On Mon, Dec 9, 2024 at 11:34 AM swkane <[email protected]> wrote:

> ... I think any scientifically relevant benchmark that is independent of
> the AGI system's architecture and hardware, and is general enough is worthy
> of looking at.
>

"Relevance" is relative to the value function.  "Human knowledge" is
relevant by virtue of the fact that humans value knowledge.  Science, in
fact, is not value neutral in this respect.  Only once you have selected
the dataset can science pretend to neutrality via the Algorithmic
Information Criterion for model selection because truth is operationally
definable only relative to the data selection.  The beauty of Wikipedia as
"the" data selection is that it encompasses not only the range of purported
"truth" (ie: the distillation of human observations) but, precisely because
it is biased, it also exposes the nature of the human values motivating the
bias.

... you could feed every problem on kaggle into an AGI system, just as an
> example.
>

If one is going to go beyond Wikipedia, a much better data selection for
"relevance" is the approach I take with Hume's Guillotine
<https://github.com/jabowery/HumesGuillotine>:

By-county (and by-country) longitudinal measures provided by established
entities such as the United States Federal Government and the United
Nations.

Charles Sinclair Smith -- the guy who financed the second neural network
summer of the 1980s from the System Development Foundation -- told me that
while he was establishing the Energy Information Administration under
President Carter, at least 90% of the expense of data curation was in data
cleaning.  Expanding the size of the corpus from 1GB Wikipedia to 1TB time
series tabular data invites an explosion of argument over the data
selection.

Nick Szabo's insightful phrase "argument surface" applies in spades to such
an explosive increase in selected data and resources dedicated to
losslessly compressing it.  The Algorithmic Information Criterion provides
a single, unimpeachable metric to minimize the "argument surface" over
*the* macrosocial
model in the presence of enormously motivated reasoners -- with the
residual argument surface being over the data selection criterion.

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T705ed500a1a7e589-Ma5ae633142515ede71d09b9d
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to