Daniel Okonkwo
Contributing editor, AI Worth Knowing
18 articles
How It Works
The training objective is almost embarrassingly simple, and the interesting argument in the field is about…
· 3 min read
History
Two research traditions spent decades disagreeing about whether counting could substitute for…
· 3 min read
Jargon
The number quoted alongside every model describes how many adjustable values it contains, which is a…
· 3 min read
Jargon
The term sounds like deduction and means something much more mundane, and the confusion it causes is worth…
· 3 min read
How It Works
The text a model is built from was gathered, filtered and arranged by people making choices, and almost…
· 4 min read
Limits & Risks
A conversation that appears to build on itself is being reassembled from scratch each time, and the various…
· 3 min read
Limits & Risks
Information from training is distributed across billions of parameters rather than stored in a record, so…
· 3 min read
History
The imitation game was a move to replace an unanswerable question with a testable one, and almost…
· 3 min read
History
For half a century progress was measured against board games, and each victory turned out to demonstrate…
· 3 min read
Jargon
Storing each parameter with less precision makes a model smaller and faster at some cost in fidelity, and…
· 3 min read
Jargon
The most common architecture for grounding a system in real material is a search step bolted to the front…
· 3 min read
How It Works
Combining independently trained predictors reliably improves accuracy, the reason is a piece of arithmetic…
· 4 min read
How It Works
The two dominant shapes of machine learning task differ in how error can be measured at all, and treating…
· 4 min read
Jargon
The term describes an architecture rather than an ability, several quite different arrangements are sold…
· 3 min read
In The World
Deciding what may stay up is done at a scale that forbids reading everything and requires context that no…
· 4 min read
Limits & Risks
Judging open-ended output automatically has become standard practice because the alternative is expensive…
· 4 min read
Limits & Risks
Reproducing a result requires the data, the code, the settings and a system that has not changed underneath…
· 4 min read
Jargon
The term has a precise technical history, a loose current usage and a marketing usage, and telling which…
· 3 min read
