Skillful Dreams

@Ysqander @dwarkesh_sp @dylan522p RL rollouts are model-generated outputs used for training. Generating them means doing inference. Bigger models usually require more compute per generated token, so rollouts are slower and the RL loop takes longer.

@dwarkesh_sp @dylan522p Very unclear post. Why is the RL roll out slower with bigger models?

You'd think the race to AGI would mean training the biggest possible model. But parameter scaling had stalled for a long time after GPT-4's trillion+ parameters, and only now are models getting bigger again. What gives? Partially it’s RL scaling, as @dylan522p explains. A 5T parameter model takes 5x longer to generate RL rollouts than a 1T model. Even if the bigger model is 2x more sample-efficient, the smaller model finishes RL faster, gets deployed to research sooner, and starts helping build the next model before the big one is even done training.

Premise 1 claims determinism entails "whatever can be done actually is done." But this obscures a distinction between two types of possibility: alternative physical futures branching from the same past, and unactualized deliberative possibilities represented within a mind. Determinism denies the former but not the latter. A deterministic brain must represent possibilities in order to act. It models multiple actions that practically "can be done," weighs them against goals and reasons, and picks one to actually do. The unchosen options are unperformed but causally real as representations inside the process that produces the outcome. They are things that can be done that aren't actually done. P1 defines these possibilities out of existence and calls it determinism. But a rock falling and a person choosing are both determined, and only one represents options, weighs them against goals, and responds to reasons. That difference is what grounds agency, responsibility, and moral judgment. Erasing it by definition amounts to assuming the incompatibilist conclusion, rather than arguing for it.

@PAHoyeck The unexamined peach is not worth eating.

Another way to put this: Deliberating among possibilities, most of which will never be actualized, is itself an essential part of what causes the action I do take. The fact that only one option becomes actual does not mean the others were irrelevant or unreal in the sense that matters for agency. They existed as deliberated alternatives within my decision-making process, and weighing them against each other is precisely how my eventual action was produced. And this is the relevant sense in which something ‘can’ be done that isn’t actually done: as a deliberative possibility in the mind of a decision-making agent.

Your argument fails because premise 1 is not the definition of determinism. It smuggles in a controversial sense of “can.” Determinism says: given the exact past and the laws of nature, only one future is physically possible. It does not say: the only thing an agent can do is whatever the agent in fact does. That stronger claim already assumes incompatibilism. This matters because the “can” in “ought implies can” is about agential ability: whether the agent had the relevant capacities to recognize reasons, deliberate, exercise self-control, and act accordingly. It is not about the power to alter the past or suspend the laws of nature. So the issue is not whether determinism makes alternate futures impossible in the absolute, physics-level sense. Of course it does. The issue is whether it removes the kind of ability relevant to moral assessment. Compatibilists say no. If I choose the orange over the apple, it can still be true that I could have chosen the apple in the relevant sense: it was a live option within my practical capacities, and had my deliberation or reasons gone differently, I would have taken it. Determinism says that my actual reasons-process had a determined outcome. It does not show that the unchosen option was never really available as an object of choice. In fact, meaningful agency seems to require exactly this kind of determination. My action is free not because it is uncaused, but because it is caused in the right way: by my own beliefs, desires, values, and deliberation. If those had no causal role in what I did, my action would be less mine, not more. So your conclusion does not follow from determinism alone. To get it, you need a further premise: that if only one outcome was ever going to happen, then the other options were not real possibilities in the sense relevant to deliberation and choice. But that is exactly the point under dispute. So the argument does not derive incompatibilism from determinism; it simply restates incompatibilism as premise 1.

I don't think determinism is compatible with moral judgment

This has been known since The Information's article from Nov 22, 2023: "OpenAI Made an AI Breakthrough Before Altman Firing, Stoking Excitement and Concern" "For years, Sutskever had been working on ways to allow language models like GPT-4 to solve tasks that involved reasoning, like math or science problems. In 2021, he launched a project called GPT-Zero, a nod to DeepMind’s AlphaZero program that could play chess, Go and Shogi. The team hypothesized that giving language models more time and computing power to generate responses to questions could allow them to develop new academic breakthroughs. Lukasz Kaiser, one of the co-authors of the groundbreaking Transformers research paper, which describes an invention that paved the way for more sophisticated AI models, held a key role on the GPT-Zero project, this person said. Among the techniques the team experimented with was a machine-learning concept known as “test-time computation,” which is meant to boost language models’ problem-solving abilities, this person said. Earlier this year, Sutskever and his team discovered a variation of that method that prompted far greater results in their efforts to train more sophisticated models. Sutskever’s breakthrough allowed OpenAI to overcome limitations on obtaining enough high-quality data to train new models, according to the person with knowledge, a major obstacle for developing next-generation models. The research involved using computer-generated, rather than real-world, data like text or images pulled from the internet to train new models. In the following months, senior OpenAI researchers used the innovation to build systems that could solve basic math problems, a difficult task for existing AI models. Jakub Pachocki and Szymon Sidor, two top researchers, used Sutskever’s work to build a model called Q* (pronounced “Q-Star”) that was able to solve math problems that it hadn’t seen before, an important technical milestone. A demo of the model circulated within OpenAI in recent weeks, and the pace of development alarmed some researchers focused on AI safety. The technical breakthrough, spearheaded by OpenAI chief scientist Ilya Sutskever, raised concerns among some staff that the company didn’t have proper safeguards in place to commercialize such advanced AI models, this person said. In the months following the breakthrough, Sutskever, who also sat on OpenAI’s board until Tuesday, appears to have had reservations about the technology. In July, he formed a team dedicated to limiting threats from AI systems vastly smarter than humans. On its web page, the team says, 'While superintelligence seems far off now, we believe it could arrive this decade.'" theinformation.com/articles/opena…

It’s pretty crazy that “what did Ilya see?” might have a serious answer: He saw Strawberry - the training architecture that would enable the rapid development of AI coding agents, starting with o1

@Aella_Girl It would be funny if your anti-traditional persona turned out to be a contingent workaround for a few removable aversive bottlenecks, rather than a deep truth about your nature

i seem to have organically started showering a more frequently after cutting off 50% of my hair (was to my hips). It turns out not having a giant sopping wet mass that takes hours to dry makes showering less aversive

@RokoMijic It's not Israel dragging the US. Gulf states have pledged $3 trillion in US AI investment, essential for US dominance. Iran, under the Islamic Republic, holds all of that hostage permanently. Strategically, this is about securing the economic foundation of American AI supremacy.

I support fast, effective regime change for Iran, but it looks like the Trump regime actually just doesn't have a plan and is mostly being dragged around by Israel with no real end in sight. Am I wrong?

@acestopmeowing @_el_nino19 @atrupar Because you don't make decisions based on the worst case alone. Otherwise you'd never do anything. The worst case of surgery is you die on the table. Does that mean you should never have surgery?

@_el_nino19 @atrupar THEN WHY DO IT IN THE FIRST PLACE

Trump: "I guess the worst case is we do this and then somebody takes over who is as bad as the previous person. That could happen."

@TheZvi So far, I have found it much more open to engaging with controversial topics compared to GPT-5. A more interesting, useful conversationalist and thinking partner by far. Similar to Claude 4.5 Sonnet.

GPT 5.1 Reaction thread. What have we got?

Schopenhauer makes the same move (noting that subject and object are interdependent) but in the reverse direction: “‘The world is my representation’: – this holds true for every living, cognitive being, although only a human being can bring it to abstract, reflective consciousness: and if he actually does so he has become philosophically sound. It immediately becomes clear and certain to him that he is not acquainted with either the sun or the earth, but rather only with an eye that sees a sun, with a hand that feels an earth, and that the surrounding world exists only as representation, that is, exclusively in relation to something else, the representing being that he himself is. – If any a priori truth can be asserted, then this is it; for this truth expresses the form of all possible and conceivable experience. … Thus, no truth is more certain, no truth is more independent of all others and no truth is less in need of proof than this one: that everything there is for cognition (i.e. the whole world) is only an object in relation to a subject, an intuition of a beholder, is, in a word, representation.” (The World as Will and Representation)

Wittgenstein thinks some philosophical ‘positions’ are actually unintelligible, not really possible views to hold or think at all. He would probably think Cartesian dualism is such a position. I’m not very confident about that. But I do think solipsism would be such an example of a purported ‘position’ which philosophers sometimes pose as something which can be entertained, and then rejected. But the form of rejection will be itself confused, if it treats ‘solipsism’ as some position that can be said to be true or false. It would take a whole essay to really justify the claim that only-the-subject-that-is-oneself-exists (solipsism) does not designate a real position. But the basic reason is that the whole idea of the ‘subject’ only makes sense when contrasted to the existence of things outside the subject. I strongly suspect this is true, but I still need to work out my thinking here further

The full article does detail what he said. He said: “I do not believe that humans have an attribute called gender. I do not believe that G-d puts people in the wrong bodies. I do not believe that it is okay to sterilize children. You have the right to your beliefs. I do not share them.” And he accused people of antisemitism who were organizing to stop Google’s technology from being used by Israel.

@LokiJulianus you know that tweet thread didn't even say *what* he said about Gaza and gender or which "side" he's on but then it didn't have to did it "all you have to say is 'divisive' and everyone will understand everything"

This is the reality of being a right-wing individual at any large scale enterprise: left-wing culture warriors get to ragecel all day about their pet causes, but you can't even speak your mind in an unobjectionable fashion. You better believe they've been emboldened by Mamdani winning too.

@mgubrud @demishassabis Another way to see it: Spinoza is doing two moves at once. He’s expanding what we think nature is (from just contingent stuff to infinite, necessary substance) and stripping down what we think God is (removing the person). They meet in the middle at ultimate reality.

@mgubrud @demishassabis “God” isn’t just about something that loves you. Spinoza thinks nature has the essential divine qualities (infinite, necessary, self-caused) that justify calling it God and relating to it with reverence. He rejects the personal qualities, like loving you.

Indeed, Spinoza had it right.

@mgubrud @demishassabis The word "God" does two things here. Definitionally: nature is infinite, necessary, self-caused. Attitudinally: grasping an infinite rational order naturally evokes awe. The reverent feeling is an appropriate response to nature's structure. Both aspects justify calling it God.

@skillfuldreams @demishassabis I think that is his rhetoric, but actually he is trying to put a warm human face on nature. Because of the way this makes him feel. Not because of logic. He is using words to try to justify this.

@mgubrud @demishassabis Spinoza keeps "God" because that's the philosophical term for whatever is infinite, necessary, self-caused, and grounds all existence. He's saying that when philosophers talk about God, something actually fits that description: nature itself, just without the person or plans.

@demishassabis If you like your glasses rose-colored, I suppose. My question to Spinoza (Spinoza advocates invited to answer): If God is just nature, why are you calling nature "God," instead of just "nature"? Aren't you really just trying to put a warm human face on nature, while denying it?

@14JUN1995 I think there might be a small misreading here. Hume says we observe finite goodness/wisdom in our minds, then augment those qualities without limit to get the idea of God’s infinite nature. The ‘unbounded’ part is the augmentation process, not the qualities we start with.

Hume: all ideas are derived from sense impressions Me: Yessir Hume: Like the idea of God— Me: huh? Hume: The idea of God is derived from the impression of self-reflecting on the unbounded liberty, goodness, and wisdom of the mind itself! Me: You what bruv

Reporter: How much do you expect Xi will push you on Taiwan? Trump: I don’t know that we’ll even speak about Taiwan. I’m not sure; he may want to ask about it. There’s not that much to ask about. Taiwan is Taiwan. But the beautiful part about Taiwan, frankly, and they understand it, is we’re bringing a lot of those chipmakers into the United States. You look at the number one chipmaker in the world — they’re building a massive factory in Arizona. The biggest in the world, actually. We’re making chips. If you look at Nvidia, Nvidia’s doing unbelievab— we’re bringing a lot of the Taiwan business back to the United States. I think within two years we’ll have 40–50% of the chip market. That’s happening — I don’t know — it’s happening because of the tariffs; it’s also happening because they’re probably, they’re maybe a little bit concerned about things, who knows. But they are… you know, I don’t know if you’re reporting it, but massive numbers of — and percentages of — the chip business is coming back right now, as we speak. They’re building plants in Arizona, Texas, and other places. Reporter: On chips, sir, have you agreed or will you agree with Xi potentially to allow the license of more exports of Nvidia Blackwells, or downgraded— Trump: So we’ll be speaking about Blackwell. Blackwell is the super-duper chip. Blackwell’s amazing. Nvidia… Jensen is an amazing, brilliant man, and it’s probably ten years ahead of any other chip. And unlike a lot of other businesses, you just can’t catch them. You know, it’s… mostly you can copy and you can do things, but you really can’t in the chip business. For whatever reason, it doesn’t work that way. Blackwell just came out. Jensen presented me with the first Blackwell chip three days ago in the Oval Office. And it’s an amazing thing that they’ve done. But we’re… you know… that’s… our country — we’re about ten years ahead of anybody else in chips, in the highly sophisticated chips. I think we may be talking about that with President Xi. Reporter: But China has actually discouraged purchase of those chips. Is that something that’s on the table? Trump: Well, China actually terminated the relationship with Nvidia. That was a big thing. But in the meantime Nvidia hit a brand-new high. As did… you know, the stock market yesterday hit the highest ever that it ever was, so, all of the markets — every one of our markets — we’re doing great. The country is doing great.

@MajorOcelot45 @amitisinvesting Here's the full audio. He starts talking about China, Taiwan, chips, Jensen, and NVIDIA at 13:21. youtu.be/FIoIm0ikLMg?si…

$NVDA SURPASSES $5T IN MARKET CAPITALIZATION OVERNIGHT. TRUMP: - JENSEN SHOWED ME THE BLACKWELL CHIP IN THE OVAL OFFICE - I’LL SPEAK WITH XI ABOUT BLACKWELL Jensen today told everyone that Nvidia has a $500B opportunity over the next *five* quarters (street was expecting $380B) and that was…without China. absolutely incredible to witness 🤯

That passage is a genealogy of moral power, not an ethnic rant. Nietzsche is describing how a priestly strategy of ressentiment generated ideals that overturned aristocratic Rome. The question is whether he’s anti-Jewish, or simply opposed to the priestly moral inversion he attributes to Judaism. Consider BGE §251. He mocks Germany’s “anti-Jewish stupidity,” then says plainly: “the Jews are without a doubt the strongest, purest, most tenacious race living in Europe,” thriving "in even the worst conditions... due to some virtues that people today would like to see labeled as vices." He says it's a fact that "the Jews, if they wanted, could already be dominant, or indeed could quite literally have control over present-day Europe." But, in fact, they "are not working and making plans to this end”; what they want instead is “to be absorbed and assimilated... they thirst for some place where they can be settled, permitted, respected at last.” In light of this, Nietzsche suggests "it might be practical and appropriate to throw the anti-Semitic hooligans out of the country.” At minimum, he has far less tolerance for anti-Semites than for Jews. He despises the moral inversion of ressentiment that he traces to Judaism and sees Christianity as universalizing, but he consistently respects the Jewish people for their vitality, discipline, and genius. In Antichrist §24 he makes that psychology explicit: “Looked at psychologically, Jews are the people with the toughest life force; when transplanted into impossible conditions they took sides with all the instincts of decadence… not because they were dominated by these instincts, but because they sensed those instincts had a power that could be used to prevail against ‘the world’. The Jews are the opposite of decadents: they had to act like decadents…” And in BGE §250 he says Europe owes the Jews “many things both good and bad, but mainly one thing that is both best and worst: the grand style in morality, the horror and majesty of infinite demands, infinite meanings… This is why, among the spectators and philosophers, artists like us regard the Jews with — gratitude.” Put that back beside GM I to see his ranking as unambiguous and contrary to your reading: Jews are his first-rate benchmark of moral inventiveness (“unique genius for popular morals”), while Germans/Chinese are the fifth-rate comparators. Jews and Romans are his opposed first-rate exemplars of, respectively, priestly vs aristocratic value-systems. But that correction isn’t the most important matter. The central issue is that Nietzsche honors the Jewish people (strength, endurance, ingeniousness) while indicting the moral revolution (the priestly, guilt-laden revaluation opposed to life that Christianity globalized). If you want to attack him for rejecting that outcome, fine, but “Jews = fifth-rate” or “Nietzsche = anti-Semite” is the opposite of what the texts say.

There's a few people pushing back against this in the comments, inevitably, but I honestly don't get it. Every time I say this, I get a bunch of people telling me that I'm badly misreading Nietzsche. Then I read Genealogy Essay 1 again and, like... fucking yikes.