It's certainly a massive exaggeration to say that a human was close to solving GRH.
This. The GRH has been around for a while, at least for 110 years or so. Most of the time, there was probably at least one mathematician working on a serious proof attempt. If we put the GRH in a reference class with Fermat's last theorem and the like, that would yield a few percent that the human proof will be correct?
Also, this is completely orthogonal to the capability claims about the LLMs. Nobody can in good faith argue: "LLMs solved Riemann? Boring and trivial. The reason why no human has gotten around to solving it yet is that no student saw it fit to waste an afternoon on it."
The only thing which would argue against LLM capabilities here would be if the LLM proof was just a ripoff of the human proof. This should be rather easy to detect, and the fact that no such claim has been made indicates that the proof strategies are sufficiently dissimilar.
I think with the four color theorem, the general reaction was "oh, turns out it was not a very interesting problem after all".
Another day, another botched execution. I mean, not as botched as the last one, they managed to kill their prisoner eventually. Due to a lack of backing of the medical system or whatever, it seems that the red states -- unlike physicians, EMTs, veterinarians, and junkies -- no longer have the capability to reliably hit a vein. Granted, a condemned might sabotage easy vein access, but the standard way a physician would work around that is intraosseous injection, most death row inmates will likely shy away from sabotaging access to their tibia.
I think medicalized executions in the US are on their way out. I know that patriotic Americans are strongly opposed to importing French inventions, but surely a guillotine could be built to imperial unit standards as a compromise? From a humanitarian perspective, a guillotine seems infinitely preferable to the effectively subcutaneous injection of pentobarbytal the US favors. A society which does not have the stomach for decapitations should perhaps simply abstain from executions.
Or if the slow and painful death is the sentence, then the jury should at least specify that. "To be killed over the course of ideally N minutes".
Sure, but that argument can easily be applied to SBF as well, not that I am claiming what they did is equivalent.
You can add God in if you want, like a cherry on top, and many people get to the Big Bang and they do.
The natural shape of the god this will gain you is a Deist's deus otiosus -- a creator god whose job ended long before the first neutron decayed. That this God would then have picked up his perfect design ten billion years later and messed with biochemistry to make sure that Their Favorite Chiralities would dominate or that homosexuality in Hominidae would be neither to rare nor too common for Their Taste or who forms pacts with particular populations or inspires the letters of Paul or contemporary US politics seems increasingly unlikely from where I stand.
Modern atheism really isn't even possible before Darwin.
I would argue that the fuse of the scientific world view was already lit with Francis Bacon 240 years before Darwin.
I mean, Darwin is one of the great injuries (the others are cosmology and the LLM, with overall technological progress, psychology etc playing lesser roles).
But it is not like we would not have converged on atheism in a world where species were immutable. Not when we upon the night sky and realized that our sun is just one of ten to the power of 24 or so stars.
a sprinkling of TDS
Is it TDS or just believing in linear extrapolations? The past year has been the most unhinged in six years of his administrations so far, with 2025 achieving the second place.
I am not psychologist, but from my perspective it does not seem that the blue wave has lead Trump down a path of reconciliation. And his cult of personality makes it unlikely that JD Vance would invoke the 25th.
Sure, he is not the only demented person who has recently sat in the oval office, but Biden had two redeeming factors: one, he had a lifetime of sticking to the mos maiorum, he was very much DC establishment. Announcing an intent to genocide a people on twitter would have been so out of line with the established US diplomatic strategy that it would have been completely left field coming from Biden. Two, Biden was likely not calling the shots.
My expectation is not that he will start WW3, but my error bars on that are a bit higher than usual. More likely his badness will be on the level of "cutting a deal with China where the US allows them to take Taiwan".
I think a similar argument can be made against buying iodide during the Cuba crisis: "There is only one country in the world which has ever used nuclear weapons in anger: the US. They will have no possible incentive to repeat that because we are not living in Imperial Japan. Hence we are safe."
Anyone paranoid enough to care about prepping would start from a capabilities analysis, not from established behavior. The list of threat actors which might plausibly have the capabilities to cause havoc by attacking IT systems certainly has more than two items. Start with the obvious ones: the USG, other western governments, China, Russia, NK. The big three LLM firms. Musk, plausibly. Possibly anyone with enough GPUs to run the latest Chinese models.
If you did not anticipate OpenAI on your list of threat actors a year ago, you would also hedge for unknown unknowns.
Russia has a big Ukraine shaped wound that the US can stamp over whenever it wants.
Any direct kinetic conflict between the US and Russia would carry a substantial chance of turning into an ICBM exchange. A rational US president (ha!) would be reluctant to escalate to such a level, perhaps over cyber attacks which starve a few million Americans, but likely not over Russia defrosting the freezers in all US supermarkets.
And "we will give more aid to Ukraine" is might be of limited deterrence value. "We will give Ukraine what they need to take Moscow" would do the trick if it was easily done. It is not easily done.
Being in good physical shape is usually not considered prepping, but it will probably help a lot if you have to live without modern comforts for a while.
Counterpoint: the most effective way for vertebrates to carry energy reserves is as body fat rather than cans of food.
Claude claims that mammals can get 4-5MJ/kg from metabolizing skeletal muscle but 32-34 MJ/kg. That is a factor of seven or eight. A kilogram of body fat will sustain you for four or five days energy-wise, a kilogram of muscle mass will last half a day.
Obviously it depends on what kind of apocalypse you anticipate. If you need to outrun a pack of Utahraptors, being overweight will likely not work out for you.
TrumpwantsJensen to sell more to China, with some restrictions and a kickback.
I have to say it is kind of funny to see a lot of the China hawks ('We don't have time for alignment due to China' etc) jump to his defense. Likely Musk et al were never serious about considering Chinese AI a dire threat to the US, it was simply the best argument-as-a-soldier to try to get rich building their own paperclip maximizer.
Questions mostly answered by, "judgment that built the most valuable company in the world".
But also "the judgement that shaved 28% of the market cap of the most valuable company of the world with his dick". The difference being that it is debatable how much he contributed to the former (my gut feeling is that one 10x is on him (CUDA), another 10x is on dumb luck, like winning the lottery), but he caused the latter all on his own.
Judgement is also not a single skill value, one would trust one's car mechanic with different decisions than one's priest. It is not clear that 10x-ing your company is the same skill set as running a very large company whose tech is a key strategic asset of the US.
Huang is more like a Gulf royal who happened to be in charge in 1962.
This. I mean, he was certainly a very savvy Gulf monarch.
I do not think he founded Nvidia with a clear plan to dominate the AI market 30 years later.
His decision to invest in CUDA was excellent in hindsight, but I think it would have been defensible even with just the simulation people as a market. I see no indication that he anticipated that the market for GPUs would explode due to Etherium (which was memory-bound) or AI.
In 2017, Nvidia made 1.9G$ from HPC (half of that likely for traditional simulations, not AI) and (unofficially) a billion or two from miners. Plus a few billions from actual gamers. Miners were lucrative for Nvidia, but not transformative. Then the LLMs came along and that was transformative for the company, of course.
I am with @netstack on this one, it seems to me that the corruption refers to his kleptocratic tendencies, not to his alleged morally corrupt sex escapades at Epstein's island.
There was always a swamp in DC, but most of the swamp-dwellers had the decency to stay out of the sight of the public. Think a revolving door between regulators and industry, for example.
Corruption exists on a spectrum, and not all forms of it are equally corrosive to society. As an analogy, consider violent crime. Society can deal with some murders, it can deal with some suspects getting acquitted even, and with the knowledge that there is a dark field where we do not even suspect that a murder took place. But if we allow people to kill with impunity, 'that neighbor is doing a successful youtube channel about his weekly serial killings, that one stabbed his wife and then called a cleaning company to clean up the mess, that one is the three-time national school shooting champion', that is not compatible with having a working society.
With Trump, the corruption is all out in the open. It is common knowledge. The only way to make it more blatant would be for the DoJ to publish a price list for presidential pardons. This has broader effects on his administration. If everyone around you is selling secrets to prediction markets, you would be a fool not to partake, right? Just a perk of the job. This culture quickly drives out anyone who is not a grifter and attracts the most ambitious would-be grifters of the country in turn.
Snarky answer: while we are it, we should also get rid of weight classes in judo, weightlifting and the like. And restrictions on biomodding and doping. Having a steroid-free league is as silly as a vegan league. We might as well ban ebikes at the tour de france! If you are unwilling to sacrifice a few dozen QALYs to become objectively better at your sport through steroids or HGH, make room for someone who will. If other drivers are to chickenshit to drive an F1 car powered by a jet engine, they can always compete at go-carts.
I will say though that when I was in school, I noticed the absurdity of taking a subject where genetics largely determined aptitude and then trying to impose a grading scheme to counteract this. Math grades are not fair, some students have way more aptitude than others, but (at least in Europe) we openly accept this unfairness and just grade everyone on the same curve. PE is just as unfair as math in the uneven distribution of aptitude, only that there we fix the leading order unfairness (sex dependence). Why stop exactly there and not also grade differently for testosterone levels, or height, or weight? Not that I ever cared about sports grades very much. The only time I ever worked on a physical rating of me was for unfit for military service, my proudest achievement.
For any competitions people voluntarily engage in, my take is that the solution is to not legislate it either way. Simply do not use public funds for it, and then let people form associations as they please and have the market decide, archipelago-style. If people want to found the Gentile Chess League because they get tired of getting crushed by Ashkenazim, let them. If the male NBA players make 100x what the WNBA players make (or whatever the factor is), that is just capitalism in action, and it should not be more upsetting than female pornstars outearning male pornstars. If some association accepts everyone who renounces being a man, and another carefully checks the D2:D4 ratio and T levels of all players and a third bans everyone who is not XX genotype and a fourth simply has a panel which will judge if a candidate is cute enough to be a player, and a fifth exclusively allows trans players, let them all do their thing and convince the public that their competitions are interesting.
For the most part, US legislatures wisely leave it open to various sports associations to make rules if you can touch the ball with your hand or your foot or a racket or whatever, and this works very well, there is no culture war between soccer and handball over what the rules should be. Just let the associations make their own rules about the balls of the players as well.
But even so, it's worth noting that AI which is smarter than people in the same way that a car is faster than a Cheetah would have a massive massive impact on the world.
Yes, this part I completely agree with. At the moment, we have not reached a market-place equilibrium. Even if Claude 5.5 Opus was the most capable model which anyone would build forever (which it is not), I think it might be sufficient to render most of the white collar workforce obsolete once we figure out how to best prompt it.
Sure. There are different standards for criminal liability and for being a good fit for a professional role.
If a kindergarten teacher got acquitted of being involved in the murder of half her class because of a lack of evidence, we would not shrug and say "well, it seems she did nothing wrong, we will just let her continue to do her job, let's just transfer more kids to her class".
Well, between 1955 and 1965, the airspeed records increased by about 500m/s. If the MIC had secretly been keeping that pace for another 60 years, the current record for an air-breathing, non-rocket-assisted plane would be Mach 13 or so. Most if the records in that era were linked to interceptors, e.g. actual combat aircraft.
While the airspeed of the current generation of fighter jets is a bit uncertain, the gist seems to be that they are in the range of Mach 1.6-2.
The theory that all current US combat planes can secretly achieve Mach 8 and the US is successful in gaslighting the world about a factor of four in capabilities seems only slightly more plausible than the US having FTL capabilities due to reverse engineered extraterrestrial tech.
Maybe there are non-additive effects in Neumann's intelligence so difficult for natural selection to bring this level of intelligence to whole species even if it sometimes produces it accidentally.
I doubt it, personally. Consider the Ashkenazim intelligence hypothesis. Within just a few thousand years and mild selection pressure (bright kids tended to become clergy, and rabbis generally had a higher parental investment in their kids than Christian priests), they got an ethnic bonus to average IQ over the gentile population which is probably in the range of 5-15 points, at the cost of higher disease prevalence. If these extra ten IQ points would have been generally useful in the African savanna or in the iron age, it seems likely that life would have found a way to grant them to all of us.
I think one thing working against human evolution just increasing human intelligence is the risk of alignment failure. An IQ 100 man or woman might have been more likely to procreate than one with an IQ of 120.
I am also not convinced that he was in fact committing treason against the US in the way the Rosenbergs were.
I mean, if he was caught making a backup of the nvidia R&D servers, or installing known spies in technical positions, that would be arguably almost as bad (but even then, GPUs are more dual-use than nukes).
Of course, if your CEO has direct access to the layouts of your chips, your security would also suck.
If fucking spies of hostile powers was treason, than half of the Bond movies would end with his execution. It seems likely that the NSA was tapping into his pillow talk, and if he had actually disclosed any secrets he would already be in a federal cell.
Of course, there is plenty of space between "not treason" and "good judgement". Having sex with enemy agents on your own initiative would likely get 007 fired or demoted to office duty in the real world. Likewise, as an officer in a US company dealing with export-restricted goods there are billions of hot people you can easily fuck with little effort on your part. Nobody would care if he was getting pegged by hookers inside the cadaver of a whale, as long as neither the whale or the hookers were linked to China.
Agreed. We are not using formally verified code at the moment because formally verified code is very expensive in dev time to write. It will remain much more expensive than non-verified code, but with LLMs the cost of dev time trends towards zero.
Obviously one hard task is defining the security constraints: if the specs are wrong, it does not help that your code follows the specs.
Luckily, I think the exploits mostly plaguing us at the moment (RCE, memory disclosure, escalation, etc) can likely be formally specified without too much trouble. "There is no way that the instruction pointer will ever point to something which the attacker sent" leaves some loopholes (perhaps the attacker might still be able to construct shellcode on the victim through other means than transmitting it as a payload), but would still rule out an entire once common class of exploits.
An IT-sec future where our main worry is what the best attack against Rijndael might actually be seems good enough for my taste.
At this point, it seems the only thing that would stop superintelligence from forming is an unknown unknown, as in there is some fundamental constraint or that hasn't been identified, or an emergent phenomenon that only reveals itself in the future.
We do not know what resources are required to scale intelligence further up. This is a known unknown.
As an illustration, look at airplane speed records over time. From 1900 to 1945, they rose linearly to the speed of sound. Then between 1955 and 1965, they exploded to 925 m/s. A decade later, the SR-71 offered a modest improvement to almost a kilometer a second. That was 50 years ago, and no new records have appeared since. (Yes, I am skipping rocket planes here, their fuel economy makes supersonic jets look efficient, which is why they remain a curiosity. As are scramjets.)
Nor is this because humanity got bored with airplanes. We still invest untold billions in the R&D of combat aircraft, yet we have run into fundamental limitations where going for a higher speed is just nor worth it.
I want to make two points with this analogy.
First, it is very possible to hit a technical limitation which causes your linear extrapolation to plateau.
Second, for solving problems in the real world (like building faster airplanes), sometimes raw intelligence is not all that useful once you have enough of it. The 1960s airframes were calculated on tube computers, and could arguably have been calculated with slide rules (with longer R&D times and higher safety margins). Since the 60s, we had five decades of Moore's law boosting the effective intelligence of aerospace engineers. This helped a lot with some aspects like fuel economy, but little with maximum airspeed. Obviously it is impossible to say "this problem can't be solved by throwing more intelligence at it" without trying, and it is always possible that the critical breakthrough is just behind the next +5IQ point increase. For airspeed, I am not optimistic that GPT 20 will be able to design a hydrogen fueled, air breathing, useful plane which has a specific impulse of 50km/s at Mach 10 (which is not ruled out by chemistry).
If we ever want to get off this rock, a space elevator would be really useful. On paper, sp2 carbon barely suffices if it has no defects. It seems entirely plausible that the only gain from going from Claude 5.5 to Claude 20.2 will be in the confidence with which the latter will reject the practicality of this.
Likewise, intelligence helped a lot in the human ancestral environment, but only so much. If von Neumann would have had a 5% higher inclusive fitness than your average IQ 100 person (e.g. by being able to better outsmart sabre tooth tigers or whatever), humans would have his IQ as a baseline, it is not like his brain took twice the resources to run.
Already today, the AI companies are arguably struggling to find use cases where the intelligence of their marginal models really help. This might in part be due to them having peculiar strengths (like software and math) which are not all that relevant to most human endeavors. While I use Claude a lot, I am not paying for Fable. I tried it a bit when they included a free demo, and simply do not see the improvements over Opus being very useful to me.
If we are lucky, RSI will not yield sufficient algorithmic gains to pay for itself and the industry will converge on models which merely replace PhDs rather than models which bring forth superintelligence.
Of course, this is only conjecture, there is a finite probability that right now some research LLM is making algorithmic progress which will enable me to train Fable-class models on my gaming PC in a week or so, in which case we likely will have a singularity (in the sense of "I can't even speculate what might happen afterwards") next year.
And if you assume that it will be the undisputed leader but only in the relatively small market of coding assistance, it's wildly overvalued, even at $1T.
This seems to be disputable? Enterprise software development revenue in 2020 was around 241G$. AI makes software development more cost-effective, but this does not need to lead to less total spending, it might mean more spending as more projects become cost-effective. If Anthropic can corner the AI software assistant market long-term (which seems unlikely, but it's your hypothetical), then it might well have a revenue upwards of 100G$/year. Again, in 2020 MS had a revenue of 143G$ and a market cap around 1.5T$.
I think it boils down to them putting a lot of investment in a rather small package, with little in the way of benefits for the surrounding town.
If someone wants to build a botteling plant for 10M$, the potential downsides for the municipality are limited. At the worst, they can just buy it for that price (on credit) and bulldoze it. If someone wants a 1G$ data center, then that option simply does not exist. They will insist that the data center will remain in operation come hell or high water. If the data center runs smoothly, the town will barely notice in terms of indirect economic effect, but if there is a problem, for example unemployed masses trying to burn it down or military adversaries bombing it, they will be in the thick of it.
In a perfect market, this would lead to an equilibrium where the data centers pay handsomely for the privilege of seizing control of town politics in all matters which affect their operation. The second problem is that nobody is trusting local politicians to make good Faustian bargains with large multinationals. They are easily bribed or hookwinked even by lesser entities. And the companies do not actually need to fool or buy the median mayor, but merely the 95% percentile stupidity or corruption mayor.
So if you hear that your town is making a deal to build a data center, you might ask yourself if you are willing to bet the future of your town on the company having randomly offered a fair deal to your mayor because they want to establish themselves as a trustworthy partner in coming decades rather than them having offered every town the same seemingly beneficial deal and your mayor was one of the few who thought it would be a good idea. I know which side I would err towards.
This sounds terribly inefficient. I mean, at the moment, people could be assembling assault rifles in their garage. Yet 99% of the rifles are assembled in factories instead.
There are some steps in the production chain from raw resources to explosive drones which are hard to do in your garage, which is why drones are more recent than garages. Some of them have few other civilian uses. If people build their own drones for warfare, governments will just throw the model airplane people under the bus and ban the sale of certain types of electric motors (which would be completely reasonable).
So the big story is Jensen Huang had an affair with now a second woman
Unless she was Mossad or FSB, I doubt that anyone will care much. Nvidia is not selling spousal fidelity.
The first affair story was juicy merely because the woman allegedly working for the Chinese, so there is the possibility that he leaked some intel (but likely more boring high-level business stuff than chip designs) or let his affair influence his decisions.
obviously consent is an issue here,
Yes, obviously. Nobody cares if you are fucking another tinder/grindr date every night. If you care about your spouse not having sex with others, contract law seems sufficient to set incentives. If you are a sex pest, that is quite something different.
- Prev
- Next

If we are talking about symbolism, I think there is something to object in most methods.
Gas chambers: Were used to murder more than two orders of magnitude more innocents than Robespierre did.
Hanging: Was widely employed to murder Black men in the South in the decades after the civil war.
Firing squads: A favorite of genociders who do not have the infrastructure to run proper death camps.
With AI posed to take most intellectual labor, we might want to avoid automating what does not really need to be automated. Decapitation by a
headmanheadperson wielding an axe are probably among the traditional methods least tainted by association.Or, given that most proponents of the death penalty are conservative, the US could RETVRN to the best and manliest society ever and bring back crucifixions. I am sure the present SCOTUS would permit Hegseth to crucify condemned along Pennsylvania Avenue, and it would be a beautiful celebration of his beliefs.
Sadly for him, Trump would probably chicken out again, as he did with the public firing squad plan.
More options
Context Copy link