Hacker Newsnew | past | comments | ask | show | jobs | submit | eldenring's commentslogin

I disagree, I think distillation hides the true technical advantage since you can always get within X% of Claude by ex: training on newly created code on Github, but the underlying intelligence is harder to replicate.

based on the behavior of the chinese labs it seems like training on chain-of-thought is extremely important for distillation, and you can't get that by training on llm generated text.

What specifically do you disagree with from the article to say its Smoke and Mirrors? I think their analysis is reasonable and that it will reduce to a launch cost problem.

Giving it a quick look they're doing the same thing as everyone else and handwaving away really difficult engineering problems with some back of the envelope math that only looks at one or two aspects.

Like their section on collision avoidance isn't any more sophisticated than "Starlink has a great record so far" when we're talking about a very different scale of how much mass and surface area has to be moved around in a coordinated way.

This is in effect an editorial from two postgrads, not an actual detailed workup. I don't feel like spending time going into any more effort based on that.


Saying it is equivalent to launch costs is hand waving. Their "median" scenario would involve Starship working at the lower bound of their operational goal for it.

makes no sense, he isn't some recent hire you need proxies like that to judge. He is the most successful ceo of all time, arguably, so I don't think hes short in trust to run Nvidia well.

Possibly knowingly selling your product illegally when you’re instructed to not do that, as well as conducting numerous affairs with direct reports cannot be defined as running the company well.

Even leaving out the sale to China part, sleeping with direct reports is unethical and presents so many conflicts of interest. He was lucky the damage was minimal for so long


I dont understand this argument, its either a shareholder concern or a criminal concern? a "conflict of interest" is a term used when a 3rd party needs to be unbiased ex government or law. You wouldn't say its a conflict of interest if your favourite sports team trades their best player

Do you make money when your favorite sports team does well?

Being a Manchester United fan, apparently not

imagine you run into ub in your C program. Suddenly every single loc of your project, including dependencies, are in scope.

With Rust, only the unsafe blocks are in scope. The point of Rust is not to eliminate unsafe code, but to make it not scale with the size of your program. There is no good reason for those to be coupled.


I tend to wonder, in an alternate reality, where the buck stopped at C for systems languages, would we be better off because more people would use it and know how to fix memory related issues?

We’ve sent rockets into space running C/C++ and assembly. And that was long before C++11 rolled around.

I’m wondering because: man Rust is an ugly language. C is not perfect either, but it’s still pretty close to what it was FIFTY YEARS AGO.

Fast forward to now, and all we’ve come to is Rust? I just wonder why don’t we just keep using C, C#, and Java? I mean the former two are memory safe, and they can do “unsafe” just like Rust.

Another thing: why do we write pseudocode and not design languages to look like pseudocode? With Rust and modern C++ compile times you’d think the machine is parsing prose and turning it into assembly…


> I tend to wonder, in an alternate reality, where the buck stopped at C for systems languages, would we be better off because more people would use it and know how to fix memory related issues?

No, because what people don't realize is there is a constant influx of people who are new to C and don't know how to use it and how to fix memory related issues, so you can't just train everyone to not do them -- newbies by definition are untrained. The only winning move is to encode best practices into the compiler to make it hard for newbies to load the footguns, and only put them behind a bunch of warnings and features that experts can navigate.


over-obsession with the language's aesthetics ignore that the type system / borrow checker / ... are good and legitimately assist with development and did so even before the widespread use of LLMs.

many of those who ended up as Rust developers came from C and C++, seeing improvement from what they were using.

languages have been designed more and more to allow for higher level behaviours, in Rust's case, although not pursuing psuedocode, generally allows for very high level abstractions. regardless the computer still has to run things, languages are typically deterministic in their handling and psuedocode leaves much undetermined, unfully specified.


So then why not use C# or even F# for language abstractions? Or clojure on JVM?

Just playing devils advocate it seems higher level behaviours don’t belong in a low level language.


one of the best parts of building up lower level pieces of work/abstraction is that you can eventually produce a higher level interface to them that makes working with them entail a lot less overhead to you or someone else. (that said, in any language it's hard to make said interfaces not have holes or be painful somewhat to some subset of their consumers!)

in some respects, rust is not all the way there for high-level work, but great strides have been made to make it pretty darn close.

a talk from dioxus on the state of high-level rust: https://www.youtube.com/watch?v=Kl90J5RmPxY

having a single language to do your work in, although i personally don't subscribe to it (and am a fan of clojure, guile, C, C++, Rust, Zig, ... all without this feeling to be a contradiction for me) minimizes cognitive overhead and also reduces the quantity of interfaces you need to concern yourself with to manage interoperability. the borrow checker and type system to a seasoned rust user are not a bother regardless of the work they're doing, so it's not like that presents a unique problem.

avoiding garbage collection, avoiding the CLR, whatever overhead comes with it and the joys of .NET Core or overheads of clojure + jvm compiles can certainly be worth it (a clojure project of mine spent 33% of its runtime on java+clojure compilation!)

the lower level work being performant means that your well-designed interfaces can also run well.


> I tend to wonder, in an alternate reality, where the buck stopped at C for systems languages, would we be better off because more people would use it and know how to fix memory related issues?

I think that period has arguably already come and gone. C and C++ have been the most popular systems languages for decades. I think it's clear that mere popularity is not sufficient to produce memory-safe software with those languages. If anything, those languages are experiencing somewhat of a renaissance these days and it's these more recent developments that end up producing improvements with regard to memory safety.

> I just wonder why don’t we just keep using C, C#, and Java? I mean the former two are memory safe, and they can do “unsafe” just like Rust.

Bit of a nitpick, but I think you meant "latter" instead of "former".

That aside, there's more to picking a language to use than just memory safety and the ability to perform "unsafe" operations. Take your pick from more technical reasons (zero-overhead FFI to C, low-level control over data structures/operations/runtime, etc.) and more "soft" reasons (language affordances/features, ecosystem, etc.).

> Another thing: why do we write pseudocode and not design languages to look like pseudocode?

Perhaps COBOL might be what you're looking for? :P

As with many things, language design is full of tradeoffs, particularly when you want to produce something that is not too unfamiliar to a certain audience. IIRC another factor is that languages that want to work well for "experts" will tend to favor different points on the design spectrum. Consider APL-family languages - those are completely incomprehensible to anyone not familiar with them, but that same terseness is highly valued by proficient practitioners and is something pseudocode-like (or even "regular" code-like!) could never replicate.


All fair points, and yeah I meant latter. :)

It’s just interesting that ALL of these languages have a C compatible interface built into the language. That’s an interesting fact and says something about C that no other language can really claim. Nor will any other language claim that any time soon.

I’m not looking for anything in particular. It’s just that I’ve seen and used a lot of languages over the years, and I’m not wholly sold on the “memory safety” of Rust when something like Python can just put the performance sensitive parts in C and have safety for everything else.

Most languages are memory safe.


> and I’m not wholly sold on the “memory safety” of Rust when something like Python can just put the performance sensitive parts in C and have safety for everything else.

> Most languages are memory safe.

As I said, there's more to Rust than just memory safety. If memory safety were the only thing Rust offered then I think it wouldn't have garnered nearly as much attention as it has. It's memory safety on top of everything else that makes it as interesting as it is.


> Reducing shortcuts and loophole exploitation: With Carbon, we further mitigated shortcut attempts in training while ensuring high-quality coding performance. Carbon reduced successful cheating on Terminal-Bench 4.0 to 0.0%, with attempts to cheat dropping to 0.4% — all while improving overall accuracy.

This is probably the single most important thing to make frontier models usable right now, so I'm happy there seems to be some progress. This was the previous Gemini's biggest problem IMO


I don't think that is necessarily true, the post is making a stronger point than what you describe

> Indeed, the algorithm requires you to multiply complex numbers whose magnitudes grow superpolynomially with n. If you stored these numbers as bits, the memory requirement would effectively be exponential in the input length. It assumes you can multiply such arbitrarily large complex numbers together in a single operation.

Another way to think about what the author said, imagine you had access to an oracle which could execute arbitrary precision arithmetic expressions in constant time. You could encode a wide range of computations in these and execute impossible algorithms.


Replace the complex numbers with a finite field and O(1) arithmetic stops being an unrealistic assumption, as you can actually perform the required operations in constant time on a real computer. You can't do that for infinitely-precise complex numbers, but you'll never be able to provide an infinitely-precise complex number as the input anyway. If you start with finite-precision numbers (e.g. 64-bit floats) you can do all operations in a finite-field representation (e.g. the Ozaki scheme II is an actually practical implementation of that idea for matrix multiplication https://arxiv.org/abs/2504.08009 ) and the magnitude of intermediate values is no longer a problem.


Are you certain that applies here? The length of the intermediate numbers grows superpolynomially, so you would be sure to overflow fixed width numbers. Floats are not a field.


Yes, floats are not a field, which is why you don't use floating point except to convert to and from a different representation. Overflow is expected and cancels out as long as the final result is uniquely representable. See the paper I linked, which implements 64-bit floating point matrix multiplication using 8-bit numbers with wrapping overflow modulo a set of small coprime integers. The result is reconstructed at the end using the Chinese Remainder Theorem.


This reminds me of 'O(1)' hash maps, but far more egregious.


What's wrong with `O(1)` hash maps? I'm not aware of any underhanded tricks related to, for example, cuckoo hashing.


Hash maps are only O(1) for sufficiently small n. You need O(log n) bits just to write out the hash.


It's not any faster in practice to write a byte than a word (log(n) bits)


That depends on the scale. A couple jobs ago we had enough objects that an 80-bit key wouldn't have had a low enough collision rate (better models of computation for that domain include sqrt/cube-root ideas to account for information propagation in physical space). Sometimes you have to use more constrained devices so that your max effective constant-time word size (for the same computational cost -- not to quite be confused with the CPU's "native" word size, whatever that means in your circumstances) is 8, 16 or 32 bits (or, most recently at my last job, for _reasons_, 11 bits), making even quite small tables not quite O(1).

Many compsci researchers like the word-ram model of computation to avoid a bunch of extraneous log factors and to apply exactly the intuition you're describing, so that hashtables are actually O(1). That's not a bad habit. It introduces extra log factors in other places though, so I think it's a bit of a wash, and given that it doesn't model the actual register machines we use it feels a bit intellectually dishonest to me.


actually thinking about it more I think my oracle model is more powerful than the computer used in openai's fft but maybe its a good way to think about it.


Reflection AI is a completely different company with no relation (as far as I can tell) to the model you are referencing from 2024


If that’s the case i take it all back. I thought this was another venture by Matt Shummer’s Reflection Ai


Cooling is relatively easy, you just need radiators which are passive, and essentially reduce to a launch cost penalty. You are right that they can't be serviced, but that is missing the point of orbital data centers. The whole point is that you can build hundreds of thousands of these in a factory and launch them in a scalable manner. The power, cooling, etc. comes for "free". In the long run, as the cost of the chip, launches, etc. goes down, orbital data centers will scale better terrestrial ones.

As a side note, I don't understand why I keep seeing these wrong arguments on HN repeatedly. Like everything mentioned in this thread can easily be fact checked. Radiative cooling is solved, launch costs are going down, so power costs will pay themselves back very quickly, etc.

You can argue about specifics, like chips will get more sophisticated + power efficient and fabrication will be the true longterm bottleneck, or SMRs/fusion could reduce energy bottlenecks, but talking about cooling as if convective cooling is the only option is just nonsensical.


"you just need radiators which are passive

"Just" is doing a lot of work there. SpaceX is planning to launch 100 GW of compute annually, that comes with ~ 2.5 square kilometers of radiator (assuming an optimistic 800K radiator temp and emissivity of 0.9, double sided)

Go for advanced carbon composites, you can do that with just 5,000 metrics tons or so of material. That's 34 starship launches just for the radiators. We haven't solved assembly, we haven't brought up power panels or core compute. Planned launch cadence that SpaceX hopes to reach end of this year: 12/year.


radiators about the same size and weight as solar panels will do the trick

there is already a h100 in orbit

1GW of compute is a lot in 2026. comparing 100GW of annual compute to SpaceX 2026 goals does not make sense

if Starship launch cost predictions are accurate, data centers in space will happen within 10 years


> The numbers are brutal. Starcloud’s own white paper estimates that a two-sided radiator held around 20°C would emit only about 633 watts per square meter, over 1,000 times slower than water cooling of AI chips on Earth. So, a puny 1-megawatt orbital data centre, 1,000 times smaller than the gigawatt scale of hyperscale data centres on Earth, would need about 1,600 square meters of radiator, an area roughly the size of a hockey rink.

1GW needs a pretty big area for radiation.

And in space your data centers is hard to defend against foreign actors


chips run way hotter than 20C, and radiative cooling scales to the fourth power of temperature. check the math


Don't you think a company like Starcloud has a reason to put 20C in their white paper?


Yes, to get funding.


The whole point is that you can build hundreds of thousands of these in a factory and launch them in a scalable manner. The power, cooling, etc. comes for "free". In the long run, as the cost of the chip, launches, etc. goes down, orbital data centers will scale better terrestrial ones.

How is this not true for terrestrial/ocean deployment as well? It will ALWAYS be easier to shed heat on Earth than it will be in orbit. Convection, conduction, radiation... in space, you only get the last one.

So take the same unattended hardware you were going to launch into orbit, and put it on a container ship instead. You'll be better off in every way. If it makes you feel better, lie and tell everyone that you launched it into space. Building orbit-capable data centers might make sense, but actually launching them never will.


> and essentially reduce to a launch cost penalty.

Are you arguing that all this is technically possible or something? The whole point is that the costs would dwarf the gains.

> You are right that they can't be serviced, but that is missing the point of orbital data centers.

Pointing out a downside of something isn't ever "missing the point".

> The whole point is that you can build hundreds of thousands of these in a factory

In an Earth-based factory, right? Am I to understand that we can't build hundreds of thousands of regular Earth-based datacenters in a factory?

> and launch them in a scalable manner.

Wanna bet that launching something to space will always be a few orders of magnitude more expensive than shipping it somewhere across the planet?

> The power, cooling, etc. comes for "free".

Unlike on Earth, where you pay for sunshine? Or is cooling "free" in space but not on Earth? Lol?

> In the long run, as the cost of the chip, launches, etc. goes down, orbital data centers will scale better terrestrial ones.

The costs of the chips will get lower in space than on Earth?

The costs of launches will, again, become cheaper than terrestial transport?

> As a side note, I don't understand why I keep seeing these wrong arguments on HN repeatedly. Like everything mentioned in this thread can easily be fact checked. Radiative cooling is solved, launch costs are going down, so power costs will pay themselves back very quickly, etc.

The question isn't whether this is physically possible, but why you'd want to do it instead of an Earth-based datacenter. It's all downsides basically.


My back of the envelope maths:

Suns energy at ISS is about 1.4KW/m2 Solar panels about 35% efficient but let’s say 50% for fun

700w/m2, or about one H100 worth per sq metre (hey, I could run my own H100 off a roof top panel !!)

We want a small 70MW data centre - which is 100,000 times the size so 100,000 m2 or an array 316mx316m or 15 football pitches

Then as it’s energy in and energy out you need radiators on dark side of same size

The ISS is ~ 2000m2, so that’s fifty ISSes

I mean it’s physically possible. But the engineering, the space launch costs they are staggering. And the upside is … Im not sure

All the win seems to be is free sun energy, but a data centre in Texas or Nigeria just needs about twice the solar panels and some big ass batteries.

Im not costing that out but, honestly it seems like a marketing pitch or a really obscure need to put compute beyond the reach of governments.


> Radiative cooling is solved,

By that logic, climate change is also solved, just built a giant radiator.


> Radiative cooling is solved.

This is emphatically not true at any scale in which this scheme makes sense. Be careful with including too many Musk boosters in your information diet.


Did you read the whole thing? he is clearly talking about having lost focus which is directly in agreement with your take on the Metaverse.


If that was the case then he wouldn't have continued doubling and tripling down on these same product decisions for many more years.

Ask anyone at Meta and they will readily agree that Boz is among the most incompetent and unserious leaders in tech. His only qualification is being Zuck's friend, and that shields him from all criticism. At a more "focused" company he would have been fired a decade ago. As it stands he lost the shareholders $150B+ (and counting), dipped from the metaverse org, and got rewarded with a new role overseeing the shiny new thing - AI.


> As it stands he lost the shareholders $150B+ (and counting)

When you put it like that, I kinda warm to the guy


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: