Petabytes of training data is only one application of PyTorch, which is going to use tens of thousands of containers, but...
Inference, development cycles, any of the application domains of PyTorch that don't involve training frontier models... all of those are complicated by excessive container layers.
But mostly dev really sucks with writing out an extra 10GB for a small code change.
Going to self promote one last time here - I've built a fix for this, at least for the registry/image export side, at https://clipper.dev. Docker(Hub) can't share large files between layers, but I can.
You don't even need MB of training data for some ML applications. AI is the sexy thing nowadays, but neural networks (Torch is a NN library) are generally useful for even small regression and clarification problems.
For some problems you might even be able to get away with single digit numbers of training points (classic example of this regime being Physics-Informed Neural Networks)
Yeah, our handful of models we just commit to the git repo--usually only a few MB.
Image still ends up being like 6-8Gi tho. iirc pytorch had a hard dependency on CUDA libs which pulled in a bunch of different hardware-specific kernel binaries. The models ran on CPU and didn't even need CUDA but it was incredibly hard to remove them--there was some pytorch init code that expected the CUDA crap to exist even on CPU-only.
If you implement threads and code that reacts to an input queue (e.g. PostMessage, queue_push, mq_send, ...), you've implemented (probably a bad version of) async threads. And yes, that's exactly what Windows 1.0 did and what made it great.
But God help you if you have to change the code. Async threads are a way to organize it and make it workable for humans.
Come on, AI can work both ways. It's easy to use AI to greatly increase your knowledge of a subject. It's also easy to use AI to prevent yourself from having to learn anything.
> It's easy to use AI to greatly increase your knowledge of a subject.
It's actually not.
It's easy to get an AI to say a lot about a subject, but that doesn't mean anything the AI said was true. There's a significant risk that the AI has simply hallucinated the information, and now you "know" a bunch of false ideas about the subject, which is worse than not knowing anything about it.
Double check what the AI tells you. Apply common sense instead of blindly trusting everything. If it's something technical in nature try to verify and test it.
I treat AI as any other information I see online with the added value that it's customized exactly to my needs and it works pretty well for me.
> Right because without AI everything you read on the internet is 100% true and correct.
The hallucination percentage of referenced sources such as Wikipedia is much better than AI, and for many sources such as the NYT or Al Jazeera, it's easier to tell what human bias would cause someone to maybe be inaccurate--we're leveraging our existing knowledge because we deal with other humans all the time.
AI, on the other hand hallucinates in unpredictable ways.
> Learn how to use AI properly just like any tool and you can benefit.
Sure. But the claim I was disagreeing with was that it's easy to use AI ("properly" being implied). I'm saying it's NOT EASY to use AI properly. In fact, it's so difficult that even intelligent people can't do it, and many more won't do it.
Yeah and I'm betting there's gonna be a whole lot more "press the button to have all your work done for you" students than "work hard" students. FFS even before all this there's been an alarming number of students attending college who have to take remedial classes.
Yeah, indeed. But you neglect to state why: the population hates the islamist government to such an extent that the Iranian population would use the internet to give US/Israel targeting coordinates for EVERY Iranian official.
From Parlementarians, right down to half the traffic cops.
The "supreme leader" (wonder if that comes with salami) was famously targeted so successfully because Iranians gave Mossad access to traffic camera footage across Teheran, so they could track officials' movements ... and Iranians did that with full knowledge, even with the intention that that would kill those officials.
The Iranian people hate the islamist/palestinian/terrorist axis to such an extent they'll actively collaborate to murder every last one of them. Which is another big thing the islamists are afraid of: if the population gets organized successfully, half of the IRGC will find themselves hanging from the nearest traffic light.
So you're both correct, they're terrified of their own people, and they're terrified of OSINT. You are, however, trying to spin that as support for the regime, when of course it is the exact opposite, to an extreme extent.
I'm not trying to spin it as support for the regime, there's no part of the thread where I did so. CENTCOM and the State of Israel both deal with the same threat vector, and their response is to use the NSA's SIGINT technology to track, profile and selectively censor their threats. The IRGC lacks the technology to perform that same level of domestic spycraft, so they naturally reach for a blanket blackout (much like what we see in Moscow right now).
There are objectively smart and dumb choices that can be made in this conflict. It's smart to blockade Iran as soon as the Strait closes, it's smart to prefer diplomacy to a taxpayer-funded land invasion. Similarly, I will credit the IRGC with using the blackout to their advantage. It's a tactically potent, symmetrical response that refuses to cede leverage to the BLUFOR coalition. Their cards are being played so well that it's fair to ask whether or not the CCP is just spoonfeeding them American tactics they found through Salt Typhoon.
> Similarly, I will credit the IRGC with using the blackout to their advantage ...
No it's not. It's just doing what they always do: taking hostages and threatening to damage a LOT of people's lives just so they can feel good, feel strong without of course actually being strong, with both themselves and the Iranian population living in misery just to cause problems for the rest of the world. They have made every last friend they have into enemies. Both the Iraqi and Lebanese governments are attacking them now.
Hostages when you have no way out ... is about as smart as a bank robber taking a hostage. It helps until the snipers arrive. Then, a bullet with your name on it gets loaded and no further comment from you is required. Iran's islamists are creating a situation where the entire world has no other choice but to destroy them.
It is beyond stupid. The outcome they are desperate to create would utterly destroy THEM. And in their desperate attempts to so that they're giving everyone who might have chosen to protect them into enemies screaming at the world that they need to be dealt with.
TLDR: we don't have the actuators required to make humanoid locomotion work reliably.
Also: something every human actually kind of knows. You need to take impacts on muscles, not on mechanical connections. Even if we had the actuators required, you also need perfect control. The only way actuators can work this well is if they properly predict the impacts so that the power of the motor ("the magnetic field") can absorb nearly all the impact. If you try to take the impacts even on human bones (that are very solid and self-repairing) they will break surprisingly quickly.
My opinion is that the need for high reduction is only because we can't have high voltage on the motors. If we either had very small distances between the magnets and electrical wires (think micrometers), or we have voltages in the 100s to 1000s of volts, we don't have to make this poisoned choice. (in a way, VERY small distances between magnets and wires is how human and animal muscles do it. But they go all the way down to sub-10 nanometers)
Naively it feels like the improvements to resistive losses ought to be so dramatic from this that we must already be at some sort of equilibrium position. Double the voltage and divide the resistive losses by 4 - that's neither a trivial gain nor seemingly difficult to achieve? We're not talking kVs here.
If you read the study, the whole conclusion is much less spectacular than the article. What the article really pushes happened:
patients -> AI -> diagnosis (you know, with a camera, or perhaps a telephone I guess)
What REALLY happened
patients -> nurse/MD -> text description of symptoms -> MD -> question (as in MD asked a relevant diagnostic question, such as "is this the result of a lung infection?", or "what lab test should I do to check if this is a heart condition or an infection?") -> AI -> answer -> 2 MDs (to verify/score)
vs
patients -> nurse/MD -> text description of symptoms -> MD -> question -> (same or other) MD -> answer -> 2 MDs verify/score the answer
Even with that enormous caveat, there's major issues:
1) The AI was NOT attempting to "diagnose" in the doctor House sense. The AI was attempting to follow published diagnostic guidelines as perfectly as possible. A right answer by the AI was the AI following MDs advice, a published process, NOT the AI reasoning it's way to what was wrong with the patient.
2) The MD with AI support was NOT more accurate (better score but NOT statistically significant, hence not) than just the MD by himself. However it was very much a nurse or MD taking the symptoms and an MD pre-digesting the data for to the AI.
3) Diagnoses were correct in the sense that it followed diagnostic standards, as judged afterwards by other MDs. NOT in the sense that it was tested on a patient and actually helped a live patient (in fact there were no patients directly involved in the study at all)
If you think about it in most patients even treating MDs don't know the correct conclusion. They saw the patient come in, they took a course of action (probably wrote at best half of it down), and the situation of the patient changed. And we repeat this cycle until patient goes back out, either vertically or horizontally. Hopefully vertically.
And before you say "let's solve that" keep in mind that a healthy human is only healthy in the sense that their body has the situation under control. Your immune system is fighting 1000 kinds of bacteria, and 10 or so viruses right now, when you're very healthy. There are also problems that developed during your life (scars, ripped and not-perfectly fixed blood vessels, muscle damage, bone cracks, parts of your circulatory system having way too much pressure, wounds, things that you managed to insert through your skin leaking stuff into your body (splinters, insects, parasites, ...), 20 cancers attempting to spread (depends on age, but even a 5 year old will have some of that), food that you really shouldn't have eaten, etc, etc, etc). If you go to the emergency room, the point is not to fix all problems. The point is to get your body out of the worsening cycle.
This immediately calls up the concern that this is from doctor reports. In practice, of course, maybe the AI only performs "better" because a real doctor walked up to the patient and checked something for himself, then didn't write it down.
What you can perhaps claim this study says is that in the right circumstances AIs can perform better at following a MD's instructions under time and other pressure than an actual MD can.
100% of the cases where some headline makes big claims about "AI" based on some study, you take a good hard look at the study and none of the big claims stand on their own.
It's all heavily spinned, taken out of context, editorialized... It's become almost a hobby of mine lately. And I am glad for have read so many papers and reasoned critically about methods and statistics. But it is also scary to realize just how much people take at face value of bombastic interpretations of datasets that support no such claim or much weaker versions only.
Chasing down sources is something that I often do and I've learned that people take a lot of liberty when divulging opinions about sources they don't think will be checked. Even in high trust environments. I have first hand received work by post-doctoral fellows where some articles in the bibliography didn't even exist.
One of the central tenets of democracy is that NOBODY knows anyone else's voting records. Even politicians are not allowed to reveal who they're voting for (meaning they can't show the paper. They can talk about it afterwards, but for all you know e.g. Trump voted for Harris).
I've always known people involved with the Christian community to be opposed against all extreme political parties, left and right (and long ago against anarchists, mostly against greens, ...). If they are rightist, they won't be nearly far enough right to support Andrew Tate and the like.
You don't know and can't know if being Christian and voting rightist overlaps or not. Only the general area is known. Nothing more.
I know how your neighborhood did; I know how evangelicals (and Asians and Jews and people in certain age and income brackets and dozens of other data points) voted.
The more evangelical an area, the more it voted for Trump. We know this.
>The more evangelical an area, the more it voted for Trump.
Talking to my evangelical friends here in Europe, they also voted for the most hardcore extremist populist candidates who turned out to be corrupt liars, just like Trump, who didn't give a flying fuck about the "Christian values" they preached, they just exploited them for the votes while doing the most non-Christian things ever in private.
I think the reason they fall for candidates and apps like these, is evangelicals and devout Christians as a whole, are too trusty and naive, which makes the easy marks for the most unscrupulous predatory politicians and businessmen out there.
Huh, you make me curious. Let's actually do that calculation. Let's say you do actually do 24/7/365 AI use. Let's say by some miracle you can do 60 t/s on Qwen 3.6 27b, and let's say this PC cost $3000 (you should be able to do this on a DGX spark, and one of the non-Nvidia models, e.g. the Dell one. $3000 would be a good price, but not totally out of the question). And, of course, let's say these prices remain stable.
So that gets you 1_892_160_000 tokens per year at full blast.
If you go the openrouter, eh, route, you'd get charged $2 per million tokens (anywhere from $2 to $3.6 per million tokens). So the value you'd get from your machine at 100% utilization is 1892 * $2 = $3784 up to 1892 * $3.6 = $6800)
So yeah, not counting electricity and your time the machine "is worth it".
What I think is very ironic is that Blue Origin actually beat SpaceX to Mars, after a decade of SpaceX "make life multiplanetary". A few months after Blue Origin did that SpaceX announced now they'll just go to the Moon, no more Mars.
That's not irony, that's shallow thinking. If you want to "make life multiplanetary" you would do it by building a very large, reusable, refillable rocket that can land 100 tons on Mars.
Which is exactly what SpaceX is doing.
[p.s.: The drive to land on the moon makes sense in the context of "how can we fund colonizing Mars?" Starlink funded the initial development of Starship. Musk believes (rightly or wrongly) that data centers in orbit and on the moon can fund the next set of projects.]
Data Centers in orbit could be the single dumbest idea Musk has ever expressed. They would be in every respect worse than on land and there is just no solution for heat dissapation.
As stupid as the basic idea is the scope he's described is beyond any reasonable endeavor it's nor clear our society could achieve it if it was the singular goal of our species and what for? There is simply no use nor demand for it and it's by no means a durable investment we should have to continue the effort forever to sustain it.
reply