index · 124 entries · filtering is local — nothing leaves the page

Every entry.

The whole site on one page. Type to filter titles, hooks and section names; without JavaScript this is simply the full index.

Machine Learning Foundations

What a neural network isweighted sums, bent, stacked deep
How models learnwrongness, measured, then walked downhill
The kinds of learningwho provides the answers — and at what price
Where data comes fromthe model is the dataset, rearranged
Overfitting & generalizationmemorizing the test is not passing it
Models grading modelsthe eval that scales, and the biases it brings
Learning by buildingreading gives you vocabulary; building gives you judgment
Keeping upa field that outruns its own textbooks
Calibration & uncertaintya number that means what it says
Evaluating modelsthe score is a proxy; act like it

Applied & Classical ML

Tables still rulethe models that run the world are not transformers
Features & leakagethe bug that makes your model look brilliant
Recommender systemsretrieve thousands, trim to hundreds, rank, show ten
Forecastingthe future is a different distribution, on purpose
Imbalance & anomaliesone in ten thousand, and the ten thousand keep changing
ML for sciencewhere prediction started producing discoveries

Large Language Models

Transformers & attentionevery token weighs everything before it
Tokens & tokenizationthe atoms of model perception
How LLMs are trainedfrom internet prediction to instruction following
Fine-tuning in practicea thousand good examples and an adapter
Models in other languagesthe same model is a different product in Thai
Context windows & memorythe model's whole world, measured in tokens
Hallucinationfluent, confident, and wrong by design
Reasoning modelsbuying accuracy with thinking time
One model, many sensesbolted-on sight became built-in perception
Beyond the dense transformernot every parameter has to work on every token
Scaling laws & small modelspredictable gains, surprising arrivals, shrinking costs

Prompt Engineering

Anatomy of a promptinstructions early, context middle, format last
Few-shot & chain of thoughtexamples teach, scratch tokens think
Structured outputschemas, tools, and grammars over hope
Context engineeringthe prompt grew up and became a budget
Designing for uncertaintythe interface is where reliability actually happens
Optimizing promptsstop tuning by hand, start searching
Why prompts failambiguous, buried, injected, and untested

AI Agents

What makes an agenta model in a loop with consequences
Tool use & function callingthe model asks, your code acts
MCP: a standard for toolsone protocol between models and everything
Computer usewhen the tool is the whole screen
Agent memoryeverything it knows tomorrow, you had to write down
Multi-agent systemsa team of models, at a price
Coding agentsthe application where the loop actually closed
Evaluating & containing agentstrust the state, not the summary

RAG & Embeddings

Embeddingsmeaning as a point in space
Vector searchfinding neighbors without checking everyone
The RAG pipelinelook it up before answering
Agentic retrievalstop retrieving once; let the model hunt
Evaluating retrievalmeasure the ceiling before you polish the room
RAG vs fine-tuningknowledge in context, behavior in weights

Generative Media

Diffusion modelsstart from static, subtract the noise
Image generation in practicedirecting a very literal artist
Video, voice & musiccoherent for seconds, drifting by minutes
3D & world modelsfrom generating pictures to generating places
Judging generated mediathere is no accuracy score for a picture
Provenance & detectionassume detection fails, prove the real

Computer Vision

How machines seefrom a grid of numbers to the idea of a cat
Detection & segmentationnot just what — where, how many, which pixels
Vision-language modelspictures and words, one shared space
Video understandingpictures were easy; time is the hard part
OCR & document AIturning paper into data you can trust
Vision in the wildwhy the demo dies in the parking lot

Speech & Audio

How machines hearsound as a picture of frequencies over time
Speech to texttranscription got shockingly good — measure where it isn't
Whose voice it hearsone average error rate, several different products
Text to speechvoices are now synthesized, cloned, and directed
The voice looplisten, think, speak — under a second, interruptible
Audio beyond speechwho spoke, who is that, and what just broke

AI Safety & Alignment

What alignment meansthe gap between what you optimized and what you meant
How models are alignedpreferences in, constitution over, refusals shaped
Bias & fairnessthe harm that arrives without an attacker
Jailbreaks & prompt injectionthe model does what the text says — whoever wrote it
Privacy & memorizationthe model remembers more than anyone intended
Evaluating safetya passed eval is a lower bound, not a verdict
Interpretabilityreading the weights instead of trusting the output
When it goes wrongthe incident response nobody drills until they need it
The long gameoverseeing systems smarter than the overseers

Reinforcement Learning

Learning from consequencesno answer key — just a world that scores you
Value functions & policy gradientstwo ways to turn experience into behavior
Reward design & specification gamingthe agent optimizes the number, never the intention
RL for language modelsthe quiet engine inside every frontier assistant
Learning from logsyou have data, not a simulator
When RL winssuperhuman at Go, struggling with a doorknob

Robotics & Embodied AI

Why robots are hardchess fell in 1997; the doorknob still holds
Sense, plan, actthe classical stack, and the learning that's eating it
Learning to movedemonstrations, simulators, and internet-scale priors
Manipulation & the humanoid bethands are the frontier; the body is a business case
Self-drivingrobotics' biggest bet, graded on public streets
Robots at workwhere embodied AI already earns wages — and why pilots stall

AI Infrastructure & MLOps

The hardware layermatrix multiplies, memory walls, and very expensive plumbing
Training at scaleone loss curve, ten thousand chips, months of not crashing
Serving & inferencethe meter runs per token — engineer accordingly
Local & edge inferencethe model on your own silicon, no meter running
What it costsunit economics before the invoice teaches them
What it costs the gridthe energy question, with the arithmetic actually done
The ML lifecyclethe model is 5% of the system; this is the other 95%
Securing the stackthe attack surface that isn't the model
LLMOpsprompts are config, evals are CI, tokens are COGS

Policy & Regulation

The regulatory problemgoverning a technology that ships faster than law can read it
The EU AI Actthe first comprehensive rulebook, tier by tier
The patchworkfifty rulebooks, three philosophies, one internet
Compliance in practicewhat the paperwork actually looks like from inside
AI and workthe question everyone asks, answered honestly
Open questionsthe debates that will define the next decade's rules

History of AI

The founding dreama summer workshop names a field and mortgages its credibility
Winters & thawstwice the field promised, twice the funding froze
The quiet decadesstatistics, GPUs, and a million labeled images
The deep-learning decadefrom AlexNet to ChatGPT in ten compounding years
The open-weights turnthe release that split the field into two ecosystems
Reading the presentusing seventy years of pattern to see this moment clearly

The Lab — interactive

Train a neural networkwatch a decision boundary get carved in real time
Walk the loss landscapedrop a ball, pick a step size, watch it learn or explode
Train a tokenizerwatch a vocabulary assemble itself out of characters
Drive an attention headqueries, keys, and a softmax you can feel
Sample from a language modelthe temperature knob, attached to an honest little model
Teach an agentdraw a maze, watch Q-learning adapt — then betray it
See your voiceyour microphone, rendered the way models hear
Steer the denoiserthe guidance knob, from ignoring the prompt to frying it
Grow an embedding spacetype text, watch words find their neighbors
Route to the expertswatch a mixture-of-experts eat itself
Does 0.9 mean 90%?a reliability diagram you can break and repair
Leak a modeladd a column, watch the score soar and the model get worse
Break an agentplant a sentence in a webpage and watch it reach the send button
Move the fairness thresholdthree definitions, one dial, no way to satisfy them all
Break a RAG pipelinethe answer is in the document — will it reach the model?
Hunt for neighborsexact retrieval vs the fast guess, recall meter running