[Blogs] Martin Ziqiao Ma

Blogs

Let me know what's on your mind :)

Compression is intelligence. -- Ilya Sutskever
Visual compression is spatial intelligence. -- Hao Tan
Language is compression from infinite meaning spaces onto finite, compositional symbols, so using language is by definition intelligent.
Mapping is understanding.
Grounding maps words to the world, so language grounding is language understanding.
Forget about storytelling, there are only three true problems in AI: machine learning, system engineering, data science.
Symbolic pattern understanding is a result of scaling language models.
Space-time understanding is a result of scaling visual models.
Intuitive physics understanding is a result of scaling world models.
Intuitive psychology understanding is a result of scaling human models.
We should get rid of dataset and call it data. Humans have no data boundary.
Benchmarks are sampling functions over the problem space.
Cognitive science is not about human intelligence, it's about intelligence.
AI for cognitive science is science: it studies the drivers of shared human-machine cognition and the mechanisms behind their differences.