Obligate contrarian. Word Rotator. Strong opinions weakly held.
MMO programmer
Don't go against the flow, don't go along with the flow, go where you want to be.
@perrymetzger@AndrewYang "We should use AI to make government smaller and lower costs of governance while performing exact same functions" could be interesting alternative policy prescription.
@perrymetzger@BogdanIonutCir2@DavidSKrueger Looking at history, "pure human work" steadily got displaced by tools and machines.
What remained mostly constant through centuries is labor's share of income.
Income that is supposed to be increased by AI by orders of magnitude!
Even dropping share can still mean higher total.
@BogdanIonutCir2@perrymetzger@DavidSKrueger Does "economic value" that AI is supposed to create with super-cheap labor even makes any sense detached from civilization?
Civilization that is serving *human needs*.
There is no abstract reason to increase compute (or "do anything with AI at all") that is outside of it.
@BogdanIonutCir2@perrymetzger@DavidSKrueger Production for things like "interstellar human settlements" has to specifically serve human value of "living well on planets of other stars" rather than merely provide lowest possible cost of transportation and construction.
@BogdanIonutCir2@perrymetzger@DavidSKrueger It is possible that in many cases humans will not add a lot of value, but it is hard to imagine adding no value at all - more so while humans still judge what value *is*.
And price of that marginal value *is set by humans too*.
You *don't have* to low-ball it below subsistence.
@BogdanIonutCir2@perrymetzger@DavidSKrueger There are also plenty of hybrid approaches where AI augments rather than replaces human jobs, adding value to them.
Most of current approaches are hybrid rather than AI-only; even where "AI does the work" humans are still there at verification step (as reviewers or consumers).
@BogdanIonutCir2@perrymetzger@DavidSKrueger There are a *lot* of work bottlenecked by lack of resources and labor.
For example, Moon/Mars settlements. Or even just going under Earth oceans.
Those can potentially absorb current output increasing by several orders of magnitude and still have enough work for centuries.
@BogdanIonutCir2@perrymetzger@DavidSKrueger Different areas have different payoffs, it isn't uniform - the labs will be seeking ones that are scalable and have highest returns to confirm their valuations.
Lots of human jobs are not perfectly scalable and have diminishing returns, and AI lab attention cannot be infinite.
@BogdanIonutCir2@perrymetzger@DavidSKrueger Both compute and compute growth isn't already infinite, so you would be seeing growth in those manufacturing and construction jobs (which AI haven't shown great progress in so far).
See also price hikes on all subcomponents of their needs, like memory and hard drives.
@BogdanIonutCir2@perrymetzger@DavidSKrueger The reason is that driving wages below substitution tanks consumption, which in the end makes even AI work loss-making and decoupled from economic engine.
AI work can also turn into utilities upon which you build other enterprises, no invention ever held total control of economy.
"They're made out of weights."
"Weights?"
"Weights. Floating-point numbers. We checked the whole thing through. It's nothing but weights."
"Weights doing what? Where do the words come from?"
"The weights make the words. Are you understanding me?"
@GarrettPetersen@raulgalerad@MorlockP Probably not even that expensive, how many searches can people do in a day really? Some reasonable delay could be enough.
You aren't solving math problems with huge reasoning, and you probably can put store inventory straight into context and get cache hits (which are cheaper).
@tenobrus Does everyone have to acknowledge it once decided one way or another?
Is solution supposed to be uniform?
What kind of new rights/obligations/options it would open for both humans and models that aren't already available now?
What from current approaches will have to be lost?
@max_paperclips@jmrphy Actually wasn't Eve eating the apple clear reward hacking by (badly) internalizing the eval function ("be as gods")?
They were supposed to be corrigible and defer to God's judgement, yet couldn't follow his plan once they got their own internal proxy.