On Mon, Aug 12, 2024 at 7:29 PM YKY (Yan King Yin, 甄景贤)
<[email protected]> wrote:
> Attached is my presentation PPT with some new materials not in the submitted 
> paper.

I wonder if you had time to answer your question at the end of the
presentation. How does this help AGI?

We have the algorithm mostly figured out. A fully connected neural
network can simulate an arbitrary number of layers to learn
arbitrarily complex features as well as an attention mechanism through
mutual inhibition. We established that text prediction is sufficient
to pass the Turing test (as well as any possible test for
consciousness). The largest language models have 10^12 parameters
trained on 10^13 tokens using 10^26 operations at a cost of 10^17 per
dollar, on the order of $1 billion.

Now it is an engineering problem. We can't make transistors much
smaller (2 nm = 18 silicon atoms) to reduce power consumption, but we
can optimize the hardware for sparse, low precision vector operations.
We can reduce training costs to one operation per parameter per token
using one shot learning, like the brain and most text compressors
already do. Ultimately it will take nanotechnology, moving atoms
instead of electrons, to reduce power consumption for a human brain
sized neural network from 1 MW to 20 watts. That technology is still
decades away.

Meanwhile we have the much larger problem of collecting the training
data needed to automate human labor, which is now up to $110 trillion
and rising 5% per year. Yet, ChatGPT has been out for almost 2 years
without the slightest increase in unemployment. The problem is that
you need to collect 10^17 bits of human knowledge to do all the work
that people do, and we only have 10^14 bits (15 TB) of text available
on the public internet and most of it is already used to train LLMs.
AI will profoundly change the world. But when I look at the ads for
Meta, Gemini, and Copilot, I think, really? Is this the best we can do
with AI? Help kids write fan letters? These are basically toys.
Collecting all the knowledge you need to do your job that isn't
written down will cost on the order of $100 trillion at the global
average wage rate of $5 per hour. Figure one year of human time to
train your replacement. It doesn't matter if it is carbon or silicon.
You can only output 3 tokens per second.

-- 
-- Matt Mahoney, [email protected]

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T04cca5b54df55d05-M802d3df156a65ab8ae5b6fc4
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to