Hi Matt, Doesn't the "predictor" actually contains trained models as well?
-- ----------------------------- At 2018-09-10 23:25:54, "Matt Mahoney via AGI" <agi@agi.topicbox.com> wrote: >On Mon, Sep 10, 2018 at 8:10 AM <johnr...@polyplexic.com> wrote: >> Why is there no single general compression algorithm? Same reason as general >> intelligence, thus, multi-agent, thus inter agent communication, thus >> protocol, and thus consciousness. > >Legg proved that there are no simple, general theories of prediction, >and therefore no simple but powerful learners (or compression >algorithms). Suppose you have a simple algorithm that can predict any >computable infinite sequence of symbols after only a finite number of >mistakes. Then I can create a simple sequence that your program can't >learn. My program runs your program and outputs a different symbol at >each step. You can read his paper here: >https://arxiv.org/abs/cs/0606070 > >This has been the biggest pitfall of AGI projects. You make fast >progress initially on the easy problems, thinking the solution is in >sight, and then get stuck on the hard ones. > >> Doesn't Gödel Incompleteness imply "magic" is needed? > > No, it (and Legg's generalizations) implies that a lot of software and > hardware is required and you can forget about shortcuts like universal > learners sucking data off the internet. You can also forget about self > improving software (violates information theory), quantum computing > (neural computation is not unitary), or consciousness (an illusion > that evolved so you would fear death). > > How much software and hardware? You were born with half of what you > know as an adult, about 10^9 bits each. That's roughly the information > content of your DNA, coincidentally about the same as your long term > memory capacity according to Landauer. (see > https://www.cs.colorado.edu/~mozer/Teaching/syllabi/7782/readings/Landauer1986.pdf > ). All this debate about nurture vs nature is because for most traits, > it's both. > > The hard coded (nature) part of your AGI is about 300M lines of code, > doable for a big company for $30 billion but probably not by you > working alone. And then you still need a 10 petaflop computer to run > it on, or several billion times that to automate all human labor > globally like you promised your simple universal learner would do by > next year. > > Or maybe you could automate the software development. It's happened > once, right? All it took was 10^48 DNA base copy operations on 10^37 > bases over 3.5 billion years on planet sized hardware that uses one > billionth as much energy per operation as transistors. > > I believe AGI will happen because it's worth $1 quadrillion to > automate labor and the technology trend is clear. We have better way > to write code than evolution and we can develop more energy efficient > computers by moving atoms instead of electrons. It's not magic. It's > engineering. > > -- > -- Matt Mahoney, mattmahone...@gmail.com ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T9c94dabb0436859d-Ma1dd71b4e4781df28ef394d6 Delivery options: https://agi.topicbox.com/groups/agi/subscription
