Archive
Everything we have published, newest first — 100 articles.
August 2026
Learning, in a neural network, means adjusting numbers until the errors…
Training a model and running one are two different kinds of work
A token is not a word, and several odd behaviours follow from that
Predicting the next fragment is a narrower job than the results suggest
Attention is a weighted lookup and the name oversells it considerably
Ranking systems were the machine learning most people met first
Machine translation cleared a threshold and left its hardest problems behind
A model trained on past decisions will reproduce the reasoning behind them
Recognition systems perform unevenly and the unevenness is not random
When a first draft costs almost nothing the bottleneck moves to checking it
Being confidently wrong is the default behaviour, not a malfunction
What a benchmark score measures and what it quietly does not
Correlation carries these systems a long way and then stops at the edges
July 2026
A system’s account of its own reasoning is a reconstruction, not a record
Training data has an end date and the model cannot feel the edge of it
Symbolic AI was not a mistake, it was answering a different question
The perceptron controversy that shelved neural networks for a decade
An AI winter is a funding event before it is a scientific one
Expert systems made the field money and then became its cautionary tale
The statistical approach won by outgrowing the argument rather than…
Parameters are not knowledge and the count is not a score
An embedding is a position, and the geometry is what does the work
June 2026
Fine-tuning changes a habit far more readily than it adds a fact
Overfitting is memorisation wearing the costume of success
Inference is a borrowed word for the part that runs every single time
A training corpus is a collection with a history, not a sample of language
Past a certain scale more text stops helping and cleaner text starts
Sound and images reached the same toolkit by completely different roads
A guardrail is at least four different mechanisms sharing one word
Training on human preference optimises for what a rater can notice
Open weights are not open source, and the difference decides who can check…
These systems have a physical footprint and it is built from concrete and…
The cost of training decides who gets to build, and that has already…
In the sciences these methods succeeded where the answer could be checked
May 2026
Proving a recording is genuine is harder than producing a convincing fake
These systems have no memory, and every workaround is something else…
Chain enough reliable steps together and the chain becomes unreliable
A model cannot be made to forget, and that breaks how deletion is supposed…
A deployed model decays because the world moves and the parameters do not
The human reviewing the output is doing a harder job than the design assumes
Turing proposed a test to dodge a question the field then spent decades…
The field acquired its name at a summer workshop, and the name has been a…
A simple pattern-matching program showed people would supply the…
Games became the field’s yardstick because they could be scored, not…
April 2026
The hardware that made this possible was designed for drawing pictures
Latency and throughput are different clocks and improving one usually costs…
Quantisation is lossy compression applied to a model’s numbers
Distillation trains a small model to imitate a larger one’s behaviour
A context window is a budget, and everything competes for the same space
Retrieval means the model was handed documents, not that it looked anything…
March 2026
A model is a file of numbers, and the file is not the whole system
Ask the same question twice and the answers differ, for two separate reasons
Averaging several models beats polishing any one of them, up to a point
Picking from a fixed list and composing from nothing are different problems
Forecasting a sequence of numbers is an older discipline with different…
Somebody wrote down the answers first, and that turned out to be a global…
Where the computation happens decides what the system is allowed to be
Training on your data means several different things, and only one is…
Physical tasks lagged because the world does not arrive pre-labelled
A regulated setting asks questions a demonstration never has to answer
February 2026
Accurate and useful are separate properties and they come apart often
A system that can be fooled on purpose is a different problem from one that…
Finding the rare case is hard for a reason arithmetic makes unavoidable
Capability is a claim about something nobody can observe directly
Interpretability has produced real findings and not the thing people wanted
January 2026
Cybernetics asked most of the questions first and then lost the name
Machine translation was announced early, defunded in the middle, and…
A shared dataset organised the field more effectively than any theory did
The training algorithm everyone uses was discovered more than once
Anything that starts working stops being called artificial intelligence
Zero-shot and few-shot describe the input, not an ability of the model
December 2025
Alignment names a goal and quietly skips the question of whose
A batch is a group processed together, and its size is a real decision
Multimodal means one system takes more than one kind of input
Foundation model was a naming decision, and it was argued about immediately
Training a model on another model’s output is a real technique with a known…
An image generator works by learning to remove noise it was taught to add
A recommender has to guess the question, and that decides everything about…
Learning from consequences is a different arrangement from learning from…
November 2025
A small model built for one job can beat a large general one, and size is…
Moderating a platform is a judgement problem that automation can only…
A dataset licence is a document with terms, and most people never read one
Answering the question instead of listing the links unpicks the bargain the…
Who owns a generated work is unsettled, and the reasons are older than the…
Automating a task is not the same as automating a job, and the difference…
Using one model to grade another is convenient, and it inherits every bias…
People attribute a mind to these systems reliably, and the design does not…
A model can be corrupted during training in a way that shows up only on a…
A great many published results here cannot be run again by anybody else
The expertise needed to supervise a system is built by doing the work it…
A network for reading handwriting worked commercially long before anybody…
October 2025
The field published everything, and then some of it stopped
Recognising speech took fifty years and three complete changes of method
Early robots tried to think before moving, and the reaction against that…
Two communities worry about this technology and they have been arguing for…
A hyperparameter is a setting somebody chose, and the choosing is where…
Agent is a word from one field being used to sell something from another
Inductive bias is what a model assumes before it has seen anything
Ground truth is a confident name for something people had to agree on
A scaling law is a curve fitted to past runs, not a promise about the next…