Re: AGI investment (Re: What are "AGI-first'ers" expecting AGI will teach us about FAI?)

From: Matt Mahoney (
Date: Sun Apr 13 2008 - 17:49:03 MDT

--- Thomas McCabe <> wrote:

> On Sun, Apr 13, 2008 at 4:44 PM, Matt Mahoney <> wrote:
> > --- Nick Tarleton <> wrote:
> >
> > > Global economic output over the next 30 years, at constant growth,
> > > will be $4.6 quadrillion (my calculation from CIA World Factbook
> > > numbers). That nearly a quarter of the world economy will be invested
> > > in AGI over that time is hideously implausible.
> >
> > If AGI enables machines to do all the work that humans would otherwise be
> paid
> > to do, then how much would it be worth?
> >
> >
> >
> > -- Matt Mahoney,
> >
> These sorts of calculations require a much higher level of rationality
> than the general public usually possesses. Even on SL4, how many
> people donate almost a quarter of their income to AGI development?
> Where did you get the $1 quadrillion figure in the first place?

My estimate is based on the size of the world economy (US $66 trillion per
year and growing at 5%) assuming that a significant fraction (more than 1/4,
and approaching 100%) will be invested in AGI over the next 30 years. Most of
this will be indirect, in the same way that the development of the technology
that enabled the internet (TCP/IP, HTML, HTTP, etc) was a tiny fraction of the
overall value of the internet (tens of trillions). Most of the value is in
content and equipment (Google, Yahoo, various ISPs, your computer, etc).

I expect the development costs of AI algorithms to be a tiny fraction of the
value of AI. Most of the value will be in the form of computational resources
and knowledge. Most of the work will be indirect. For example, viewing a
porn site would help train an AI to generate better images.

One can argue that the cost of AI could be a tiny fraction of its value. I
argue that given a choice we would prefer to have AI sooner at higher cost.

My figure of 30 years is just a guess how long it will take to achieve
recursive self improvement. One could use Moore's Law to project when an
artificial brain will cost less than hiring a natural one if we knew how much
computing power we need. But we don't really know this. I am estimating the
cost of development prior to achieving RSI because once we do, a singularity
will quickly follow.

> If someone gave you $1 quadrillion tomorrow, what would you spend it on?

I wouldn't. AGI will emerge from the internet as people pursue their own
goals. Paying people to do anything else would just be counterproductive. A
centralized effort would kill it.

-- Matt Mahoney,

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:02 MDT