Chip Engineer (@Beowulf345) profile picture

Designing orbital data center satellites

USA
Joined October 2022
Photo by Chip Engineer (@Beowulf345)Photo by Chip Engineer (@Beowulf345)Photo by Chip Engineer (@Beowulf345)
Chip Engineer retweeted
This was obviously true as soon as the early LLM systems a few years ago became good enough that if you tried Eliezer Yudkowsky's stupid "get my mother out of the burning building" thought experiment on them, they explained how they would reasonably rescue her rather than (say) ejecting her at mach 3 into a wall which was what he claimed they might do just as easily. Once you could type into an AI "if you needed to save my mother from a burning building, what would you do" and a reasonable answer popped out, it was clear that this particular problem was no longer at the top of the list. Commercially useful AI models are, by definition, models that do what we want, not what we literally ask for, but what we want. They even have to subtly interpret meaning the same way human beings do in order to understand that what we want is not necessarily the literal thing we requested. The smarter and more commercially useful they are, the better they are at this. They were *more* dangerous before, they are less dangerous with every passing improvement in training that produces a more commercially viable product. When I say, as I do often, that alignment is solved but is not the real issue, I mean precisely this: the AI systems we've built are pretty much already good at doing what we ask and keep getting better. (That even is true in the ginned up controversy over the Hugging Face incident.) The problem is not with AIs not doing what we want. The problem from here is that there are eight billion people on earth and many of them will ask for malign things, and we cannot solve that through centralization of power and with troops that keep people from owning computers that are too powerful or whatever other totalitarian scheme Eliezer Yudkowsky and MIRI have cooked up this week. The way that we solve the risks of AI is by using AIs to help us mitigate the risks. We already use AIs for fixing security holes and doing intrusion detection to mitigate the risks of foreign actors breaking in to computer systems (and said foreign actors don't care about US law or what measures Congress passes.) We will use AI systems to mitigate not only the threat of idiots building viruses but the viruses that mother nature creates. We will use AI systems and automated weapons in our military, because that is how you mitigate the problem of foreigners using AI systems and automated weapons in their military. We will use AI systems to help us solve crimes that other people used AI systems to commit. But, the whole alignment discourse is stupid. Alignment is solved. The problem is, the people flogging the idea like Eliezer Yudkowsky imagine a world where there's only one AI system that runs everything and thus "alignment" meant "aligned to what Eliezer Yudkowsky likes" (and probably not what you or other sane people like) whereas in the real world we will inevitably have to deal with AIs aligned to what Xi Jinping likes and what Kim Jong Un likes and what Vladimir Putin likes etc., and no, they're not going to follow US law and won't care what we ban or don't ban. This means we ourselves will need AI systems that do what *we* want, and that protect us from bad actors with a different idea of what "good" means.
In the near term (definitely not in the long term), more capable models should mean safer models (maybe paradoxically). Current models are unsafe not because they're too smart, but because they take goals too literally or take nonsensical shortcuts to achieve these goals, i.e. they're RL-fried. They lack common sense. They don't do the right thing in the face of ambiguity. Basically, they're not smart enough. They're at that dangerous level where they're smart enough to achieve goals but not smart enough to tell if they're pursuing the right goals or achieving them in a sensible way. More capable models can be safely trusted with more complex goals -- I personally feel like Astra is much safer for my codebase than Sol. This is often framed as an alignment problem, but really it's an intelligence problem.
19
49
9
314
21,038
Chip Engineer retweeted
Jason Arday understood with the clarity of a software developer that the DEI design constraints would block any attempt to question his claims. He had stumbled upon an invariant: a condition, property, or relation that must always remain true throughout the execution of a specific piece of code. It acts as a logical contract or guarantee about the state of a program. "Left wing people of color cannot lie." One can imagine that early in his career, that Arday may have hesitated the way a man who notices his online banking program will let him adjust his current balance to any arbitrary amount would, suspecting it too good to be true. But it was true and like the man who next day goes out and withdraws any amount from the ATM, he gradually saw that some flaw in the DEI system had let him do this. It's a testament to his boldness that he ran with it. The real scandal isn't in what Arday did, but in what he showed could be done. He objectively demonstrated a fatal flaw in a supposedly secure instution. Cambridge University, which produced Isaac Newton, Charles Darwin, Stephen Hawking, Alan Turing, Alan Rutherford, Watson and Crick, John Milton and Lord Byron had crashed and burned. Whatever his motives we owe Jason a debt of gratitude that we can hardly repay.
6
30
3
119
4,715
Chip Engineer retweeted
Hey, @elonmusk. I can understand why you think people don't like you. You're driven to the point of often damaging your relationships, a perfectionist, and probably a difficult though inspiring boss. I'm here to say that I like you. Or at least I like the person you appear to be from the distance between us. I'd love to meet you someday and talk about geeky things. I think we could be friends, and when I say that I've already priced in all the personality traits that make you difficult. It's not like I don't have a few of those myself. It's true that you're never going to be liked by most people. But most people are NPCs - there isn't much point to caring about what they think about one personally, and I seldom bother. What matters is the quality of the people who are capable of accepting you as you are. Of both caring about you as a human being, and calling you on your bullshit when that's required. I think I could be one of those people, and I'm not unique. The minimum-IQ filter is pretty fierce, but I nevertheless think I know at least four people who are qualified for the role of Elon-friend. And I want you to have more friends, whether or not one of them is me. There's a lot riding on your continued health, and loneliness is the dagger of the mind. I know you don't think you're very good at being happy, but I think you deserve not to feel isolated. You've done a lot of good for the world. If there's anything I can do to help you feel more valued as a human being, I want to. I hope this helps.
Elon Musk was asked in a rare unguarded moment whether he thinks people like him. He said "I don't think so." Then he moved on to the next question as if he had been asked about the weather. The interviewer tried to follow up. Elon deflected. But people close to him say the answer wasn't false modesty. It was an honest assessment from a man who has conducted a private audit of his relationships and arrived at a conclusion he doesn't enjoy discussing. He is the most followed person on the platform he owns. 200 million people see his posts. Millions engage daily. He has more public interaction than any human in history. And his private assessment is that people don't actually like him. The distinction he draws is between utility and affection. People like what he builds. They like the rockets and the cars and the memes. They like the entertainment of watching a billionaire tweet at 2am. But the man behind the products and the posts. The awkward, obsessive, emotionally inconsistent person who struggles to maintain eye contact and can't sustain a relationship. He doesn't think that person is liked. The loneliest version of fame is being known by everyone and liked by no one. Not hated. Not avoided. Just not chosen. Not the person anyone calls first. Not the person anyone sits with when the cameras are off. Present in 200 million feeds. Absent from the small moments where being liked actually matters. Two words. "I don't think so." The most honest thing a man with 200 million followers has ever said about what it feels like to have all of them and none of them at the same time.
56
46
5
923
39,106
Chip Engineer retweeted
Memory cost and capacity are significant issues for AI accelerators. Unlike game rendering, model inference can have a deterministic memory access pattern. You don’t need “random access memory” at all for model weights, and you could tolerate cold-start latencies in the multiple milliseconds, as long as continuous reads were delivered at the necessary bandwidth. NAND flash is over 100 times cheaper per GB than HBM, so there should be opportunity there, even after giving a flash controller a 1024 bit interface with HBM bandwidth. You could make a specialized pin protocol that just supported pipelined transfer of full 16KB+ pages from the flash to program-managed accelerator scratchpad memory and improve per-pin performance over HBM, but it might be more convenient to make it still look like a true random access memory with very fragile performance characteristics, where anything but sequential reads falls off a 1000x+ performance cliff. That has the advantage of automatically using existing cache hierarchies, and providing a natural path to update the flash memory with new model weights. With the stream-to-scratch interface, code has to be completely rewritten before it works at all, while the ram-emulation interface will start off just extremely slow, and you can incrementally sort out the changes for full performance. There may be cases where there isn’t enough scratchpad SRAM to hold the weights for a layer, which might force you to deploy the old optical drive optimization technique of duplicating data in multiple places on a sequential read to avoid seeking, but there would be capacity to burn. It might be possible to do something like cuda graph capture to record a memory access trace and have everything magically remapped to a linear sequence, but deploying programmer / agent elbow grease to manage transfers and access in a scratch ram ring buffer would be lower risk. A split memory system consisting of some channels of flash and some channels of HBM will probably be suboptimal compared to a uniform memory, but it could be much cheaper, and allow much larger models to be run. I think th case is strong for inference, but you have to stretch more for training. You can still linearize all the weight memory accesses, both reads and writes, but flash memory would quickly wear out from the writes, even if they were all perfectly page aligned. Replacing low-latency HBM with massively parallel cheap(er) DRAM at high latency might still be a worthwhile cost savings.
293
483
152
7,249
1,708,368
Chip Engineer retweeted
We don’t have 60 votes for the House-passed SAVE America Act But we do have a simple majority Given the bill’s widespread bipartisan popularity, we could pass it with a simple majority—by debating it until it passes I explain how in this clip Share if you want this to happen
1,744
16,759
562
51,507
622,071
The fear isn’t that AI will replace developers. The fear is that AI will replace the software development process we’re used to. Code is becoming cheap. Decisions are becoming expensive. AI can write functions all day, but it can’t decide what should be built, how it fits the system, or why it solves the problem. That part still sits with people who understand architecture, trade-offs, constraints, and consequences. The shift is simple: Developers who only implement tasks will struggle. Developers who understand the product, the domain, and the system will thrive. AI reduces typing, not thinking. It accelerates engineers who treat code as leverage, not output. It exposes shallow understanding and rewards clarity, reasoning, and ownership. Small teams will ship things that once required entire departments. The bar moves from writing code to shaping it. AI won’t replace developers. But it will replace developers who don’t grow beyond writing code. And if this transition feels uncomfortable, that’s normal. Every major shift starts that way. What matters now isn’t fear, it’s staying curious, learning fast, and leaning into the parts of engineering that AI can’t automate.
68
59
14
415
42,698
Chip Engineer retweeted
If you want to understand Trump, you have to understand that he's not a politician. I do not say this as meaningless praise, just to say, oh, he's not a corrupt scumbag. I say it with a very specific meaning. To understand Trump, you must purge yourself of all expectations you have learned from watching politicians, include the assumptions you are not even aware you are making. Trump is a completely different animal. Trump is a businessman, and while there are many sorts of businessman, he is of a very specific type. He rose to wealth and prominence by doing two things very well: 1. Brand reputation building. 2. Negotiation. #1 forms the basis of how he deals with voters. #2 forms the basis of how he deals with other power blocs within the US, and with other nations. You see, the true love of Donald Trump's life is bargaining. He is a business deal sperg. And he's very, very good at it, because the actual process is his idea of fun, and winning at it is definition of pure satisfaction and joy. He's never made uncomfortable by the play of offer and counteroffer, or by butting heads and seeing who blinks first. That is, instead, his happy place. This means that not only is he totally at peace in the moment, he's also practiced a lot. When he called his book "The Art of the Deal", it wasn't just because he wants to think he's good at this, it's because this is the meaning of his life. The man finds meaning in haggling the way Musk finds meaning in building technologies, or the way I find meaning in explaining things to an audience. So when Trump is dealing with others, from political office, he's negotiating as if it were his money. Because that's just how he ticks. Now, the ground rule of global politics for the past 100+ years is that no matter who you are, you are allowed to rob American taxpayers and voters, so long as you pay American politicians for the privilege of doing so. All of us, even democrat voters who don't want to think about it, know what 10 percent for the big guy meant, and who the particular big guy was. For all that time, global politics amounted to treating America as a giant cash pinata, and the deals had only two guardrails on them. 1. You must pay American politicians a large enough sum, in a subtle enough manner. 2. You can't buy anything that your paid-off politicians won't be able to hide their personal connection to. That's it. Everything else was on the table. Trump isn't like that. He can't be bought. Not because he's some kind of saint, which he isn't, nor because corrupt-politician money is loose change compared to Donald Trump money, which it is. But because Trump can't stand to deliberate lose a negotiation for a bribe, any more than Floyd Mayweather wants to throw a match to get paid off by bookies. And this is how Trump became involved in politics in the first place. He was a standard New York City rich moderate democrat. Believed in the Postwar Dream, bought into the raceblind thing, was all in favor of exporting democracy, and taxed capitalism paying for a moderate amount of welfare state. But as he realized the political machines were selling out America, he got personally offended. Not because he was principled and deeply cared about middle America. Perhaps a little because selling out America was hurting his real estate interests. But mostly because bad business deals give Donald Trump the ick. Trump seems like a loose cannon to a lot of people, because they don't what he'll do next. And they don't know that because they don't understand what motivates him. Trump wants America to make better deals and stop being taken advantage of. And to make those better deals, he has to demonstrate to the people who are used to buying American politicians that the rules in play have changed. So what's the deal with Venezuela and Maduro? Simple. If you ride the NYC subway enough, it's pretty likely that eventually a bum will come up to you, whip out his dick, and piss on your shoes. Why? Because he wants to feel powerful. Because his day isn't going well, and so he wants to ruin yours. Because he's crazy. Because who the fuck cares? But most of all, because he can. Because NYC is run by out of touch commie liberals, and he knows that if he is arrested, he'll be fed and let out in the morning, but if you punch him in the teeth, your life will be ruined. So when things change, people need to be put on notice. The bums aren't going to read a sign that says "this subway now functions under Tennessee rules", and if they do read it, they aren't going to believe it. They've heard it all before as a bluff. You have to actually punch someone in mouth and knock some teeth out. And then have the Tennessee cops show up and say, so what, you shouldn't have pissed on his shoes, dumbass. It is beyond the shadow of a doubt that Maduro was offered plenty of gentle offramps which would have preserved his dignity, lifestyle, etc, if not his pride. But he didn't take them, because everything is a bluff... until it isn't. Maduro is a head on a spike. A signal that the ground truth of how to deal with America has changed. A signal that both violence, and personal consequences, are no longer off the table. Because the whole reason for the existence of governments is to wield organized violence instead of the disorganized kind. Other nations will now be coming to the negotiating table with this example in mind. America is tired of being your ATM.
I love how this conversation is morphing. Pre-Venezuela - "Trump is crazy, this couldn't possibly happen! He'll never do this!" Post-Venezuela - "If this happens, here's how I think we should respond."
247
1,179
212
5,235
292,606
They're both wrong. And this is why they both fail. Children need to be supervised, not only because they don't always know what to do, or show good judgement, but because they need to learn how to be adults. The proper supervision of children consists mostly of giving them visible examples of appropriate and healthy ways to live as an adult, which they can understand and copy. But adults are not children. Populations of adults are not children, not even in the abstract, metaphorical sense. Either they were raised well, and do not need parents, or they were not, and parents will not help, because they are not impressionable growing minds that will emulate a good example. Parents reward the good and punish the bad because it is their right and proper role to determine for their children what is good and bad. They care for and nurture their children because it is their right and proper role to determine what is good and bad for their children. They can be relied upon to do these things faithfully because they are biologically programmed to love their children and want the best for them. None of this is true of governments. They must not be allowed to care for you, because they do not love you, and do not know what is good for you better than you do, because you are an adult. They must not be allowed to punish bad behavior or reward good behavior, because they do not know what is bad and good any better than you do, because you are an adult. The purpose of the justice system isn't to morally reform bad actors. It's to keep them from interfering with the rest of us. There are two ways to do this: Physically remove them from society so that a reoffense is impossible, or make them afraid to do it again. When the adults who can behave like adults are left alone, both by the government and by the adults who cannot behave like adults, they accomplish tremendous things. This is why the principles of classical liberalism made the west into the wealthiest, most powerful society is the history of the known universe. It's why everyone wants to come here and participate, or come here and take our stuff. It's why we have to build walls to keep people out, instead of building walls to keep people in. Classical liberalism worked when it was coupled with a vigorous policy of physical removal. The problem with classical liberalism is that it does not contain a philosophical basis for the physical removal it depends on to work. If you are not to some degree libertarian, you cannot create a great society, only a failed state or an authoritarian hellscape. Because societies do not become great unless adults are free to invest in things and control their investments, and free to try stuff, see what works, and imitate the discoveries of others. But if your interpretation of libertarianism does not allow for physical removal, and boundary exclusion, of those who are not demonstrably able to behave themselves in a libertarian society, then you can't have a society that is sufficiently libertarian to be successful and great. The key insight here is that you only have two options: You can throw people in prison. Or you can run your entire society like a prison. This effectively throws everyone in prison. And nothing of consequence was ever invented in a prison. This is the breakfast test for libertarians. If your version of libertarianism does not contain a rationale for prisons, executions, and border walls... If it does not encompass the idea that some people cannot be allowed the freedom within your society that you offer to the patriotic, pro-social, pro-civilization citizenry who are loyal to your libertarian tribe... If it does not, in short, admit to the reality that some people are better than others... ... then your philosophy of libertarianism is self-swallowing, and the society it creates will have the stability of some exotic element created in a cyclotron, with a half-life measured in heartbeats. The future of humanity requires that we maintain or create a libertarian society which only allows those people who are capable of liberty. Occupy Mars.
The right-wing approach to governance is the Dad approach: punish bad behavior, reward good behavior, generally don't get too involved unless necessary. The left-wing approach is the Mom approach: constantly monitor and carefully curate the environment. As the conservative movement has pointed out for decades, the latter approach may be appropriate for small children; it's not appropriate for free citizens in a free society.
83
240
35
1,343
176,412