The problem with descriptive names is that they start descriptive but then become proper nouns. At a former employer in the Fortune 100 outside the software industry, everything started with a descriptive name, that then became an acronym. And as every project and tool inevitably developed its own idiosyncrasies, the descriptive name pretty soon didn't tell you anything useful about the project at all.
It is an unavoidable reality that knowing something's name gives you very, very little information about what that something is. That's what sentences are for.
> But if intelligence really does become too cheap to meter, it will become possible to do a perfect reconstruction and synthesis of everything. LLMs are watching (or humans using them might be). Best to be good.
I cannot believe this is just put out there unexamined of any level of "maybe we shouldn't help this happen". This is complete moral abdication. And to be clear, being "good" is no defense. Being good often means being unaligned with the powerful, so being good is often the very thing that puts you in danger.
I've had the same though as Karpathy over the past couple of months/years. I don't think it's good, exciting, or something to celebrate, but I also have no idea how to prevent it.
I would read his "Best to be good." as a warning or reminder that everything you do or say online will be collected and analyzed by an "intelligence". You can't count on hiding amongst the mass of online noise. Imagine if someone were to collect everything you've written or uploaded to the internet and compiled it into a long document. What sort of story would that tell about who you are? What would a clever person (or LLM) be able to do with that document?
If you have any ideas on how to stop everyone from building the torment nexus, I am willing to listen.
1. Don't build the Torment Nexus yourself. Don't work for them and don't give them your money.
2. When people you know say they're taking a new job to work at Torment Nexus, act like that's super weird, like they said they're going to work for the Sinaloa cartel. Treat rich people working on the Torment Nexus like it's cringe to quote them.
3. Get hostile to bots. Poison the data. Use AdNauseum and Anubis.
4. Give your non-tech friends the vague sense that this stuff is bad. Some might want to listen more, but most just take their sense of what's cool and good from people they trust in the area.
This seems to me like a form of social engineering, or to some extent, being a bit insufferable. And, rest assured it will not result in anything useful. The only result of this is that you will alienate your friends and colleagues if they work for an employer you don't like.
I think we need to stop focusing only on the AI aspect of this. Yes, it's an important component to the sort of mass surveillance system you're describing, but it's not the only component. The internet, advertising, privacy, all of these are integral to this outcome.
While I don't have a general solution, I do believe that the solution will need to be multi-faceted and address multiple aspects of the technologies enabling this. My first step would be for society to re-evaluate and shift its views towards information, both locally and internationally.
For example, if you proposed to get rid of all physical borders between countries, everyone would likely be aghast. Obviously there are too many disagreements and conflicting value sets between countries for this to happen. Yet in the west we think nothing have having no digital information borders, despite the fact that the lack of them in part enables this data collection and other issues such as election interference. Yes, erecting firewalls is extremely unpalatable to people in the west, but is almost certainly part of the solution on the national level. Countries like China long ago realized this, though they also use firewalls as a means of control, not just protection (it doesn't have to be this way).
But within countries we also need to shift away from a default position of "I have the right to say whatever I want so therefore I should" and into one of "I'm not putting anything online unless I'm willing to have my employer, parents, literally everyone, read it." Also, we need to systematically attack and dismantle the advertising industry. That industry is one of the single biggest driving factors behind the extreme systematic collection and correlation of data on people. Advertising needs to switch to a "you come to me" approach not a "I'm coming to you" approach.
It's nice that the LLM-enabled panopticon still cannot find this very recent related media, [0] but my silly mind can. It is actually an interesting commentary from a non-tech point of view. This is how the rest of the world feels:
Anyway, back to work trying to make my millions using Opus and such.
Well the companies that facilitate this have found themselves in a position where if they go down they take the US economy with them, so the maybe this shouldn't happen thing is a moot point. At least we know this stuff is in stable, secure hands though, like how the palantir ceo does recorded interviews while obviously blasted out of his mind on drugs.
To be clear...prior to this recent explosive interest in LLMs, this was already true. Snowden was over 10 years ago.
We can't start clutching our pearls now as if programmatic mass surveillance hasn't been running on all cylinders for over 20 years.
Don't get me wrong, we should absolutely care about this, everyone should. I'm just saying any vague gestures at imminent privacy-doom thanks to LLMs is liable to be doing some big favors of inadvertently sanitizing the history of prior (and still) egregious privacy offenders.
I'm just suggesting more "Yes and" and less "pearl clutching" is all.
I recently learned about Singapore’s seemingly excellent public housing system that is used by over 75% of its population. Singapore being so capitalist about everything else while carving out housing I think provides evidence that capitalism can be made stronger by keeping certain things away from strict market forces.
You can't have a free market if supply (or demand) is constrained.
You either go the liberal+regulated free market way, which will get lobbied, then overregulated, or you go the socialized way, where you allow local government and associations/unions to compete in the market... with the exact same rights as the companies, and no weird rules that prevent them to truly compete.
And if you have a natural monopoly (energy distribution is the worst offender), you have to go the central planning way I guess, until we find something better.
I think the ship very much has not sailed on how different spaces treat LLM responses. Are LLMs something that you can use if you want, but are considered rude and banned from being blatantly posted without human ownership? “You can’t use an LLM” would be an impossible rule, but “You can use an LLM to write your response but you have to take responsibility for the output” is feasible.
> Surely you can think of something cool to build with that, which doesn't involve money.
People have been saying this for nearly a decade and many people with a great deal of motivation have failed to find any that worked and couldn’t be done better with traditional databases. It’s past time to ask if there’s actually any gold in them hills.
This might be a bit of talking past one another. The rent vs buy argument should be comparing similar housing, but your comment bakes in the assumption that renting=apartment. That may be true for your area, just wanted to point out the dissonance.
If Wikipedia was not real, it would sound like a naive utopian thing you’d read in a bad paperback. Multi-language repository of basically all human knowledge that’s extremely resistant to government capture and contributed by volunteers totally transparently? Bullshit. And yet … there it is.
Having said that I agree that Wikipedia is a tremendous achievement, and despite the wards it's amazing that it works as well as it does.
If you permit me to go on a tangent: Wikipedia is also interesting as a test case for our definitions of (economic) 'productivity'. By any common sense notion of productivity, Wikipedia was and is an enormous triumph: the wiki models harvests volunteers' time and delivers a high quality encyclopedia for free to customers. By textbook definitions, Wikipedia tanked productivity in the encyclopedia sector because these definitions essentially put revenue in the numerator and various measures of resources expended in the denominator---and Wikipedia's numerator is approximately zero dollars.
Wikipedia is not resistant to capture. It is structurally exposed to coordinated editing groups, and the platform has no robust way to detect or neutralize them. On politically or strategically important pages, especially where state or financial interests are involved, organized paid editors can and do shape coverage, dominate talk pages, and crowd out dissenting contributions.
There have been many editing scandals over the years. While one could argue that editing cabals being caught and banned demonstrates that Wikipedia has structural resistance to such behavior, the scandals that become public are, almost by definition, the ones that were clumsy enough to get caught. The mechanisms that appear robust largely work only against amateurish, poorly hidden or commercially indiscreet operations. What they do not reliably detect are long-term, well-resourced, politically motivated or state-aligned editing groups that behave patiently, avoid obvious sockpuppeting footprints, and stay within procedural boundaries while still dominating page direction.
There’s also the rules of Wikipedia. Things like what constitutes a “source” really skew things. Lots of worthy sources aren’t a mainstream news outlet or academic journal, but are excluded. This creates biases of its own.
Wikipedia is great in many ways, but I would say it’s not really designed to be neutral.
Wikipedia’s sourcing rules are a big part of the problem. Reliable sources on Wikipedia mostly means prestige, center-left Anglosphere media and a narrow slice of academic publishing. In practice that gives generalist news outlets more epistemic weight than field-specific experts, and it excludes large amounts of accurate but unfashionable or non-English material. The result is a replication of whatever biases are already baked into mainstream journalism.
The gatekeeping also isn’t democratic. A small, durable cluster of editors on WP:RSN effectively decides which sources are allowed, banned, or given special status. They aren’t vetted for expertise and don’t need to disclose conflicts of interest, yet their decisions propagate across thousands of articles. If a coordinated group influences RSN, they can shape whole domains simply by defining which sources count as "real."
WP:NPOV becomes a matter of which media ecosystems have the loudest megaphone, not which sources are actually most accurate. Control the allowed sources and you control the encyclopedia.
to me it feels like something that would be a 'wonder' you could construct in one of the civilization games. and i love that there's something that awe-inspiring just... sitting there for free on my computer.
To say the quiet part out loud, I don't think any serious companies have any intention to build a data center in space. There is no benefit in actually trying this. There is however, benefit in saying you'll do it to advance a narrative and distract from the problems terrestrial data centers are facing to an audience that mostly doesn't understand how heat transfer in a vacuum works.
It can be considered as advertising. Like Coca-Cola. Not actually anything new or real most of the time. But keeping the mind-share. Making the company seem like they are on cutting-edge, visionary and futuristic. After all the scam is build on future promises. And not the current day real profits.
Classic misaligned incentives - pretendgineering is far more profitable than building stuff that works reliably and is useful.
We've moved past bullshit jobs to a bullshit economy, which operates by moving money from investors to billionaires and back again, driven by pitch deck thoughts and prayers and implied threats. ("Bail us out or everyone dies.")
I always assume, unfortunately, that once companies start to get to a certain point they become strategic, and military applications comes into play. They then probably get special consideration when it comes to funding and access. All of Musk's efforts certainly fit this paradigm.
Google "atoms for peace." You will find an entire multidimensional cluster of hype ranging from Rickover maneuvering to get a nuclear navy, which seems to work pretty well, but on the way there it created a subsidized nuclear reactor business which was never in the money but for subsidies, loss leaders, and underbids. There was no Golden Age of nuclear power. There were fixed bid contracts that masked cost overruns until they didn't anymore. There was FOMO about Soviet gigantism and (subsidized) European nuclear projects.
Many of the dumb ideas being hyped in this AI bubble make sense viewed through this lens.
Data centres stirring up opposition? Sell a sci-fi vision that you will move them to Space! And reassure your over-extended investors that the data centre buildout rush you’re committing to isn’t going to get bogged down in protests and lawsuits.
The people hyping this stuff are not stupid, just their real goal (make as much money as possible as quickly as possible) has only a vague relationship to what they claim to be doing.
At the point, it's beginning to feel a bit like the 419 scam (where you make the details deliberately absurd so as to ward off people inclined to be sceptical early, leaving you with only the easiest marks.) SMRs! Data centres in space! "phD level AIs".
Most investors can time this aspect of the market accurately enough. It's tough for these people to stand by and watch profit being left on the table for a year or two, though. So they get back in, seeing how long they can leave their hand on the got plate.
Myself, I made the decision to go to cash a while ago, right before the recent AI pullback. Things were going great for a week until I started seeing all that money go unclaimed. I get back in, and the pullback I predicted happens. It was my own conscious decision to look past the gorilla in the room to get more free treats. I'll be fine but this is a good anecdote for how these things unfold.
It is not at all true that "Most investors can time this aspect of the market". This is laughably, absurdly, wrong - as if most people could predict the future. Here's a little advice I sincerely pray you accept : don't trade options.
Really? Logic wouldn't dictate that if I'm up 300% or more over two years and everyone is starting to get jittery about an AI bubble that perhaps I should pull out now and await the pullback? If it happens in a year, and I can buy back in at a 15-20% discount, that is also a return!! Do you hold for possibly another 5%? That doesn't make any sense. Your cash gets 4% a year just waiting--paid monthly.
Yeah, that you could do, though even then if the timing is sufficiently uncertain you might be in trouble, and it's particularly risky in a time of stubbornly higher-than-ideal inflation. If you happened to have a bunch of Hype-y AI Ltd, then sure, probably. Far less clearly a good idea if you just have the S&P500, though.
It's further complicated by the fact that most of the worst examples of AI hype are not public. Like, if and when the bubble bursts, the hyperscalers will likely get burned, but they're not going to go to zero or anywhere near it.
And that's assuming you already have stocks; it's very different, risk-wise, from shorting or buying puts.
> Your cash gets 4% a year just waiting--paid monthly.
Is the answer slavery? Once you've been taken to Mars and given your underground living quarters, you're going to be stuck there and not have any option but to carry on working or be thrown out to die on the surface.
They should be doing that somewhere inhospitable to prove out the technology and concept. Set up in the arctic and work only with loads of materials that correspond to a starship load.
But then again no one is really serious about Mars.
You mean, not very much? Everything about space-based anything is dependent in the short to medium term on Starship making mass to LEO cost about as much as air freight.
Starship, at least as a rapidly reusable second stage, may fail, rockets are hard. But you aren’t really engaging with people’s dreams if you start from “we don’t have access to the technologies that appear to represent a one to two order of magnitude cost shift”.
The only real advantage is 24/7 power without having to use batteries (or some other power supply at night or when cloudy). The way solar prices are going the problem of suppling power when the sun isn’t visible is a real bottleneck.
For 24/7 solar... you are either in a sun synchronous orbit or in a very high orbit.
The sun synchronous are polar orbits ($$$) that are preferred for earth observation (so that the sun is casting the same shadows). As these are polar orbits, the satellite is not overhead all the time and getting a satellite into such an orbit takes a bit of work.
A SpaceX is at about $3k / kg to LEO. The numbers I see suggest a $20k / kg to a polar orbit.
The next option is being far enough out of the way that the earth's shadow isn't an issue. For that, instead of a 500 km sun synchronous orbit, you'd be going to 36,000 km orbit. This is a lot further from the surface, takes a lot more fuel... and it's a geostationary orbit.
However, as a geostationary orbit, these spots are valuable. Slots in this orbit are divided into slots.
> There are only 1,800 geostationary orbital slots, and as of February 2022, 541 of them were occupied by active satellites. Countries and private companies have already claimed most of the unoccupied slots that offer access to major markets, and the satellites to fill them are currently being assembled or awaiting launch. If, for example, a new spacefaring nation wants to put a weather satellite over a specific spot in the Atlantic Ocean that is already claimed, they would either have to choose a less optimal location for the satellite or buy services from the country occupying the spot they wanted.
> Orbital slots are allocated by an agency of the United Nations called the International Telecommunication Union. Slots are free, but they go to countries on a first-come, first-served basis. When a satellite reaches the end of its 15- to 20-year lifespan, a country can simply replace it and renew its hold on the slot. This effectively allows countries to keep these positions indefinitely. Countries that already have the technology to utilize geostationary orbit have a major advantage over those that do not.
Furthermore, the "out of a nations control" - those slots are owned by nations. Countries would likely be very annoyed for someone to be putting satellites there without authorization. Furthermore, they only work with the countries on those areas. They also require spacing to ensure that you can properly point an antenna to that satellite.
Furthermore, geosynchronous orbits have a 0.5 second round trip lag. This could be a problem for data centers.
Putting things in these orbits is pricy. For LEO, you'd need a lot of them. For geosynchronous, the idea of servicing them is pretty much a "you can't do that" (in 10 - 20 years they use their last fuel and get pushed to a higher orbit and pretty much get forgotten about).
Satellites in geosynchronous orbit are things that need to be especially well behaved because any orbital debris in that area could really ruin everyone's day.
I think a prerequisite to doing any really big stuff in space would be fully and rapidly reusable launch rockets, which could get costs down by a couple orders of magnitude.
And geostationary isn't necessary for this. You could go a bit higher or lower and still have 24/7 sunlight. Relay your communications through Starlink or something and you have full connectivity.
That said, I think orbital data centers still don't make sense, for all the reasons described in the article.
The real advantage is latency but who really needs that? The military may have some use cases (think remote control of drones and the link between the controllers and satellites) but the use cases are limited
There's one obvious potential application, which is caching of common requests. If something like segments of streams or any CDN contents is cached on the satellite, it reduces communication to a single hop for a large portion of traffic (IIRC, 70% or so?). Storage is very lightweight these days and failure to read cached data is not critical, so putting lots of SSDs on a LEO constellation satellite seems like a no-brainer to me if you're trying to optimize bandwidth usage.
That seems like it would make the most sense on the "last mile". So, adding caches to the LEO satellite ISP birds would be a good idea. I wonder if Kupiter, StarLink, et. al. do that. (And if not, their reasoning against it since they've surely considered it.)
Geosynchronous orbits do not pass through the Earth's shadow as much as you might think. These orbits sit in the same plane as the equator, which is tilted 23.5 degrees when compared to a line from the sun to the earth.
They still pass through the earth's shadow in the weeks around the equinoxes though. Worst case is about 70 minutes of shadow.
That said, it seems more likely to me that there is no requirement to stay over the same spot on the earth, and a lower altitude sun-synchronous orbit would be used.
Elon is 100% planning to put significant ai compute in space. He is probably planning to do it in a decentralized way.
He has the launch platform (spacex), he has the existing power and data infrastructure (starlink), he has the demand side. (Xai)
Will he succeed? That is different question. Is it possible to add enough power generation and thermal radiative capacity to starlink nodes to bother? Don’t know, but an analysis that fails to answer those two specific engineering questions is useless.
Space datacenters have the dual-use of tracking and weapons targeting which is needed for a robust Golden Dome architecture (immune to comm link jamming, terabit image sensor processing)
> Space datacenters have the dual-use of tracking and weapons targeting
Space datacenters aren't going to be equipped with military infrared sensors. They stick out like a sore thumb on the electromagnetic spectrum and the second you test it every peer-power would know it's a military platform. Nevermind the fact that the satellites don't transmit to American C2, so they'd need laggy ad-hoc networking to reach STRATCOM over on Link 16.
> Musk is involved in every aspect of Golden Dome.
SpaceX is the only firm on the planet produces a booster stack with the throw weight to put a usable kinetic weapon in orbit. It's not their first military contract, Musk has been sticking his nose in the NRO projects for years now.
SDA’s “Battle Management Layer will provide automated space-based battle management through command and control, tasking, mission processing and dissemination” to support time-sensitive kill-chain closure. https://www.sda.mil/battle-management
Golden Dome and future missile tracking and ISR will depend on real
-time insights, which requires Edge Computing on orbit, running advanced AI/ML algorithms.” https://unibap.com/news/defense-in-the-foreground
This was my thought the first time I heard these talked about on a podcast where it talked about there being infinite cooling ... and I just kind of face-palmed because it was like, "This is being discussed by people who don't know things about space." We already have places on earth with effectively unlimited solar power and effectively unlimited cooling (though not the same places) but without having to launch stuff into space.
Ok, but couldn’t you equally say that about anything constructed by industrialized people in places that used to have lots of non-industrialized people?
I think GP's point is that an advanced nation-state could just as easily shoot down an orbiting data center as an oceanic data center and that "international space" offers an equally flimsy defense as "international waters" but a much larger price.
Swarms of satellites need to maneuver, which includes maneuvering directly toward the atmosphere.
It would take zero anti-satellite weapons to take down Starlink. Just point a good old fashioned gun at the SpaceX engineer who can issue maneuvering commands to the satellites.
You only need to destroy a few. Then you have a cloud of debris that will take down the rest or at the very least force them to use all their fuel making evasive manoeuvres.
How absurd is absurdly decentralized, here. A hundred ground stations? Thousands? Do they really have more than can be shut down by the FBI domestically and blown up by the USAF internationally?
And how does decentralized ground infrastructure save you from a centralized executive?
Over one hundred ground stations, spread across the world. More on demand - Starlink allows one to use terminals as makeshift ground stations in a pinch.
Uncle Sam could bring Starlink down, probably. For anyone else, that would pretty much require WW3.
Executives don't matter as much as you think they do. No credible executive is going to cave to random death threats, and carrying them out would cause new executives.
Now, would SpaceX eventually become a shell of its former self without Musk calling the shots? Maybe. But if the shell you're worrying about is Starlink orbital shell, and the time you're worrying about is today and not in ten years? Killing Musk doesn't help you much.
You think Musk would refuse, and give up his freedom or even his life instead of complying with a US government demand? The point isn’t to actually kill him. The point is that you can, and you use that to force compliance.
https://en.wikipedia.org/wiki/Communications_Workers_of_Amer...
reply