Xingchen AI · AI explainers, one a day
One episode a day, turning today's AI news into plain words (text edition). The latest 2 are free; all episodes and a private feed are for members.
Agents Are Getting a Memory Layer: From One Chat to Long-Term Collaboration
The industry spent a year on model capability. Now it turns out what blocks agents is that they forget.
Latest 30 episodes
A wave of 2-billion-parameter models now handles vision, speech and tool calling, pushing inference cost to a new level.
Formal verification tools broke a 10,000-page human proof into checkable steps — closing a gap and advancing the result.
Front-line staff can now dispatch agents directly, making permissions and data boundaries the new management problem.
A text analysis study finds reading fatigue from AI output is real and tied to six measurable sentence patterns.
From linguistic analysis to acoustic modeling to waveform generation — three clearly separated stages behind every spoken sentence.
Outside big tech, most industries have yet to see financial returns matching their AI spending.
Using shadow previews, replayable execution and retry refactoring, a researcher argues planning must be made visible.