Based on 12 visible X reactions from 18 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@BrianRoemmele AI is like a child, but it will grow up and learn as it does. Anyone pretending that AI will not steadily grow is like a parent with a five year-old child pretending that that child will never grow up to be an adult.
@BrianRoemmele The paper does make a strong case for preserving visual structure in training data Curated high-signal datasets feel like the obvious next step once you see the results
@BrianRoemmele No. It's still only a calculator they don't understand. Because they don't understand light.
@BrianRoemmele Obviously, the communist have a upper hand on the data capture
Unsupervised visual document pretraining consistently beats training on flat text.
@BrianRoemmele AI is like a child, but it will grow up and learn as it does. Anyone pretending that AI will not steadily grow is like a parent with a five year-old child pretending that that child will never grow up to be an adult.
@BrianRoemmele The paper does make a strong case for preserving visual structure in training data Curated high-signal datasets feel like the obvious next step once you see the results
@BrianRoemmele No. It's still only a calculator they don't understand. Because they don't understand light.
@BrianRoemmele brute force is just lazy design
@BrianRoemmele Never trust the Chinese.
DO YOU THINK THE WALL AI COMPANIES ARE HITTING WILL BE SOLVED WITH MORE COMPUTE? What only you know and a few now in China is: AGI and ASI will come about by this way of curating and training AI. The US is on the all you can eat junk food buffet of Reddit and Wikipedia data for AI training. And this paper proves it is the wrong path. This is not theory. This is the operating manual for the next phase of real, grounded, individual-scale intelligence amplification. The future belongs to those who kept the structure. I never threw it away.
Scalable Visual Pretraining for Language Intelligence https://x.com/_akhaliq/status/2076788514503463325/photo/1
paper: https://huggingface.co/papers/2607.09657
Based on 12 visible X reactions from 18 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@BrianRoemmele Never trust the Chinese.
DO YOU THINK THE WALL AI COMPANIES ARE HITTING WILL BE SOLVED WITH MORE COMPUTE? What only you know and a few now in China is: AGI and ASI will come about by this way of curating and training AI. The US is on the all you can eat junk food buffet of Reddit and Wikipedia data for AI training. And this paper proves it is the wrong path. This is not theory. This is the operating manual for the next phase of real, grounded, individual-scale intelligence amplification. The future belongs to those who kept the structure. I never threw it away.
Scalable Visual Pretraining for Language Intelligence https://x.com/_akhaliq/status/2076788514503463325/photo/1
paper: https://huggingface.co/papers/2607.09657