About
Next Token records how AI got here, one piece a week. The name comes from how a language model works: it picks the next piece of a sentence — the next token.
Two kinds of writing
An event checked against a source link — exactly one per record. Numbers a company reports about itself are marked as company-reported. If the source can't be checked, it doesn't go in.
My own write-up after reading several sources, without source links. Notes carry a yellow tag and can be filtered on the timeline.
How records are checked
- Every Monday morning I go through a fixed list of places — paper repositories, lab and company blogs, government records — and collect new events.
- For each event I open the primary source and match the date, actor and content. If there's only a secondary report, the record is labelled “Press” or left out.
- Claims that may change later (a bill awaiting signature, say) are noted and checked again. If they change, the record is corrected.
- I read the week's records and write three threads and one question.
Primary sources first; secondary reports are labelled “Press”. 87% of the 1,295 records rest on a primary source.
Source labels
Each record shows what kind of source it rests on.
A paper in a repository or journal
An announcement by the company, lab or organisation itself
A record from a government, parliament or international body
A secondary report — the primary source couldn't be found
Records what someone said, not a fact
Records a trend observed in a community
Learn
Five-Minute AI Lessons open one a day for 60 days, for students aged 13–18: one experiment, three quiz questions graded on the spot, and slides as a PDF. The AI History Module retells the same record at three levels, and the Part Reviews sum up each part at three depths.
Follow along
New records and each day's lesson are in the RSS feed. The whole record can be downloaded as a dataset.
Who makes this
factnotes — an automation and AI engineer at a semiconductor plant in Korea. Blog: blog.naver.com/factnotes (in Korean).
Using the content
Text and data are CC BY-NC 4.0: free for non-commercial use with attribution. The dataset is available as CSV and JSON.