Transformers and LLMs
Attention as geometry: queries, keys and values moving through embedding space, scaled up into language models.
- Lessons
- 8
- Study time
- 5.1 h
- Access
- First lesson free
Syllabus
- 4.0
Inside an LLM: from one neuron to the next word
Follow a real, tiny GPT from a single neuron to an MLP, an attention head and a stacked decoder, until it turns a prompt into next-word probabilities and generates text one token at a time.
45 minFreeLive 3D lab and Python challengeintermediate- Trace a token from a single neuron through an MLP to an attention head with queries, keys, values, a causal mask and a softmax
- Follow the residual stream through multi-head blocks stacked into a decoder
- Turn the final hidden state into next-token probabilities and run the generation loop
- 4.1
Self-attention geometry and scaled dot-product in 3D
Project tokens into query, key and value spaces, fire attention beams between them, and see exactly why we divide by √dₖ.
42 minFree previewLive 3D lab and Python challengeintermediate- Compute Q, K, V projections and their tensor shapes
- Explain why unscaled dot products saturate the softmax
- Read attention weights as a soft lookup over keys
- 4.2
Tokenisation and embedding spaces
Byte-pair encoding from scratch, and how embedding tables turn token ids into directions with meaning.
30 minProLive 3D lab and Python challengeintermediate- Train a BPE tokenizer on a small corpus
- Measure similarity in embedding space
- Add positional information with sinusoidal and rotary encodings
- 4.3
Multi-head attention and the transformer block
Heads as parallel subspaces, residual streams, pre-norm and the MLP that stores most of a model's facts.
38 minProLive 3D lab and Python challengeadvanced- Split and merge heads with reshapes and transposes
- Trace the residual stream through a block
- Count parameters and FLOPs of a transformer layer
- 4.4
Pretraining, sampling and fine-tuning LLMs
Next-token loss at scale, temperature and nucleus sampling, and parameter-efficient fine-tuning with LoRA.
40 minProLive 3D lab and Python challengeintermediate- Implement temperature, top-k and top-p sampling
- Explain scaling laws at a high level
- Fine-tune with low-rank adapters
- 4.5
Embeddings, vector search and retrieval-augmented generation
Embed a corpus, fire a query into it, and follow nearest-neighbour search through an HNSW graph before the results reach the prompt.
38 minFree previewLive 3D lab and Python challengeintermediate- Retrieve with cosine similarity and explain the cost of exact search
- Trace an approximate search through an HNSW index and its recall trade-off
- Design a RAG pipeline: chunking, re-ranking with MMR, and retrieval evaluation
- 4.6
Alignment: reward models, RLHF and DPO
From instruction tuning to preference optimisation: reward models, the KL leash of RLHF, and the closed-form shortcut DPO takes.
36 minProLive 3D lab and Python challengeadvanced- Train a reward model from pairwise preferences with the Bradley-Terry loss
- Explain the KL-regularised RLHF objective and reward hacking
- Derive the DPO loss and compare it with PPO-based RLHF
- 4.7
LLM agents, tool use and evaluation
Reason, act, observe: the agent loop, function calling, planning budgets, and how to evaluate a system that takes actions.
34 minProLive 3D lab and Python challengeintermediate- Implement a ReAct-style loop with tool calls and observations
- Bound cost and failure with step budgets, retries and guardrails
- Evaluate agents with task success, trajectory checks and LLM-as-a-judge