John Carmack’s ‘Different Path’ to Artificial General Intelligence

> So, about four years ago, I went on one of my week-long retreats, where I just take a computer and a stack of reference materials and I spend a week kind of reimplementing the fundamentals of the industry. And getting to the point where it’s like, ‘All right, I understand this well enough to have a serious conversation with a researcher about it.’ And I was pretty excited about getting to that level of understanding.

As much as I respect Carmack as a computer graphics expert, I really doubt his competence in machine learning. He doesn't have a single notable paper published. If he really thought that implementing gradient descent and basic stuff in a week long retreat gave him the chops to have serious conversations with AI researchers, he is really deluded.

Unless he can produce something that outdoes stable diffusion, chatgpt, alphago etc he should just hand over technical leadership of his start up to a leading AI researcher. Even Yann Le Cun at Meta is struggling to make any progress and is keeping himself busy by calling every other research labs output pedestrian. We cannot take any of Carmacks AGI predictions seriously, he simply lacks any expertise in the field.

Yep - he also failed with his lean, simple, first-principles approach with Armadillo Aerospace. The guy is proudly uncreative, and so he could have never come up with Scaled's design of a variable geometry rocket ship launched from a jet mothership.

> As much as I respect Carmack as a computer graphics expert, I really doubt his competence in machine learning. He doesn't have a single notable paper published.

Publishing papers is the way the academic/scientific world measures notability and/or competence. It's not the way the engineering world that Carmack comes from measures it. They measure it by building. But you're right, we kind of have to just trust that he has the expertise he says he does by his statements since he has not built any modern AI programs (that I know of at least).

> If he really thought that implementing gradient descent and basic stuff in a week long retreat gave him the chops to have serious conversations with AI researchers, he is really deluded.

This is not an accurate account of how he said he developed his knowledge base. Just how he got started so he could have conversations. He said that he spent a retreat learning the basics and then later in the interview he said he took the time to understand the 40 most essential papers in the field as related to him by a well known researcher. He has since largely put the last 4 years of his professional life into this. While we have no proof of his knowledge, given his intelligence and high competence in computer programming and math, I have no doubt that if he did put in the work he could achieve an understanding equivalent to that of your average AI researcher.

That said, of course it makes sense to be skeptical.

> So I asked Ilya, their chief scientist, for a reading list

Just curious, if this reading list is available somewhere.

'Kill Your Darlings' is one of the most fundamental skills in anything creative. It is a hurdle that, once taken, allows you to ascend above a plateau of mediocrity.