On Wed, Mar 21, 2007 at 12:00:09PM -0700, rooftop8000 wrote: > Trying to make a seed AI is the same as hoping to win the lottery.
Winning the lottery is an unbiased stochastical process. Darwinian co-evolution is a highly biased stochastical process. Seeds are one-way hashes (morphogenetic code expands the small seed into a structure which is appropriately positioned (environment-shaped) to extract knowledge from the environment (aka doting parents). Such seeds can be relatively tiny, see fertilized human eggs (the womb does not seem to contribute noticeable amounts of complexity). Hence they contain far less complexity than an adult, winning which by stochastical process, unbiased or otherwise, takes terrible odds. > You're just hoping you only have to do one thing so you can forget about > all the other stuff that is required. No. I don't think that other stuff required can be done. This is the same reason I don't subscribe to SENS. I thought this was unlikely when I was a 15 year old, and I still think it's unlikely as a 40 year old. > And if i could pick things that wouldn't be needed in a seed AI, it would be > real-world vision and motor skills. I agree that understanding movement and Learning from the environment takes navigation in and active manipulation of the environment. The G in AGI doesn't stand for domain-specific. > diagrams and figures is essential to thought, but why would a computer > program > need to do recognize > a picture of a chair or a picture of a horse or be able > to track a flying bird in the sky? I don't think that's required > for most problems. I also don't see how you get to all other thoughts from > there? > (Not that it can't be useful to have in your system..) > > >Not necessarily the other way round. I understand some > > consider self-modification a specific problem domain, so a system capable > > of targeted self-inspection and self-modification can self-modify itself > > adaptively to a given task, any given task. I think there is absolutely > > no evidence this is doable, and in fact there is some evidence this is > > a Damn Hard problem. > > I agree. you can only do some minor self-modification if you don't fully > understand your inner workings/code. I have reasons to suspect that a system can't understand its inner workings well enough to do radical tweaks. Well, we can (in theory) mushroom our cortex by a minimal genetic tweak. That's a trivial modification, which doesn't reengineer the microarchitecture. Live brain surgery on self or a single copy doesn't strike me as a particularly robust approach. Add a population of copies, and a voting selection or an external unbiased evaluator, and you're already in Darwin/Lamarck country. > I still think most of this AGI will have to coded by hand, and it will I don't think this is doable by mere humans. This is a few orders of magnitude below of what the maximum complexity ceiling is (tools only take you that far). If AI is numerics, Fortran+MPI would be enough. C would be arguably less painful. If AI is not numerics, you're equally screwed, whether this is Lisp, Erlang, Ruby or Fortran. > be a lot of software engineering and not the romantic seed AI or minimal To clarify, I'm only interested in ~human equivalent general AI, and only in co-evolution from a reasonable seed pool in a superrealtime virtual environment heavily skewed towards problem-solution as fitness function as a design principle. The only reason for this is that it looks as if all other approaches are sterile. You're of course quite welcome to prove me wrong by delivering a working product. > subset of 10 perfect algorithms... Seems like people don't seem to > want to put in all the energy and keep looking for a quick solution My estimate is several % of yearly GNP for several decades for a likely success by above design mechanism. If you call that a quick solution, many will disagree. -- Eugen* Leitl <a href="http://leitl.org";>leitl</a> http://leitl.org ______________________________________________________________ ICBM: 48.07100, 11.36820 http://www.ativel.com 8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=303
