> because that is not a verifiable task that can be scaled in RL like math or coding
Is it not, or are the feedback loops just longer? I could imagine being able to build a pretty good simulator of business operation for various kinds of businesses. (Some decent ones already exist as games.) From there, if you can build a good enough simulation, I can't see why you couldn't RL it.
Of course it's easier said than done to generalize it to the real world, where the possibility space is large. But I don't necessarily see an impediment to getting there eventually.
The feedback loops are quarters and years, so no you cannot scale it like you can scale math or coding. This is the reason we are seeing the most progress in those domains. Even ML is harder to scale because test runs are expensive and consume significant calendar time, so research "taste" is still important.
A full simulation would be, but I suspect even a partial simulation of evolving a software project though multiple "checkpoints" over a quarter would improve model performance
"Running a business" includes lots of real-world interactions that would be hard to model. An AI trained on an accurate simulation would probably get torn to shreds by scammers in the first six months. The models that exist today have enormous blind spots.
Google's ad network, the one that drives tens to hundreds of billions of annual profit for them? Also one assumes that the companies using it to advertise expect it to be profitable or they wouldn't do so. What's your evidence that it's not profitable for almost anyone?
OpenAI is in minute one of building out its ad network.
See: what Facebook's ad network looked like in the early days.
If Meta can support a $240 billion ad network. GPT usage can support a $80-$100 billion ad network right now. They're new to building ad networks, and they've chosen not to be in a hurry about it because they were getting cheap money elsewhere. Their sense of urgency about it will increase as the AI bubble valuations decline and the cheap money becomes far more expensive - that's all happening now.
Why would the emissions be less? Whether you're burning the gas in a flare or burning it to produce power, it's still burned, right? Do the power plants burn it more completely, or are they doing carbon capture?
Or do you mean less net emissions, assuming that bitcoin mining would have otherwise been powered by other means?
Presumably it's the latter. The useful work of bitcoin mining now being powered by the otherwise wasted flare. It's kind of great actually. But if it's so easy to generate electricity from these flares, why aren't we doing flare-powered electricity generation all over the place and just pushing net-power back to the grid? Oil derricks must be wired for power, so why aren't we plugging stirling engine generators into the grid and feeding that power back?
I think because the grid connections would be uneconomical. These sites tend to be somewhat remote, and you'd need significant transmission lines to feed back the amount of power generated, which would ultimately be obsolete once the well was dry. (And just using it to power the oil operation itself wouldn't use enough power to be worth it or to use up the flare gas.)
But if you can relatively cheaply set up a large consumer on site, then it makes sense. I wonder if we could do mini data centers like this.
So the math usually breaks down that as expensive as power is for a data center, having equipment idling is even more expensive.
We've had many talks over the decades about people moving server load around from areas of high power cost and/or temperature to areas of low cost and/or temperature based on live tracking, but every time someone has showed me the math that it's not even close to economical.
Then add the extra complication in the US of CAPEX versus OPEX, which I still don't understand why we've made it so complex, as it seems to completely eliminate some business models.
What are the fundamental differences between data center and crypto workloads that makes it work for the latter but not the former? I guess the former would need better network infrastructure, for one thing. But both seem like they should be fairly similar on the CAPEX vs OPEX front, no?
I'm also not sure the associated gas from oil wells is actually that intermittent? I think at least in some cases it can produce power pretty consistently. Just not forever.
There are different kinds of stranded power. A gas flare has, I presume, an expected average rate of production per day and if that is either stable or you provide a reservoir, you can install enough hardware to consume the median output, and then cross your fingers it will be depreciated before the well is exhausted. Not knowing much about wells (mostly only as pertains to GG emissions), I can't say if that's likely or virtually guaranteed, but I think we can deduce it's at least very likely, since it's already being attempted.
Then there's stranded solar and wind, and hydro, where either the grid cannot use the current power, or the transmission lines aren't rated to handle peak power. Those would all idle your equipment quite often, particularly in the case of peak shaving. Or conversely, if there's a power station you want to run as base load and keep running even when demand sags, and so you get a discount based on the cost difference between the producer's other options.
Perhaps some day we will have a thermal battery that is sufficient to vary the power draw of a datacenter by banking or expelling heat on a cost/benefit basis, but the heat pumps are only a fraction of the total power draw, so getting more than a 25% discount on net power draw seems unlikely, and then subtract travel time for maintenance and administration, theft prevention, etc, probably doesn't work out.
Generation does not need to be high voltage. Just like distributed solar, power pushed back gets to benefit from the local transformer's boost. It's exactly the same wires that are already there.
The problem of course comes later when the wasteful industry has proven that the stranded resource can be recovered but has long term contracts with enough locations that newer, better uses for that resource cannot do so.
But then you get into the, "If a tree falls in the forest" problem because would anyone else have actually moved on this if they hadn't?
AI is still the new big thing. Remains to be seen whether it will eventually just become a relatively stable part of the substrate, like the internet, or if the curve will keep bending upward - in which case it will likely remain top of mind as long as human minds are relevant. My bet's on the former, but we'll see.
Interestingly, HN by and large was extremely skeptical of crypto once it went from "cool technical project showing how you can implement a digital currency using shared consensus" to a bunch of speculation schemes. I recall numerous comments stating how "HN is biased against crypto" whenever someone pointed out all the problems with the tech.
One of my personal favorites was when Brian Armstrong got on here pitching his idea that eventually turned into Coinbase and everyone shit on him. Being a bit of a skeptic to crypto myself at the time, I was lukewarm on the idea, but man I wish I had joined him. For me personally it was a really good learning experience to stay open to technology no matter what my personal biases are.
That being said, when the AI stuff started making an appearance, I fully embraced it. This went against my gut feeling being a mid-level curmudgeon software engineer who had cut my teeth on code more nights than I can count until 4am.
I suppose that demonstrates a third possibility, that the bubble will burst and the hype train will end. I'm personally more bullish on its usefulness than that, but again, I guess we'll see.
I too doubt it will pop the way crypto did. The difference being, if you ignore the weird financing and ya know, societal collapse from AI, it's doing something useful!
But it sure felt like every HN story for a bit was crypto/blockchain/NFT related.
There are likely valid points made here, but I think the conclusions are considerably too strongly held. For one thing, it's already clear that humans value the achievements of other humans outside of their pure economic value. We are impressed by the athletic, artistic, and intellectual abilities of other humans, even in fields where we are surpassed by machines or computers. Cars have been faster than humans for a century, but we still value Olympic sprinters. Computers have been superior at chess for decades, but we've all still heard of Magnus Carlsen. AI can now produce art that only the most talented humans could hope to replicate, but most of us aren't particularly interested, because there's no human mind behind it.
So assuming we did somehow manage to create a post-scarcity world, I think people would still want to learn and to improve their skills. And we would still value those skills. There would also still be work to do; work that either people do because they want to, and/or that we don't want computers doing for us. Much of it work that in today's world is undervalued.
In that utopian world, there would also be room for hobbies that could fill some of the human need to produce value that work currently does. Lots of people would love to be able to learn a craft or play a sport but don't currently have the time, money, or energy to do so.
I have plenty of concerns about the impacts of AI on the world, but IF we do somehow manage to get the good outcome where it truly is transformative and the benefits manage to be at least somewhat evenly distributed... well, I agree that we won't all just hang out, but I do think we'd be ok.
The problem is that pursuits you mentioned are 'winner take all' examples. Who knows the name of the 29th best chess player in the world? Probably not very many. And that's still a very formidable bar to cross for most humans on earth.
So then, what does a laid off business analyst who built and edited spreadsheets most of their life are to do now? Pick up athletics in hopes of making it at the seniors' olympiad? Shake their ass on TikTok in hopes of becoming a 55 year old "influencer"?
I've always wondered who the fans of minor league sports are, but I assume there's enough of them beyond the athletes' families and the major league scouts trying to farm the best players for their own teams. There's a niche for anything.
We celebrate the highest achievers most, but that doesn't mean that the rest of us don't feel satisfaction in similar pursuits. Many people gain satisfaction from improving their skill at chess or sports or other crafts. I play ultimate (frisbee) twice a week, and I can assure you it's not because someone might eventually pay me to do so. I do keep trying to improve though. I'm glad that my work schedule leaves me the time and energy to do that.
What? We’re talking about a post-scarcity future and you think someone will be depressed because they don’t get to edit spreadsheets anymore? Poor guy only has infinite opportunity to pursue their limits (or not!) in the area of their choosing…
Do I play hockey with the end-all-be-all of becoming the next Gretzky? Or do I play because I enjoy it and want to push the limits of my mind and meat-suit?
Maybe I'm cynical and have low status friends but when I look around only a minority pursue hobbies with any great dedication or consistency. People pursue careers to project status. When success no longer depends on merit it will devolve to proxies like nepotism, scheming, intrigue etc. We see how such utopias play out in regimes like Soviet Union or North Korea. Your only assets in such places are who you know and even that is very temporary as allegiances are reshuffled with some regularity and woe is you if you bet your fortunes on the wrong one.
i mean, i think a lot of people probably know the 29th best chess player in the world. they may not be world-renowned but i bet the experience of that person is that they get a lot of respect for being super good at chess. people still play chess at all levels to try to better themselves - i think most people who play chess don't do it because they think they'll eventually get to #1 in the world (same with basketball, or any sport).
a business analyst who edited spreadsheets already wasn't particularly high status - i'm not convinced their life would be worse if the job was removed from their lives and they could do whatever they wanted instead.
Nit: I’ve never heard of Magnus Carlsen, and I can’t imagine most people have, outside of those really into chess. The only chess person I’ve ever heard of is Gary Kasparov, and that’s only because it was a big deal when Deep Blue beat him.
Is it not, or are the feedback loops just longer? I could imagine being able to build a pretty good simulator of business operation for various kinds of businesses. (Some decent ones already exist as games.) From there, if you can build a good enough simulation, I can't see why you couldn't RL it.
Of course it's easier said than done to generalize it to the real world, where the possibility space is large. But I don't necessarily see an impediment to getting there eventually.
reply