Review of Mind Children By Hans Moravec (1988)

Loading \[Contrib\]/a11y/accessibility-menu.js

I’ve had this book on my reading list for a while since this is the classic book everyone cites about predicting the singularity ahead of time and describing a ‘merging’ of AI and human minds into the future as a positive singularity. Since I had some time this afternoon, I downloaded and read it (hint: you can borrow it for free from internet archive [here](https://archive.org/details/mindchildren00hans/page/n11/mode/2up)). The book is actually pretty short (only about 160 real pages plus appendices plus bibliography) so can be read in one sitting and is a fairly easy read.

The actual premise of the book is surprisingly straightforward and I actually gained relatively little new knowledge from it. The primary steps are essentially to argue that AI technology is rapidly improving (and indeed at the time (late 1980s) it was and we were in the first AI summer), that Moore’s law predicts that we will reach approximate parity with human brain computation in the 2020s (as indeed happened!) and that this will usher in AGI (still tbd but now seems extremely plausible), and that in the end we will reach mind uploading technology and then our new silicon-based minds will continue to evolve and adapt and will drift away from their initial human forms until basically nothing distinguishes us from our AI descendants. Personally, I think this is pretty plausible and, to me, this largely represents the ‘good ending’ of the singularity, although YMMV.

In terms of ‘AI theology’, Moravec seems to have some vague thoughts about RSI but generally seems to envision a slow takeoff with humans and AI systems coexisting and our combined civilization growing and evolving into the stars. He forsees primarily darwinian processes operating on this mass of AIs and human uploads and that this will slowly shape the mind distribution towards ‘optimal minds’ causing uploaded humans to essentially forget their original human identity and function similarly to the pure AI systems. Moravec generally does not make detailed future predictions of e.g. gdp growth, or effect on current humans and broadly seems to assume that by and by large current systems continue at least long enough for humans to be uploaded into a silicon substrate and hence join the AIs on the capability frontier. He does not discuss AI alignment nor consider it seemingly a problem, at least in the book.

Interestingly, the Moore’s law projections are pretty accurate. He predicts specifically about 10 tera-ops and 100 terabits of compute and memory available for approximately $1000 in 2030, and this is also his estimate for the required compute and memory to match that of the human brain and hence reach AGI. He expects this to be possible on supercomputers in 2010. This is shockingly good estimate honestly given it was made nearly 40 years ago. For instance, a 4090 GPU (released in 2022) today has approximately 80 teraflops of fp16 performance and costs about $2000

... (truncated, 14 KB total)