Transformers represent one of the largest advances in mathematical modeling: they learn high-level predictive structure and generate continuations that conventional functions could never produce by hand.
Brian Greenforest separates that achievement from meaning itself. Prediction becomes meaningful through an agent that lives, chooses, acts, and participates in the physical world.
Understand the Machine at Kernel Level
Teaching Transformer kernels in GLSL and MLIR reveals exactly how embeddings, attention, matrix operations, normalization, and training objectives shape the output.
That mechanical clarity makes the achievement larger, not smaller. A trained function can synthesize an unexpected continuation across immense learned structure.
Connect Prediction to a Living World
Brian proposes quantum possibility or biological neurons as paths that could join predictive mathematics with sentience and meaningful choice.
Use Transformers to improve mathematical models, then bring those models into systems with physical state, consequences, and agency. The next frontier connects the greatest predictive engine with the conditions that make a prediction matter.
Read the Argument About Prediction and Meaning
The linked article sharpens the distinction between generating statistically likely language and carrying grounded human meaning.
LinkedIn status when archived: Visible to anyone on or off LinkedIn.
More actual than ever, after using it for so long, after diving into bits and pieces (thanks to Andrey and Georgiy), after teaching a student how to write and run Transformer kernels efficiently in GLSL and MLIR. I really understand what it's doing and how exacty it's trained and why it works.
We're delusional in stating that high-level predictive patterns are meaning. This thing needs quantum luck, or biological neurons attached to chips to become sentient and meaningful.
Use it to improve your mathematical models. Transformer is the biggest achievement in mathematics.
#ML #transformer #meaning #hype
https://lnkd.in/gqMTruRE
