion, the non-Euclidean geometry of open and closed universes, black-body radiation and the cosmic microwave background, the thermal history of the universe and Big Bang nucleosynthesis, and the observational evidence for dark matter and dark energy (the cosmological constant). The second half turns to the flatness, horizon and magnetic monopole problems of the standard Big Bang model, how inflation resolves them and seeds the primordial density perturbations, extending to eternal inflation and the multiverse, building an understanding of the origin and evolution of the universe from the observational data up.
Taught by Jonathan Gruber: from supply and demand, it derives consumer demand from utility maximisation under a budget constraint and models firm decisions through production functions and cost curves, then analyses perfect competition, monopoly and oligopoly and their welfare consequences (consumer and producer surplus). The second half turns to factor markets (labour and capital), choice under uncertainty and international trade, then externalities, public goods, taxation and redistribution, showing when markets work, when they fail, and what justifies government intervention and what it costs.
Taught by Ricardo Caballero from Blanchard's Macroeconomics: from the goods and financial markets it builds the IS-LM model, adds the labour market and the Phillips curve to derive AS-AD, and traces how monetary and fiscal policy act differently in the short, medium and long run. It then covers the Solow growth model and technological progress, the role of expectations in consumption and investment, and exchange rates and the Mundell–Fleming model in the open economy, applying these tools to financial crises and the liquidity trap, building a complete framework for analysing macroeconomic fluctuations and policy.
MIT's graduate course in development economics, in the J-PAL tradition of studying poverty through randomised controlled trials: from agricultural household models and market imperfections, it covers human capital (education and health), intrahousehold allocation, credit markets and microfinance, risk sharing and consumption smoothing, savings, land and labour markets and technology adoption, extending to governance, corruption and political economy. Built around close reading of current empirical papers, it stresses causal identification through randomised trials, instrumental variables and natural experiments, showing how households in developing countries actually behave and which policies genuinely improve the lives of the poor.
Taught by John Geanakoplos, it builds finance from the ground up on general equilibrium: from Irving Fisher's intertemporal model it derives interest rates and present value, applying them to bond pricing, the yield curve, mortgages and prepayment, and social security. Uncertainty follows, with risk aversion, portfolio choice, the capital asset pricing model (CAPM) and dynamic hedging, closing with the leverage cycle, Geanakoplos's own theory of how leverage and collateral amplify boom and crash, applied to the 2008 financial crisis, recasting Wall Street practice as rigorous economic models.
Taught by Robert Shiller, 2013 Nobel laureate in economics, it examines the financial system as a machine for managing risk: from probability, diversification and the principles of insurance, it weighs the efficient-markets hypothesis against behavioural finance, then works through the markets for bonds, stocks, real estate, forwards and futures, and options, and the institutional design of banks, investment banks, central banks and monetary policy, and financial regulation, with guest lectures from practitioners, showing how finance serves society and why it produces bubbles and crises.
Taught by Ben Polak, who introduces each idea through games played in the lecture hall: from strict dominance and the iterated deletion of dominated strategies, it covers Nash equilibrium and its application to Cournot and Bertrand competition, mixed strategies and evolutionarily stable strategies, then sequential games, with backward induction, the Stackelberg model, information sets and subgame-perfect equilibrium, extending to cooperation in repeated games, bargaining, signalling and auctions under asymmetric information, training the ability to formalise and solve real strategic interaction.
Taught by Nancy Kanwisher, who discovered the fusiform face area: starting from neuroanatomy and a real brain dissection, it lays out the methods of cognitive neuroscience (fMRI, EEG and MEG, TMS, studies of patients with brain damage, intracranial recording) and asks of each what it can and cannot show. It then works through the brain's functional specialisation, region by region: face recognition, scene perception and navigation, bodies and visual motion, number, music, language, and the network that represents other minds, extending to how plasticity and development shape that division of labour and to deep networks as models of visual cortex, building an understanding of how the mind is organised in the brain from the experimental evidence up.
Taught by John Gabrieli: from the research methods of psychology and the biological basis of behaviour, it covers perception and attention, consciousness and sleep, learning (classical and operant conditioning), memory (the amnesic patient H.M. among the cases), language, thinking and intelligence, then emotion and motivation, development across the lifespan, personality, social psychology (conformity and obedience in the Asch and Milgram experiments), stress and health, and psychological disorders and their treatment, reading every claim against the experiments behind it and building a full map of the mind from neurons to society.
From psychophysics (thresholds, Weber's law, signal detection theory), it follows the visual pathway from the optics of the eye through the retina, lateral inhibition and receptive fields to the lateral geniculate nucleus and primary visual cortex, then covers colour vision (trichromacy and opponent processes), lightness constancy and illusions, spatial vision and contrast sensitivity, motion, depth and stereopsis, and object recognition and attention, extending to hearing (the cochlea, pitch and sound localisation), touch, taste and smell. It shows that perception is not a recording of the world but an inference the brain makes from incomplete signals, and sets human vision against computer vision.
Taught by Michale Fee: from the biophysics of a single neuron, it builds the equivalent-circuit model of the membrane, the Hodgkin–Huxley model of the action potential, integrate-and-fire neurons, dendrites and cable theory, and synaptic transmission, then turns to the neural code, with spike trains and firing rates, receptive fields and the spike-triggered average. Along the way it teaches the mathematics of data analysis (convolution and filtering, Fourier and spectral analysis, matrices, eigenvectors and principal component analysis), closing with network models: rate models, feed-forward networks and the perceptron, recurrent networks, line attractors as neural integrators, and Hopfield networks for memory. MATLAB problem sets simulate and analyse each model, showing how computation emerges from biological neural circuits and how they compare with artificial ones.
Grounded in MIT's generative tradition, it studies language as a property of the human mind: from the distinctions between description and prescription and between competence and performance, it covers morphology (morphemes and the hierarchical structure of words), phonetics and phonology (the IPA, phonemes and allophones, distinctive features, syllable structure and phonological rules), syntax (constituency, phrase structure, movement and wh-questions, binding) and semantics (compositionality, quantification and entailment), extending to language acquisition and the poverty-of-the-stimulus argument, variation and change, and sign languages. Problem sets ask for rules induced from real data in unfamiliar languages, training the habit of analysing language like a scientist, and giving a linguist's view to set beside the formal grammars of 18.404J and the language models of CS224N.
Taught by Michael Sandel through Socratic debate with a lecture hall of a thousand students: from hard cases (the runaway trolley, the shipwrecked crew driven to cannibalism), it sets out utilitarianism (Bentham and Mill), libertarianism (Nozick) and Locke on property and consent, then tests the limits of markets against military service and surrogacy, before turning to Kant's categorical imperative, Rawls's veil of ignorance and difference principle, and Aristotle's teleological justice, carrying each into affirmative action, patriotism and same-sex marriage. It trains the habit of naming the principle behind an intuition and following it through to its consequences.
Taught by Shelly Kagan, who argues a physicalist position openly rather than surveying views: from the case for and against an immaterial soul (Plato's Phaedo among the texts), it moves to personal identity (the body, soul and personality views, and the puzzles of duplication and fission) and what it is to be dead, then asks why death is bad, through the deprivation account and Epicurus' and Lucretius' arguments that it is not, whether immortality would be desirable, and whether fear of death is appropriate, closing with the rationality and morality of suicide. It shows how far careful argument can take a question most people answer by instinct alone.
Oxford's first-year introduction, as taught by Peter Millican in lectures now freely available: from Descartes' Meditations and the sceptical arguments that open modern philosophy, it covers Locke and Berkeley on perception and the external world, Hume's problem of induction and his account of causation, then free will and determinism (and whether the two are compatible), personal identity, the mind–body problem from dualism to functionalism, and the arguments for the existence of God, especially the design argument and Hume's critique of it. Reading the primary texts closely, it trains the analytic habit of reconstructing an argument premise by premise and finding where it fails.
Sept 2026 — Present
Alliance Française
Studying French in the classroom at the Alliance Française, the network of French language and cultural centres founded in Paris in 1883: grammar, vocabulary and pronunciation built up alongside listening, speaking, reading and writing, structured around the Common European Framework of Reference for Languages (CEFR).