$1e12 = US Federal Discretionary Budget (annual) 10% = Fraction for procurement of the best unified model of society $1e11 = Unified model of society budget $4.6e6 = GPT-3 cost of induction 1e12"words" = GPT-3 corpus size 5e12B @ 5B per word average = GPT-3 corpus size $9.2e5/1e12B = GPT-3 induction cost per byte
Now, if we assume induction as intelligent as GPT-3 is worthy of Federal funding and induction cost per byte scales linearly (yeah, I know... bear with me): 1e17B/year = Socially relevant corpus size So if every political faction out there were in cut-throat competition with every other political faction to get _their_ favorite dataset included in the corpus, how "inclusive" could this model be? I mean, after all, there are _only_ 100 quadrillion bytes to go around. ;-) And, yeah, I know, "algorithmic bias" blah blah blah. First off, as I keep telling these "algorithmic bias" morons (I'm being kind here), there are two kinds of bias, both of which are dealt with by using lossless compression as the model selection criterion: 1. Inductive bias 2. Data selection bias Inductive bias is dealt with by going with the size prior of Solomonoff Induction, which is the basis of using lossless compression size as the model selection criterion. Yeah I'm sure some twerp in some podunk "studies" department will claim the 86 family's instruction set is an example of "whiteness bias", and they might even get some twerps in Congress to yammer something like that, but this does not strike me as the kind of thing that would make even the most cucked captians of industry cower for fear of being called "racist". (knock on wood) Data selection bias is dealt with by two factors that conspire to ruthlessly expose biased data _as_ bias data: 1. Biased measurement instruments are routinely audited for bias (if not outright errors), not only _directly_ by other measurement instruments but, _indirectly_ by cross-disciplinary consilience. As you increase not only the number of measurement instruments but the number of disciplines contributing measurements, the likelihood that bias will not be called out as such in a _unified_ model drops to zero. 2. Lossless compression demands that you do, so-unify, your model. OK, so now we're popping the stack back to the first "yeah I know" relating to selecting GPT-3's induction as the exemplar (and less egregiously, the linear extrapolation): Yeah I know and that's why I suggest we put that $1e11 into a PRIZE to be awarded to _any_ inductive model judged _solely_ on how well it losslessly compresses the corpus. This invokes a new "yeah I know" which is that the prior calculation of the size of the corpus is no longer _as_ pertinent since there will be 2 additional demands on compute resources: 1. Redundant inductions run by multiple competitors. 2. Decompression for judging winners. First let's deal with redundant inductions: This requires lowering the 100 quadrillion bytes -- and that makes it harder to accomodate all political factions' data in the competition's corpus. OK, fine. Let's say you can't fit them into 1 quadrillion bytes because everyone is ready to blow each other's addled brains out because of Twitter memes and such: How about we just include the IRS's entire database, the US Census's entire database, the Federal Reserve's entire database, the CDC's entire database, the EPA's entire database, the Genome-wide association study's entire database, the General Social Survey's entire database, the DoEnergy's entire database, the DoEducation's entire database, the DoL's entire database, the DoCommerc's entire database, the DoHHS's entire database, the DoT's entire database, the DoJ's entire database, the FBI's entire database, the General Social Survey's entire database... anything else? I seriously doubt we are anywhere near 1 quadrillion bytes. That means every year about 100 $1e9 awards can be issued for inductions by algorithms as resource-prolifgate as GPT-3. Sure, there will be a lot of losers out there rolling the dice with their own money and coming up snake eyes. Too bad. That's what _always_ happens with these objective, high-stakes prize competitions. It's part of the fun and it's also why they are strongly leveraged philanthropic investments. Second, let's deal with the decompression for judging winners. Now, if you haven't noticed, machine learning _inferences_ tend to be orders of magnitude less costly than _inductions_. But, ok, we'll deal with it anyway. This can be handled by the simple expedient of charging an entry fee to pay for the judging cost. QED PS: "Yeah I know" I didn't address the linear extrapolation of cost per induced byte, but there does seem to be a kind of "industrial learning curve" to learning, which means a linear extrapolation would be quite conservative. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T088b77b93050aad5-M9978ae583b39a1c741364604 Delivery options: https://agi.topicbox.com/groups/agi/subscription
