It all comes back to Searle's Chinese Room thought experiment, or a twist on it. Imagine creating a von Neumann-style self-replicating LLMs that do nothing but replicate and prove theorems. Well, then, what of it? You'd have a universe full of proved math theorems but does it all mean anything at all?
It's easy to blame Dartmouth for its shortsightedness but they have a point. Education, including graduate school, is more and more resembling vocational training and anyone paying for it, whether an individual or the university via an assistantship program, has to calculate the return on investment. With the recent FUD created by LLMs for math research, it is reasonable to ask what kinds of careers math PhDs can now expect and whether the field can absorb the current level of math PhDs.
You seem to have a weird misunderstanding of what a PhD is.
It's 3-6 years of your life spent figuring out everything everyone has ever tried to solve a narrow class of problems that you are passionate about, guided by an expert on that topic. Without the worry of being able to afford food or rent.
Around the end, you make a contribution or two to that topic and become one of the top experts on the world in the field.
People hire PhDs for two reasons, one if they really need an expert in that field, or if they want someone who is proven to be able to spend a long time on a topic while figuring out and understanding what others have been doing there for the last several years.
Smartness does not have anything to do with the second part here. Without this training to do research, you can often train someone to do something technical but all their new ideas will likely be those that have been tried and failed 20 years ago.
Of course you could find exceptional candidates who can do this without a PhD, but employers are not in the business of taking low probability chances.
I know a lot of people who have PhDs in fields like physics and math. I am familiar with their journey to reach that, and the paths they have afterwards. In all cases, they were people I would consider smart before deciding to pursue the PhD. I am not sure what you think my misunderstanding was. We seem to agree on the conditional probability that someone is smart given they have a PhD.
> In all cases, they were people I would consider smart before deciding to pursue the PhD.
That's the misunderstanding, I recon.
Most people who go through PhD programs are not smart. If I had to guess, maybe 20-30% would be smart in the traditional academic sense, and single digit percentages in the street smart/eq sense.
Out of the set I know (maybe more than 100 individuals, my school sent many to grad school) I'd say their EQ distribution was very similar to that of society as a whole, or slightly higher. At least, those were the vibes.
> Smartness does not have anything to do with the second part here. Without this training to do research, you can often train someone to do something technical but all their new ideas will likely be those that have been tried and failed 20 years ago.
Telling you what ideas have been tried and didn't work is exactly the thing AI can help with.
Smart people still need to apply that intelligence for it to mean anything. Math PhD programs provide a structured environment to develop specialized knowledge, and an outlet/goal for novel contributions to the field.
Even with an 180+ IQ you're not going to have a shower thought and solve the Birch and Swinnerton-Dyer conjecture without first learning the prerequisite concepts (and how to articulate your solution). Even very intelligent people still struggle to maintain the discipline needed over months/years to self-study with the depth/rigor needed for the equivalent degree.
I fail to see the economic case. You will need very large solar arrays and radiation hardened components, both of which make it uneconomical compared to a data center on Earth. This is another of Musk's genius ideas like the tunnels that were supposed to be carrying us under cities by now.
The problem is that experiments into this by Google have concluded that radiation hardening is unnecessary. HBM memory far surpassed their expectations for survival, exceeding 5 years in their simulated proton beam on earth. Localized shielding for consumer equipment is more than enough, especially combined with ECC.
Solar power and power dissipation are similarly solvable. The only bottleneck currently is launch costs which is why it’s expected around 2030 is when these DCs will explode
> Solar power and power dissipation are similarly solvable
So it is not outright stupid, just has a few extra costs.
So again, what's the economical benefit? Because you surely have no power and dissipation problem down on Earth, so why add a whole rocket launch into the equation? Because big fire goes brr?
Because data centers take a long time to get through local regulations. By comparison launching a bunch of satellites into orbit is much simpler and you don’t get local political opposition you have to spend a lot of resources trying to overcome (not to mention secondary cost effects of needing a lot more staff to manage all the diffuse lobbying that needs to happen). Any political opposition to space DCs is going to be more diffuse and disparate since any problems are non local. There’s also that cooling is logistically easier in space and thus cheaper as it’s a small amount of mass vs the huge cooling towers since they’re cooling a rack directly instead of needing to cool the entire DC.
So what’s the economic benefit? As Suncatcher has predicted, the economic cost benefit flips to space DCs at a certain launch price/kg which is expected to be reached commercially by early to mid 2030s (spacex will be ahead because they don’t need profitable launches and can get ahead).
We surely don't expect the same total computation throughput from one data center and a couple of satellites, right?
Like even if a single satellite launch is cheaper than.. building a whole building, I have a really hard time believing that putting that many satellites up would still be cheaper.
And even that hypothetical depends on that we need some very quickly growing need for compute, which somehow can't be managed by.. simply replacing the existing compute hardware every 5 years in data centers, and the relative slow build of new ones?
Space is big. We’re not talking about a couple of satellites eithers. It’s probably many tens of satellites in close proximity. Enough to host a single LLM model for example. Then you put up enough of these micro DCs. There’s simply no need for a single constellation with thousands of racks like a terrestrial warehouse. The only reason you do it that way today is purely because that’s how permitting and construction makes sense on earth, not because there’s some inherent advantage.
You can surely just put a single rack into every 10th office building in a city of your choice for faster connections, significantly larger compute per rack, and significantly cheaper price.
What's the point of going to space besides a scammer wanting his share price up?
And before you say "politics" and "legal", space is just not that regulated because there was no need for it. Why do you think it will not get regulated once "per kg transport" becomes cheap enough? How is the politics and legal all that bad for the data center case? Like maybe don't build it into stupid places harming the environment, but it's not that bad of a thing all together, no reason why they couldn't be built under normal circumstances.
I imagine the cooling and power challenges of that is what inhibits this idea. But if it were economical, why don’t you try building a startup around it and find out? It’s easy to argue from the couch, much harder to actually validate your ideas against reality
It's difficult to imagine any sort of compute in space given the state of the art rad hardened processor is about as fast as a computer Apple sold in 1998.
You don’t have to imagine. The starlink constellation is a working and profitable massive deployment of orbital compute. (It doesn’t use rad hardened hardware as far as I can tell)
Starlink satellites have such a low orbit that they experience non-negligible atmospheric drag, which is why they have to be replaced at such a rapid rate. All proposals for orbital datacenters propose placing them in sun-synchronous orbits for maximum solar exposure, which puts them right at the edge of the Van Allen Belt at best, meaning they'd be exposed to dramatically more radiation than a Starlink satellite.
The PV at least can potentially be much less robust than on Earth, since it doesn't have to contend with wind and weather. CdTe thin films, for example, can be just 1 micron thick (the absorption length of solar photons in that material is ~0.1 microns). So potentially the material cost of space PV can be much less than the same peak power PV on Earth.
I'd imagine sometime in the future we will build processing wafers similar to monocrystalline solar panels. Data processed sequentially over a couple dozen feet. AI models hard coded into the silicon itself. The entire data center maybe a milimeter thick sheet.
The bet that they are making is that, in the near-future, energy on Earth will be such an issue (especially if you want to scale) that DCs on Earth will cost more than components + launch costs
Why would energy be an issue? The Sun is shining, wind is blowing, the ocean has waves, nuclear is nuclear. All of these will just get more efficient in using the same constant resource.
It's not the energy per-se, it's the access to it. You mention solar, solar in orbit is constant, on Earth it's not. You mention wind, it's not pilotable. You mention ocean, it means you need to build close to an ocean. Nuclear yes, but it's expensive. And then there is the grid, and there is the cooling.
Finally you can sprinkle all of that with political things like putting datacenters in residential areas and forests.
But these might also get solved, hence why a "bet".
Musk's "genius idea" there was to (pardon the pun) derail high speed rail and other similar investments so he could sell more cars.
What are Hyperloop/Boring actually doing now? It doesn't take that much investment to human-drive a 10-20 Teslas up and down the Vegas Strip in a straight tunnel at normal speeds.
The term 'Voluntary Exit Offer' reminds me of the suicide drug Quietus in P.D. James' Children of Men, which the government offers to people who no longer wish to live.
30% was the quoted number for the acceptance rate of the earlier VRP, which was mostly long tenured employees who were at retirement age, or close to it. I would be surprised if this gets a significant acceptance from the younger employees. OTOH, older employees who didn't accept the earlier VRP and were regretting that decision may jump at this offer.
C++ is probably the largest, most complex programming language to ever exist. I, as a former C++ compiler developer, find it hard to keep up with the language changes. What chance does an average user have?
To use the subset that makes it a better alternative than C, aka "Typescript for C", and ignore the rest.
Hopefully we would have moved on into fully bootstraped alternatives, but open source has decided to build their alternative languages on top of LLVM and GCC instead, with some minor exceptions.
Note that I am starting to have a "find it hard to keep up with the language changes" similar feeling with Java, C#, Python and TypeScript.
This is already a good norm. A substantial portion of the new language features really aren't meant for business logic but instead help library developers anyway.
I have the same opinion. C++ was my first programming language I used at work. I feel that I was relatively proficient with it, but it was around 2005 I think. Since then I've mostly used other languages. Right now I feel I can use C with ease, in fact I periodically use it. But C++? I don't even dare.
I guess it's fine if one specializes on C++ programming only. But when one needs to use different languages and C++ is used sporadically, I feel like it might be substantially harder than other languages.
The problem with C are the parts that C++ already provides an alternative, regarding bounds checking, generic code with preprocessor tricks or now that _Generic thingie, stronger typed enumerations, heap allocations without sizeof math, automatic resource cleanups.
I would feel the opposite. I dare not write a program in C. The mentality of RAII and destructors for resource management is so ingrained that attempting to track resource use and lifetime in C would be nightmarish.
It's a very complex language, but I also find little reason in using any of its more niche features; however, I essentially just write C with RAII and templates when it comes to C++, so I'm probably abnormal.
For instance, I can now avoid SFINAE in C++20 and beyond, which was always the ugliest part of C++ templates to me. C++20 concepts simplified most of my metaprogramming needs.
Reading boost libraries was like reading brainfuck. In fact most of std lib source felt that way too. I never reflected on this since I stopped doing C++ 15 years ago - but yeah I have not worked with a language that's worse to read since. Maybe macro heavy lisp ?
Extreme accretion/scope-creep of syntax (and features) across all languages continues, C++ is probably one of the most extreme example of this.
Everyone's trying to bring in the shiniest new toys from all other languages. So-called "modern C++" is an unbearable menu of every possible flavor of procedural, functional, OOP, templates, strong/weak typing, metaprogramming, and so much more. End result is a Frankenstein mix of C, Fortran, Pascal, Python, F#, Scala, C#/Java, Visual Basic, LUA, TCL, Rust, and what not... and that's before we even talked about the STL/STD situation in C++.
The only correct way to work with C++ is to agree on the most minimal subset of features needed for consistency and productivity in your codebase, stick to those and avoid adopting 95% of the crap that's being brought in. Ignore the noise.
Sorry for the nitpick, but ossification means the opposite of how you're using it here. It means existing structures / practices becoming frozen in place, by analogy with real ossification which is where flexible cartilage hardens into rigid bone.
I've never used LLMs for C++ but I notice that it likes to use every available feature on other languages. Is Claude-generated C++ incomprehensible as a result?
The way to deal with a bad HOA is to get on the board. Normally, HOA boards attract the most petty minded people because most of the normal people living in the community are too busy but if a bunch of like-minded people get on the board, they can bring sanity back.
> if a bunch of like-minded people get on the board, they can bring sanity back.
This sounds good, but I don't think it generally works. The problem is that the petty minded people are also the ones who don't have anything else to do with their time, so they can expend all their effort on being petty. Whereas the normal people have other things to do with their lives, and simply don't have the time to deal with all the pettiness, and if they end up on the board, they now don't have a choice: their lives are basically hostage to the petty people who keep taking up their time with bogus issues that can't just be ignored because the structure of the HOA requires some kind of response.
They also have a manual, non motorized version. I am building this currently for eventual use, with modifications, in a point diffraction interferometer.
reply