Y'all, Taking off my "here's the way I see it" hat and putting on my day-job cost estimation hat, let's start by establishing a constructive cost estimation process. Good code-size and cost estimates are best arrived at by 3 different methods, each "tweaked" to arrive at the same number. If the number is wrong, then one of the 3 different methods invariably breaks, e.g. subroutines must be less than one statement long, etc. Note that in addition to estimating cost, this process also identifies good and bad approaches, and puts a stake in the heart of misdirected projects. I did one of these for an ADA compiler in 1981, and advised them to re-target for the HAL-9000 computer because they were 20 years from completion. They reorganized and completed 6 months behind their original schedule.
Note also that cost ceilings can be calculated. For example, it is obviously impossible to spend a quadrillion dollars on any technical pursuit, because there simply aren't enough qualified technologists to take the money. You can't even spend a trillion dollars. Further, there is a non-linear phenomenon where beyond a certain point, a little more complexity costs a LOT more, first because of the "quadratic phenomenon where the interactions between subsections eventually swamps all other factors, and then a point is reached where no one understands the entire thing, beyond which little real progress can be achieved. Experience at major manufacturers (building the most expensive yet still successful softwares) suggests that this ceiling is on the order of a billion dollars. Of course, MUCH more money can be spent on research to get ready to spend your billion dollars on writing software, which is one of my own favorite topics. Right now, I would be doing research, building machines to scan out neural constructions, working on a possible "theory of everything" that seems to be there but undiscovered. etc., rather than writing AGI software, but that is my own opinion and NOT the topic of this posting. The three methods for estimating projects are typically chosen from lists like: 1. Take another comparable project, and adjust the cost of it as seems appropriate. Perhaps the entire IBM-360/370 OS, compilers, etc., might be a good starting point. 2. Estimate the number of modules, and multiply that by the cost of design, coding, debugging, testing, etc., and then multiply that by the estimated number of project restarts. Ben, you would probably be the best source for these guesstimates. 3. Just take a WAG (Wild Ass Guess) based on experience and gut feeling. 4. Come at it from the problem complexity side, estimate the number of heuristics needed, multiply that by the estimated effort to create each one, and add in a large margin to create the system to process those heuristics. 5. Going for an n log n level of effort, describe each of the levels of abstraction, and estimate each as though it were a separate project. Sure you may change your mind later as to what the levels are, but this should give you a reasonable estimate UNLESS you end up adding a bunch more levels. Remember to add a "safety margin" extra level or two. 6. Don't forget to look for those "little" missing pieces that could be VERY expensive to try to "fill in" at the last minute. Pieces like a fast enough processor, new theory, etc. These too must be guesstimated. My own "gut feeling" (which many here routinely reject) is that either it can be build for a <$1 billion, or it couldn't (presently) even be built for a petabuck, any more than all the women of the earth could produce a baby in one month. Continuing from the prior thread... On 9/21/08, Matt Mahoney <[EMAIL PROTECTED]> wrote: > > --- On Sun, 9/21/08, Vladimir Nesov <[EMAIL PROTECTED]> wrote: > > > Hence the question: you are making a very strong assertion by > > effectively saying that there is no shortcut, period (in the > > short-term perspective, anyway). How sure are you in this > > assertion? > > I can't prove it, but the fact that thousands of smart people have worked > on AI for decades without results suggests that an undiscovered shortcut is > about as likely as proving P = NP. Not that I expect people to stop trying > to solve either of these... The present field of CS is dominated by people with large memories and little/no creartivity. Five minutes talking with the head of any CS department will convince you that they think that everything really worthwhile has already been done, and all that remains is just lots of polishing. It was these same simple minded PhDs who drove Perceptrons and Shallow Parsing into obscurity for decades. The "smart people" you refer to must either work in this very toxic environment, or work without funding. While I generally agree with your conclusion (mostly because I think that "successful" AGIs aren't worth the paper needed to print out their software), your method of reaching it is flawed, because it presumes a continuation of the present toxic process. In that toxic environment, a petabuck would not produce results. My proposal: Much like the "College of Science" at nearly all univesities was subsequently chopped up into freestanding departments like Physics, Chemistry, Biology, etc., so now CS departments need to be chopped up, at minimum to separate the small minds from the real advances. The Department of Computational Intelligence needs to be completely separate from the Department of Information Technology (that the present CS department chairman can head), the Department of Predictive Computation, etc. Only after this happens can ANY money be productively spent in an educational research environment. Steve Richfield ------------------------------------------- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
