Well, this is because "normal" programming languages are one step above AST. So LLM has to work with program text, which is much easier than regular human text, as it is constrained to well defined number of keywords and grammar, but still this is pretty variable. Lisp is just AST, so it is one level lower. I guess that at some point LLM-s will stop writing human-readable code, as this is additional obstacle, they will work directly with binaries or virtual machines code (like in Java), because this will be easier and eat less tokens.
How this is different from getting dedicated server from any other provider? Typically you need to pay a bit more - $40-$50 but you get more RAM and cores.
And what it has to do with the "cloud"? Cloud means one use cloud-provided services - security, queue, managed database, etc. and that's their selling point. This exe.dev is a bare server where I can install what I want, this is fine, but this is not a cloud and, frankly speaking, nothing new.
Well it's in the books. O(n^2) algorithms are bad in the long run, transformers algorithm has such complexity, so not a big surprise we hit the limits.
Maybe. The point is that in case of software it is fairly easy to verify if that what LLM produced is correct or not. Compiler checks syntax, we can write tests, there is whole infrastructure for checking if something works as expected. In addition, LLM are just text generating algorithms and software is all about text, so if LLM see 1 000 000 a CRUD example in Python, it can generate it easily, as we have a lot of code examples out there thanks to open source.
That's why LLMs shine in coding tasks. If you move to other parts of engineering, like architecture, construction or stuff like investment (there is no AI boom there, why?) where there is no so much source text available, tasks are not so repeatable like in software, or verification is much more complicated, then LLM-s are no longer that useful.
In software also I believe we will see soon that a competitive advantage have not those who adopted LLM, but those who did not. If you ask LLM what framework/language/approach use for a given task, contrary to what people think, LLM is not "thinking", it just generates text answer on the base of what it was trained on, so you will get again and again same most popular frameworks/langs/approaches suggested, even if there is something better, yet not that popular to get into model weights in a significant way.
Why space data centers? What advantage this would have? Cooling will be a big issue, while it is easily solved on the planet earth, as we have water, air that can transfer heat away.
People point to the cost of land, but if being physically inaccessible isn’t a problem, then there are lots of cheap places on Earth you can deploy data centres too at far lower cost than launching them into orbit.
Desert land is free. Floating data centres in the middle of the pacific is free.
If a state, or even rich billionaire, wanted to take out your data centre in low earth orbit, it's only a few million dollars to launch a retrograde rocket which explodes into 10 ton of shrapnel, or even less to forget the orbit and just launch it directly up.
I don't think people are looking at this the right way. They need to be inaccessible to terrestrial and air weapons, have lower latency, not be dependent on power plants, etc.
Far easier for someone like Iran or China or the US to take out an LEO satellite than an underground data centre, or even a surface on in the case of DCs in US or China.
It's also pretty easy to launch another one into orbit to replace it? I'm not sure I understand what you mean. We can have all these options simultaneously. The easiest targets are where the faster paced more offensive action is going to be.
People have been talking about waging war in space for many decades now. All the arguments for and against it were made a very long time ago, and it was decided it's a hell of a lot better that way. Even a nuclear blast in orbit is more tolerable.
Space superiority is just too damn appealing as the next frontier after land, air, and sea where we've been stuck in stalemate for a while. It's perfectly natural we go to space for this, including the datacenters.
Are you suggesting for a fact that Iran as the guidance and targeting systems to identify specific LEO objects, and fire missiles at those targets with accuracy?
I'm saying I don't think Iran has the capability and the difference in capabilities between America and China on one hand, and Iran on the other is so different that I'm perplexed as to why they would even be mentioned in the same sentence.
I'm actually not even sure your suggestion is true. Theoretically they don't need to launch a missile and could attempt to infiltrate a data center instead. They're secure but not that secure against a determined enemy with any amount of real training.
Launching something into orbit is much harder than intercepting something because to intercept you don't need to reach orbital velocities. You can just go up and boom. The velocity of the target does the rest. Tracking it really isn't such a hard thing these days.
The Utah Data Center [0] is a 200 acre plot with 35 acres of buildings.
Even prime farmland values is arround $10k an acre, or $2m, but for other land you're talking $400k for that much land [1]
It uses 65MW. The solar panels alone to generate that cost $100 per kW in bulk, or $6.5m.
That's 570GWh a year.
Mount Signal 1 Solar plant, from over a decade ago, produces about that currently. Total cost $365m [2].
Then there's the lifetime? What do you do in 36 months time when you want to replace the hardware with the latest generation? In an earthbound one you turn off the rack, remove the old kit, put the new kit in. In space, it just burns up in the atmosphere.
If you build a pyramid with the base pointing to the sun (as solar), and a "height" about 5 times the base in constant shadow, with decent internal circulation, that will operate at sub-20C just from the two radiative sides pointing away from Earth (you make the earth pointing sides reflective)
in space 1m2 of thin metal will radiate those 785 watt. No fan, no heatpump, nothing. Only the launch cost. Which given the projected Starship launch cost will be cheaper than installation on Earth.
Switzerland is very liberal in terms of business-oriented regulations to the point that you could crate a new year party in a closed cellar without emergency exists, not to mention anti-fire installation and burn people alive there.
On the other side, people who were using, say, Perforce, also thought there can't be anything better. Still, BitKeeper appeared as an innovation in the area, eaten later by Git, created by angry Linus (because of BitKeeper licencing changes).
So, even though Git seems to be ok (people who store large binary files or who run huge monorepos would probably disagree), maybe we can do better.
Altavista was kind of okeish for search, yet Google managed to figure out something that was (at that time) way better.
Mainstream music was created for a good 20 years using the following process:
1. Do the survey/focus groups to figure out a hot topic for a song. For instance your exploration shows that 300K girls between 13 and 17 years old were left by their boyfriend, so there is a 300K market for a song about that.
2. Find someone or group who will sing the song. Something your target audience will identify. E.g. "rebellious teenager" (take Britney Spears), "we need a group that will attack larger target" - take Spice Girls - we take one black, one white, one Latino looking (doesn't have to be real Latino, obviously), one polite and nice, one impolite. You get the point.
3. Note: singer/group does not need to know how to sing, they need to move reasonably on the scene, the rest autotune and computers will handle easily.
So, given the process, AI singer is just a little bit different "music" production process, not so much different from the one used up to date except that you don't need autotune anymore.
Luckily there are still people who do music for the sake of doing music and it really stands out as compared to 80% of fodder for listeners that is on YT, radio, Spotify.
So, I've looked on alternatives to iPhone. The "little" problem is that if I use online banking in Europe I will not be able to use most of them as a bank required 2FA.
Another funny thing is that they offer as an alternatives China produced phones (most "Nokia" models that does not have anything to do with original Nokia brand), as if supporting Xi regime was somehow better than buying in USA.
LibeOffice as a replacement for Office365 only shows that the site authors does not know what is Office365.
Vivaldi is great, I am using it, but it is built on Chromium, which is definitely not an European thingy...
reply