> On Jun 13, 2015, at 6:54 AM, Anastasios Tsiolakidis <[email protected]> > wrote: > > Which brings me to the very sticky problem of "discovering computation", as > opposed to using it. What about having the AGI-brain play scientist with > assembly opcodes, stringing together C programs and figuring out what they do > if anything, and even perform such program generation or metaprogramming > without access to the formal specifications of compilers, and/or figuring out > workarounds for compiler bugs, the Pentium Bug etc.
Discovering how to design code requires some constraints to be tractable. And like with humans there is an economic bias toward code structure that can be analyzed inexpensively. However, at every step of the process you need to have some goal, even if it is just building a collection of efficient code motifs and transforms that might be useful later, and having a metric for “good enough”, or you will boil the ocean. The product of one of the most sophisticated examples of software learning to design software (“discovering computation”?): http://www.jandrewrogers.com/2015/05/27/metrohash/ Those hash functions are state-of-the-art and computer generated, but that is not the interesting part. The amazing part is that the software that generated those hash functions learned how to design hash functions almost from scratch on its own. And now that code (though it is more about the data model it created for itself) can design an endless stream of novel hash functions that are better than any human can design. The code design expertise lives in the software and its data, learned from real-world experience and able to adapt its designs to unique environmental constraints. It is not AGI, but it is prototypical evidence that computers can learn to design novel, goal-driven code in messy environments. The underlying mechanism and technique generalizes. In this context, the physical world is indistinguishable from any other kind of computer program and could be attacked the same way (and is). Most computer scientists fail badly at translating their paradigms to the physical world because the physical world is not a graph-like data model, but graph-like model assumptions are a cornerstone of mainstream computer science. Wrong tool for the job. Andrew ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
