Pei Guo (@amazingguo) 's Twitter Profile
Pei Guo

@amazingguo

ID: 434789918

calendar_today12-12-2011 09:38:50

133 Tweet

41 Followers

238 Following

Pei Guo (@amazingguo) 's Twitter Profile Photo

On different machines, the following code returns slightly different results, even with the same torch library. It turns out that the low-level math acceleration library is different for different CPUs, and that alone can cause tiny difference that accumulates.

On different machines, the following code returns slightly different results, even with the same torch library.

It turns out that the low-level math acceleration library is different for different CPUs, and that alone can cause tiny difference that accumulates.
Pei Guo (@amazingguo) 's Twitter Profile Photo

It took decades—and many of us—to (re)discover what Hinton described as coarse coding 40 years ago. He saw it clearly through simple 2D dot examples, not noisy, overcomplicated methods. He shaped the future by seeing it before anyone else.

It took decades—and many of us—to (re)discover what Hinton described as coarse coding 40 years ago. He saw it clearly through simple 2D dot examples, not noisy, overcomplicated methods. He shaped the future by seeing it before anyone else.
Pei Guo (@amazingguo) 's Twitter Profile Photo

I wrote a blog post that recreate the invention process of VAE, Diffusion, and Flow Matching in a science fiction style. It doesn't have difficult mathematics, only intuition, logic, and reasonable imagination. peiguo.me/posts/autoenco…

Pei Guo (@amazingguo) 's Twitter Profile Photo

I'm reading On Intelligence by Jeff Hawkins, and I have this mind-numbing feeling, like he somehow traveled back from the present to write this book.

Pei Guo (@amazingguo) 's Twitter Profile Photo

I wish I had watched Chelsea Finn's 'how to do research' talk a decade ago instead of focusing on so many online resource on 'how to publish a paper'. youtu.be/FacJ_1tTSx4?li…

Pei Guo (@amazingguo) 's Twitter Profile Photo

Choosing whom to follow is an important part of one's research taste, if not the most important. If I can choose again, I would learn neural nets from Hinton himself 10 years ago instead of wasting my time studying from so much misleading online materials.

Pei Guo (@amazingguo) 's Twitter Profile Photo

I've known E=mc^2 for long but never read Einstein's original paper until recently. It only has 3 pages and deduces the equation by looking at an imaginary body emitting lights from two coordinate systems. Simple math, great intuition. sites.pitt.edu/~jdnorton/teac…

Pei Guo (@amazingguo) 's Twitter Profile Photo

Hopfield networks laid the groundwork for AI research, representing a rare and fruitful marriage between neuroscience and statistical physics.

Pei Guo (@amazingguo) 's Twitter Profile Photo

Most AI researchers have to be both theorist and experimentalist—coming up with ideas and implementing them. Now that I can delegate the coding to Claude, I get to be a pure theorist for the first time.

Pei Guo (@amazingguo) 's Twitter Profile Photo

Many times when I found something interesting, like Johnson–Lindenstrauss lemma, I can find a nice blog from Su Jianlin. spaces.ac.cn/archives/8679

Pei Guo (@amazingguo) 's Twitter Profile Photo

Both Demis Hassabis and Dario Amodei studied neuroscience during their PhDs — that fact alone seems to carry some significance.

Pei Guo (@amazingguo) 's Twitter Profile Photo

Only after this video have I realized computer vision has learned so much from human vision, and there's no reason to believe this trend won't continue. youtube.com/watch?v=Gv6Edl…

Pei Guo (@amazingguo) 's Twitter Profile Photo

New blogpost exploring interpretable features from random networks. In it I explore the cause of this phenomenon: is it a property of the network or does it reveal some limitation of the interpreting tools? The answer is both. peiguo.me/posts/random-e…