HNHacker News
TopNewBestAskShowJobs

sobellian

1,364 karma · joined February 23, 2014

submissionscomments
sobellian··on The Elizabeth Holmes documentary “You Can See Everything” is deceiving
> I cannot say what was missing from the trial because I do not know what was presented there in detail. Most of those claiming utter certainty, know as little as I do.

Okay but can you see how this is a different point than, "The access to the information required for that is not available to most people, we have no basis to judge the narrative for accuracy"? The latter implies that most people literally could not obtain access to the necessary information. The former states that you have not looked into it that much and you doubt others have either. The trial is public record; it is available to everyone.

sobellian··on Tesla Roadster
Your math is pretty good, Tesla actually gave the storage tank capacity as 250kg

ETA: this is a horrifying amount of energy to be sitting on top of.

sobellian··on Why are we even doing all this?
> This aligns with my own lived experiences — I didn’t become a theoretical physicist to learn about the universe or better appreciate the beauty of nature myself (that’s philosophy, which I would never be a doctor of).

If one truly philo's sophos, then wouldn't one at least be curious what the computer says about a difficult problem? Like, if you've been wrestling with an especially troublesome integral for years then I feel like that somewhere along the line you should have plugged it into a CAS - even if you want to clean up the derivation afterwards it would be foolish to ignore every tool at your disposal and even rail at the Mathematica authors for the temerity to write a symbolic integrator.

I feel that the article is so preoccupied with clever sarcasm that it doesn't really engage with nuanced views and is instead happy to just spend a few KB of text tearing down vibe research strawmen.

sobellian··on Imposing Sanctions on the International Criminal Court
Well I think you hit upon an excellent point and one that I'm rather close to, which is legalization. Adopting a uniform procedure to deal with various cases and accord defendants some recourse is admirable. That doesn't make it law, but legalizing those types of proceedings is still valuable. I think Nuremberg rested on shaky grounds as a court, but it was still preferable to either summary execution or disclaiming any punishment on the accused.

Likewise I think where I'm uncertain with Putin is that while he's undoubtedly a nefarious guy and we have written down rules and norms he has broken, those norms do not apply to the Russian world. We claim they apply, but we could only make them apply through force of arms which runs opposite to the rule of law. That is rather the rule of "might makes right."

sobellian··on Imposing Sanctions on the International Criminal Court
I don't really think there's a need to be so hostile. I don't really know how I feel about subjecting international citizens to US law to be honest with you. Certainly in some cases it's appropriate. But in others it seems strange to me.

And undoubtedly I agree that if Putin or Netanyahu pass through Europe they should be arrested. But I'm not sure if that's a legal decision or a political one.

sobellian··on Imposing Sanctions on the International Criminal Court
Right and you can imagine how much authority the average Russian lends to an American arrest warrant.
sobellian··on Imposing Sanctions on the International Criminal Court
If the point is that anyone can tell anyone else anything, that's vacuously true but a court typically wields the power of a state and its code of laws. For treaty countries ICC is perfectly unproblematic but for nontreaty countries there's a strong difficulty in calling it a court as it bears no legal legitimacy.
sobellian··on Imposing Sanctions on the International Criminal Court
Seems to be a rather silly, pedantic point if it even passes. GP was clearly talking about sovereignty etc. which is exactly the caveat that you granted - that ICC may have no enforcement power over non-treaty countries. For the record I have no axe to grind with the ICC, but you and GP literally agree on the main difficulty facing the institution.
sobellian··on Imposing Sanctions on the International Criminal Court
> Anyone can decide anything, the question is whether it is on their power to enforce their decisions.

You seem to have arrived at a suitable definition of "authority"

sobellian··on “Math 2.0” will need to value mathematical progress more holistically
Current models have undergone massive amounts of RL to "get shit done." There's incentive toward reasoning, but probably little toward elegance. Every engineer is familiar with the impulse to just hack things together.

But this need not be a permanent state of affairs. The model labs are probably already examining the elegance problem, since it is such a common complaint of researchers reading these machine generated proofs and papers. And I haven't seen anything disproving the notion that elegance can be RL'd.

sobellian··on Sharing AI progress in mathematics
Ah thanks for pointing this out, for some reason I had always equated radix sort and bucket sort (with 2^k buckets) in my head. But I learned today that this isn't true!
sobellian··on Sharing AI progress in mathematics
If you sort n k-bit items for a total time of O(nk logn), that scales more poorly in n than multiplying n-word integers. Of course if k is constant you can do radix sort, but I genuinely don't know under what conditions radix sort is more/less galactic than this multiplication algorithm.
sobellian··on Sharing AI progress in mathematics
I am fully braced for it to be a https://en.wikipedia.org/wiki/Galactic_algorithm

Very surprising result though! Multiplication is easier than sorting.

sobellian··on OpenTPU – An open-source AI accelerator, developed by AI
The repo claims they tested with an FPGA. To be fair it is not the same as an ASIC, and we have no idea what errata abound, but it actually did make a thing that spat out tokens.
sobellian··on The Future of Mathematics
There are plenty of very difficult math problems with practical application. Just because they don't interest some subset of academic mathematicians does not mean they don't exist. And you don't have to literally invent a field of mathematics to do valuable work applying it (and applying it may be difficult).
sobellian··on Our Project Suncatcher prototype satellite is in orbit
The radiators are what need to be very hot. IIRC it works out to, order of magnitude, a similar area to the solar panels. These satellites would likely be placed in SSO, a polar orbit that precesses at the same rate that the Earth orbits the sun. So the fraction of time they see the sun would be very high. Cosmic rays are a little bit of an unknown, we don't know exactly how a commercial GPU running an LLM deals with cosmic ray radiation. I think the intuition is that it would be very little because a single bit flip in that vast array of matrix math and many times its result might not actually change any activations anyway.

I think the big problem with the idea is that GPUs have a failure rate and even if they didn't, they become obsolete. Most orbital DC plans are really more like flying server racks with no servicing in orbit. So when the GPUs die the satellite is a flying brick. All the power equipment, all the thermal equipment, all the comms equipment now depreciates at the same rate as GPUs. Very different economics from terrestrial DCs. And that's all assuming that you can launch everything up there quite efficiently. And the satellites take time to engineer but DCs are a more known quantity.

sobellian··on Google’s Project Suncatcher to put ML infrastructure in space
Upthread comment and therefore the premise of your $8/W target is wrong. As I explained, I plugged in all the announced engineering parameters for Starmind, $8/W, and $76/kg launch. It ties terrestrial over a 5yr period (the slider default) but gets crushed over a 10yr period. Naturally a longer time window is prudent absent some world-ending calamity in 5 years that renders the need for compute irrelevant. And even the 5yr tie is very sensitive to these parameters and requires everything to go right.
sobellian··on Gemini 4 Argon
Hmm, is it though? Compilers relieve agents of the burden of much symbolic work, like many other tools that give agents value over a pure LLM.
sobellian··on September 2026: The world today, as seen by one Polish guy
I'm not going to say that NATO materiel has had zero effect but I think what Russians (at least those in power) really can't reckon with is that their initial invasion got stopped by Ukrainian AD, artillery, armor, and soldiers operating on the same Cold War stockpiles and doctrine as they did. The single most valuable intervention from NATO was simply alerting Ukraine that Russia was about to invade, allowing the Ukrainians to disperse their units.

Russia committed to too many (and often poorly chosen) lines of advance and did not have a plan to keep their units adequately supplied and moving on the approach to Kiev. This meant even in the opening month of the invasion, their plan went south almost immediately and you had infamous miles-long traffic jams on the roads leading south. Their SEAD didn't work. Their soldiers were not told ahead of time that they were invading Ukraine. They didn't want to mobilize ahead of time, so they didn't amass enough forces. Against a large country like Ukraine you don't need to invent a NATO bogeyman to understand why the initial offensive failed. Credit to Ukraine, they did not allow a defeatist narrative to take hold and fought.

sobellian··on September 2026: The world today, as seen by one Polish guy
I am curious how Russian people feel about this propaganda line that they're fighting NATO. It's obviously nonsense, along the lines of "war is peace." Russia borders NATO and there is no front there. Russia invaded a non-NATO country and immediately bogged down against an army that does not fight according to NATO doctrine. I have to believe this is something the average Russian realizes.
sobellian··on Google’s Project Suncatcher to put ML infrastructure in space
Okay so even assuming all of SpaceX's numbers, which is extremely optimistic, even if we're an angel's advocate and assume that they spend all this money on engineering and optimize their launch costs and their satellite costs and they don't hit any hiccups, and DCs on Earth stay pretty much the same and don't get any added efficiencies. Even then over every time period orbital is still more expensive. And it gets particularly punishing the longer you extend the analysis period as the satellites need to get replaced while the DCs don't spontaneously combust all their infra.
sobellian··on Google’s Project Suncatcher to put ML infrastructure in space
It's reasonable to 0th order to assume a ball of heat management and phased array antennae has a roughly constant cost per kg, yes. And as I said starlink absolutely wants to use as much power as they can, as communications from space are heavily limited by power.
sobellian··on Google’s Project Suncatcher to put ML infrastructure in space
The calculator I've cited elsewhere uses https://chatgpt.com/share/69391474-4b24-8005-bb93-ebd4340c65.... The various starlink versions have approximately the same cost per watt across very different power levels. Much of the cost is going to be ~proportional to power - solar panels, heat rejection, power electronics, etc. Listen, if they could make it cheaper to deliver the same amount of power they would. Starlink is approximately all of SpaceX's business.

ETA: power is absolutely important for Starlink. SNR is very important for shannon capacity and satellite systems are often limited in this respect.

sobellian··on Google’s Project Suncatcher to put ML infrastructure in space
Commercial GPUs already fail a lot and that's before you put them in LEO. Starlink is doing something that can't be accomplished anywhere else than space. If you could somehow have ground stations that can give wireless internet to every corner of the Earth you would much rather do that. Again, this is all physically possible, but the question is whether it's wise, whether it's saving money or if it's just a giant Rube Goldberg machine. If you assume Starlink satellite $/W then even if launch is zero, it costs twice as much as terrestrial using assumptions charitable to orbital.
sobellian··on Jevmem – automatic project memory for Claude Code, built on Jev
My pet theory is that while the pretraining -> RL pipeline achieves very impressive results, it does not reward clarity of thought or elegance. It's not obvious whether it even should for most tasks, but it does grind on me as a human who needs elegance in order to keep everything under control. You give astra/codex many tasks, it retires them all more efficiently than I could by hand. But you look under the hood and every bugfix is another codepath, it just hammers away at things with admirable persistence and vigor until the tests pass. Similarly in discussions and docs, I've noticed many LLMs like to "beat around the bush."
sobellian··on Google’s Project Suncatcher to put ML infrastructure in space
Yes it has to advance a lot. In a DC all your supporting equipment has returns to scale and can be repaired if it breaks. If a GPU breaks, a sysadmin walks over to the offending rack and swaps the card. In space all that equipment serves just a few cards (this is more like orbital server racks) and it has to work in space (so instead of using an ~infinite heat sink like the Earth, you use radiators etc). We still don't exactly know what effect radiation in LEO would have on stock GPUs - IIRC the experiments to determine this started after Elon went all in on orbital DCs. If anything breaks, you have a flying brick.

This is why that calculator, even under extremely optimistic assumptions for orbital, and even if you assume launch is zero, still cannot make it competitive with terrestrial.

sobellian··on Google’s Project Suncatcher to put ML infrastructure in space
Note that this calculator is actually quite optimistic for orbital wrt. many things including cooling and effect on launch, as:

> No additional mass for liquid cooling loop infrastructure; likely needed but not included

> Thermal: only solar array area used as radiator; no dedicated radiator mass assumed

In hardware and mfg. solvable vs. solved is a big difference. And I too believe that SpaceX's engineers know about radiator panels. But the more cynical interpretation is that whatever the SpaceX engineers think about the technical merits, they are not being asked for that. They are just being asked for a pretext that justifies the xAI acquisition. Elon is also discussing lunar satellite factories that launch the satellites via railgun. Now, is this physically impossible? No, that isn't physically impossible either and I will seriously defend the physical possibility of this. It's not going to happen though.

And you could spend all the engineering costs on building some seriously efficient terrestrial DCs, but somehow all these analyses start with "assume that launch and satellite technology advances manyfold and terrestrial DCs stagnate or become less efficient, then if you squint the two numbers get kinda close."

sobellian··on Google’s Project Suncatcher to put ML infrastructure in space
Okay any argument about why space is uniquely challenging is going to revolve around physics. Sure it's not literally physically impossible, but we need to explain to people why this is different from shipping the GPUs to Ohio.

If you want math then https://andrewmccalip.com/space-datacenters exists. The numbers are grim for orbital DC. Even if you drag the launch cost slider all the way to $1/kg (by the way this is literally sci-fi, per ChatGPT air freight of semiconductors from Taiwan to Ohio costs $9/kg and ocean/train freight costs a bit under $1/kg for a reasonable shipment so good luck with $1/kg to LEO this century) it is still more than twice as expensive as terrestrial DCs.

sobellian··on Coulomb's law remains tricky to test at home
For two different indivisible charges, this is a bit slippery but I think a=1 by definition. How do we measure charge, practically speaking? By how much force is measured between it and a reference charge. So we take a=1 as a convention. But no experiment can disprove a=2 for indivisible charges as we would simply obtain charge through new units. For assemblies of charges the forces must add linearly due to conservation of momentum, so there we know a=1.
sobellian··on Claude discovers a novel enzyme system with CRISPR-like repeats
The cure for HER2- metastatic breast cancer is a simple matter of [intensive well-funded research]

That wasn't too hard, maybe I'm superintelligent?

← PreviousPage 2 of 16Next →