Series
The AI Field Guide
Large language models are the loudest part of AI, not the only part. This series covers the rest of the field -- diffusion, encoder-only and encoder-decoder transformers, classical NLP, search and planning, logic, constraints, probabilistic reasoning -- so that picking the right tool for a job stops being a guessing game. Part of Under the Hood.
Under the Hood · The AI Field Guide
How LLMs Actually Work
Tokens, transformers, attention, and the training pipeline: what large language models actually do when they 'predict the next token', why they hallucinate, and why they're so good at code.
Read article
Under the Hood · The AI Field Guide
To LLMs... and Beyond!
LLMs are one corner of a much larger field. Diffusion models, reasoning models, multimodal systems, open-weight vs closed -- what they are, how they differ, and how to choose.
Read article
Under the Hood · The AI Field Guide
The Other Transformers
BERT and T5 are transformers too, but they aren't trying to be ChatGPT. They're trying to be the boring layer underneath, classifiers, embeddings, structured transformations, and they're often a better answer than an LLM.
Read article
Under the Hood · The AI Field Guide
The Reranker You Didn't Know You Needed
RAG explanations stop at 'embed the query, look up the nearest documents, hand them to the LLM.' That's the demo. In production, there's a second pass between the lookup and the LLM, and it's the one that actually makes retrieval work.
Read article
Under the Hood · The AI Field Guide
After the Transformer
Transformers have ruled language modelling for nearly a decade. They have a known weakness, and several research lines are trying to replace them. Mamba, RWKV, RetNet, Hyena, diffusion-for-text, what they are, what they fix, and which ones are likely to matter.
Read article
Under the Hood · The AI Field Guide
Before the Transformer
n-grams. HMMs. CRFs. The language models and sequence taggers that ran the internet before deep learning, and that quietly still do, in autocomplete, spam filters, biomedical NER, speech recognition. What they are, why they still ship, and when they're the correct answer.
Read article
Under the Hood · The AI Field Guide
The Boring Baseline That Wins
TF-IDF, logistic regression, naive Bayes, k-means, LDA. The fifty lines of scikit-learn that beat your fancy model on the small problem you actually have. Why these baselines still win, and why the correct starting point in 2026 is often the same as it was in 2006.
Read article
Under the Hood · The AI Field Guide
Rules, Grammars, and Regex
Sometimes the correct answer to 'what model should I use?' is no model at all. Hand-written rules, regular expressions, finite-state transducers. They're deterministic, auditable, free at inference, and frequently the correct tool for the job.
Read article
Under the Hood · The AI Field Guide
Search and Planning
Most of what people called 'AI' before deep learning was search. A* finding the route, alpha-beta playing the chess move, STRIPS sequencing the plan. The algorithms that run your map app, your build system, your warehouse robot, and your game opponent.
Read article
Under the Hood · The AI Field Guide
Knowledge, Logic, and Constraints
Reasoning from facts and rules, the way the AI textbooks of the 1980s thought we'd build minds. Propositional and first-order logic. Knowledge bases. Datalog. Modern descendants. SAT solvers, SMT solvers, and the production rule engines running insurance and banking. When if-this-then-that wins.
Read article
Under the Hood · The AI Field Guide
Bayesian Reasoning
Probability is the third leg of classical AI. Bayesian networks, Kalman filters, particle filters, multi-armed bandits. The mathematics that fuses sensor noise into a position estimate, that decides whether to show a user this ad or that one, that diagnoses faults from intermittent symptoms.
Read article
Under the Hood · The AI Field Guide
When Not to Use an LLM
Nine posts in, the question this series has been circling: when is an LLM the wrong tool? A field-guide-style decision walk through the alternatives, encoder models, classical NLP, rules, search, logic, probability, with the symptoms that point to each, and the symptoms that point back to LLMs.
Read articleEvolutionary Computation
There is an antenna in orbit that no human designed. A computer bred it: thousands of candidate shapes, the best survivors crossed and mutated each generation, until a bent little wire that looks like nothing on Earth beat every hand-drawn rival. That is evolutionary computation, the branch of AI that optimises by selection instead of gradients.
Coming soon