On Thu, Nov 20, 2025, 1:37 AM Quan Tesla <[email protected]> wrote:

> Sure Matt. What investors, where? On AGI forum? I've watched you guys
> struggling to get funding for your devs for years. Maybe you were trying to
> scam and now your unscientific mind is projecting? It's possible. How
> stupid do you think investors are?
>

Most of the people who were active on the AGI forum years ago, that have
since left, were motivated by scientific curiosity, not profit. Ben
Goertzel had the most ambitious project, Webmind, then Novamente, then
OpenCog, now Hyperon, a hybrid neural and symbolic probabilistic logic
approach under development since the late 1990s. The software seemed always
under development and there was no knowledge base. Pei Wang developed NARS
(non axiomatic reasoning system) as his dissertation, but again no
knowledge base. Last I heard, he was doing vision research in China. Yan
King Ying was pursuing a mathematical and logical approach but never got
past the design phase. Peter Voss had a startup to develop AIGO, a
competitor to Amazon Alexa with better language comprehension, using the
forum to recruit developers before going into stealth mode.

I had some ideas to develop AGI, but limited my research to language
modeling using text prediction (measured by compression) after estimating
that a full solution to automating the global economy (the obvious
application) to be a decade of global GDP, or $1 quadrillion. The cost has
3 parts: software, hardware, and knowledge collection. Software is the
easiest. The human genome with 3 billion base pairs compresses to the same
size as 300 million lines of code at $100 per line, or $30 billion.

The hardware needed to run 8 billion human brain sized neural networks with
10^15 parameters each at 10 Hz is 10^26 GPU operations per second, costing
$1 per 10^17 operations or $30 quadrillion per year. Just the electricity
would be 1000 TW, or 55 times global production today. I assume Moore's law
will bring the cost down, eventually using nanotechnology to reduce energy
consumption because we can't make transistors smaller than atoms. It should
be under $1 quadrillion by 2045.

After this, the most expensive part of AGI will be collecting 10^17 bits of
human knowledge. Only 10^13 bits are available on the public internet, so
the rest has to be collected from human brains at 5-10 bits per second by
speech and writing at a cost of $5 per hour (the global average wage rate).
The human brain has a long term memory capacity of 10^9 bits (based on
Landauer's recall tests in the 1970s) and I  assume 99% overlap based on US
Labor Dept estimates that the cost of replacing an employee is 1% of
lifetime earnings. Wages are rising 2-3% per year, so I expect the cost in
2045 will be about $1 quadrillion and require that we live in a world of
constant surveillance.

All of this is detailed in my 2013 paper.
http://mattmahoney.net/costofai.pdf

So of course all of these projects failed. Software was the low hanging
fruit. Language modeling is the first step and nobody was doing it because
we didn't have the computing power or training data to do anything except
old fashioned hand coded structured knowledge from the 1980s. All of the AI
development is being done now by companies with trillion dollar market
caps. It seems obvious now, but not so much a decade ago.

-- Matt Mahoney, [email protected]

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T7ff992c51cca9e36-Md206bb2d7717f2e88c5506a4
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to