I think its a mix of both of your first two points. He has hinted that he will be producing chips for Tesla and SpaceX here in the US in the future, so there will be a time where Elon will have every critical component we need for Tesla and SpaceX here in the US and not overseas, and fully in-house. Once he does this, nothing stops him from selling components. He's also insanely efficient at standing up data centers, so I would not be surprised if he gets enlisted by someone like Anthropic to setup infrastructure for him.
That's very doubtful given recent news that Colossus is running at 11% capacity and has hundreds of thousands of idle GPUs. xAI acquired too many GPUs and currently doesn't have enough customers to use them. That's why they are making compute deals with Anthropic and Cursor.
xAI is bleeding money and this compute deal with Anthropic will pay for all of xAI's capex ($25 billion) in 2 years.
11% capacity is not 11% MFU. The first is about actually using the hardware for something in the first place, the latter is about how efficiently you compute. Different things.
The original post of 11% is referring to leaked numbers about _MFU_, which has erroneously been re-reported as fraction of GPUs being used at all. The parent post is trying to correct this misconception.
Nevertheless, you are still talking about 15B / year - when the initial investment was < 30B.
That investment also probably covers things like power, salaries, etc.
Some of the proposals are laughable, some of the example calculations are. The idea of running AI training sounds extremely challenging. But the idea of inference in space doesn't seem absurd
The power budget of a starlink v2 mini satellite is estimated at around 20kW based on the known solar panel size. That also matches what a satellite of that size would roughly dissipate without dedicated radiators, using just some heat pipes to spread the heat evenly over the satellite's surface. There is nothing fundamentally preventing you from taking the same satellite design, remove most of the comms payload and instead put 15kW of GPUs there. Or about 10 GB200 including the CPUs, networking, etc that you need along with the GPUs
Now, do the economics work out for $300k worth of compute for each satellite, in an environment where maintenance is impossible and degradation will be higher than on the ground? Probably not right now, but in a couple years they might
The starlink satellites have hall effect thrusters, and presumably beefy laser and radio comms systems that will account for a lot of the energy budget. Also, they are sometimes in the earth's shadow, a naive calculation says they have 2x the solar budget they need, in order to charge up batteries for shade-time operations. There is no where near 20kw of compute in a single satellite, and thus no where near 20kw of heat to get rid of.
Furthermore, xAI's colossus supercomputer is specced at 250MW. And this seems to be a number that'll just increase over the coming years with new bigger DCs.
To match this level of performance they will have to launch what, ~15k satellites _per_ equivalent datacentre?
Regarding cooling: you can't just cover the outer surface with pipes. You cant't dissipate the heat, you need to _radiate_ it away. You need to point that surface to the deep dark cold of space. If you point it to the sun, you will heat your satellite. Think a massive "reverse solar panel" that works with infrared. You need surface area, and loads of it.
I'm not saying this is impossible. Obviously elon will prove us all wrong because he's stubborn like that. But there is no way this will ever be economically viable when competing with terrestrial based systems.
The assumption behind "spread the heat over the size of the satellite" is radiative cooling. Admittedly my numbers were a bit off, you need to make the satellite a bit bigger, or use some of your solar panels. A starlink v2 mini has 8m² of area per side, so 16m² total. To dissipate 20kW to space, at a surface temperature of 80°C and emissivity of 0.85, you need about 28m² of space. So you need to increase size a bit, or add 12m² of dedicated radiators. A bit more to deal with the real live complications (the sun exists, the earth is actually warmer than space and covers a significant portion of the sky if you are in leo, etc.).
The actual Starlink V2 Mini is has estimates for solar generation that range up to 35kW. We need a bit more, but not much more.
My numbers are a bit optimistic, but they are in the right order of magnitude. Power and cooling are very achievable in the area of putting one server on one satellite. It's the "putting one DC on one mega satellite" ideas that run into feasability issues, and for inference those aren't needed. The economics are the bigger issue, and launch costs are only moving in one direction
How is this not absurd? What is the benefit? Space is a harsh environment, with issues due to solar radiation etc, etc. And it's permanently 100ms away from any user.
You’re making some decent points here, but you’re either forgetting or ignoring the major thing that people usually neglect to mention when they want to make a case for this (crazy) idea — weight. Unless SpaceX is going to completely redesign hardware such that it is optimized entirely for its mass, it requires many (many) launches to even get a small set of racks into space. I don’t normally get up in arms about the CO2 emissions of data centers, I think there is offsetting value created by their use, but I would absolutely protest trying to put data centers in space and do my best to shut down the hundreds, if not thousands of launches it would take to achieve even a tiny fraction of an AI data center.
Have you calculated the CO2 output of Terran datacenters run on natural gas vs. space datacenters run on solar? The launch CO2 usage is one time, the datacenter energy is for the life of the equipment.
NASA had literally done it. It was never laughable, just thought to be an incredibly difficult engineering challenge.
I think compute in space suffers less from being "impossible" and more from being "impractical". It is plenty easy to put compute in space. It is just still silly expensive and by the time your equipment makes up the cost of putting it in space, it will be well out of date.
You don't have to actually have an orbiting datacenter for the idea to work. You just have to convince enough people. Once you've done that, you can claim that regional regulations don't apply to your data because the data is in orbit. Its not like somebody is gonna go up there and catch you in the lie.
Out-of-regulatory-reach is what they'll actually be selling. It can be on earth, it just has to be sufficiently hidden such that you can claim that it's in space.
I don't think that's the purpose and if it is, it's obviously trivially solvable. The US can and does simply assert jurisdiction all over the globe. There's no reason it can't or wouldn't extend that to orbit.
Space compute isn't real though, it's just a scheme to pump the value of SpaceX before IPO by associating it with AI. It's really hard to cool things in space, because there's no matter to transfer the heat away. All you've got is radiative cooling, and that's really really slow.
He believes in the value of the idea of “space compute” for attracting investors to SpaceX. But the existence of the idea of “space compute” as a better way to deploy datacenters (along with everything else Musk has claimed in the past decade) should give everyone pause as to the plausibility of literally everything else he says.
That's Intel. Probably IBM too, though they've been doing mass manufacturing with TSMC (and GloFo before) for years instead of their fabs. I wouldn't be surprised if HP did similar things back in the 80s.
Intel and TSMC are not what they are today just because they buy a very expensive EUV Machine from ASML but because they have the knowledge and infrastructure to even use these machines.
What do you mean by etching? Google does also it's own chip design with TPUs, data centers, and models but afaik only TSMC Intel and Samsung do the actual semiconductor fabrication
Here's a probably stupid question - if someone were unbounded by ethics and conceivably had enough power and connections to power to shield themselves from many consequences of their actions - and that person owned these DCs, could they in theory observe all the streams of tokens coming in and out of these models, and even exfiltrate copies of these models wholesale to have their own teams do what they will with them in the pursuit of building their own competitive models?
Or is there something fundamental in the way these models get deployed (encryption or something or than legal contracts?) at this scale that prohibits the owners of the infra from gaining this level of insight / access?
1) The situation you described would be covered under the contract between Anthropic and xAI, and that any violation of that would be subject to financial penalties and legal proceedings. The US has a robust corporate legal system, and disputes do get resolved through the court system, although in a slow and costly manner.
The contract can stipulate a penalty at a high enough amount to discourage this behavior.
2) Output from models & intra-datacenter communications can be encrypted if customers truly cared.
3) There is no reason do this, because there are far better ways to exfiltrate data from Anthropic models. Chinese companies are already doing this at an industrial scale where they are reselling Claude tokens for 10-20% of the cost while retaining the data to train their own models. https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens...
If we look at Deepseek V4-pro, created by Deepseek who Anthropic formally accused of harvesting Claude tokens at scale, it performs the same as Claude did 6 months prior.
It's an accurate statement. The US (specifically Delaware) has a world-class corporate law system. Delaware has a court dedicated to only corporate and commercial disputes, with 200+ years of case law.
It does until you're embedded enough with the surveillance system [1]. If a company was able to get immunity in the wat AT&T got it, no contract would protect the other side.
> There is no reason do this, because there are far better ways to exfiltrate data from Anthropic models. Chinese companies are already doing this at an industrial scale where they are reselling Claude tokens for 10-20% of the cost while retaining the data to train their own models.
I think you missed the part where Anthropic stopped displaying their thinking tokens over the past few months, and instead now provides “summarized thinking”, letting Haiku summarize Opus’ thoughts.
So it is now much more difficult (impossible?) to distill the models.
I also think you over-estimate how well the legal systems works in the US nowadays, and under-estimate just how much power Elon has in the government.
At this scale thinking tokens don't matter anymore.
In Feb Anthropic called out three Chinese labs for "distillation attacks", but a lab missing in their post actually had most Claude generated tokens among all Chinese labs in their midtrain data :p
It's already in the public domain (thanks to the OpenAI trial) that Grok distilled OpenAIs models. Listening to the data going into the models in the data centre would be very similar thing. There's some downsides (you're passively listening, not controlling the queries), and some upsides (way more data). But it only ever gets you to some percentage of the existing production model. It doesn't get you what Musk wants - an AI company capable of designing and deploying leading edge models. It gets you to fast follower status.
You could, but Grok is pretty high up there, it might not be "#1" but its definitely up there with the giants, people seem to overlook it. Gemini has a similar problem, it was #1 once, and it seems like Google isn't hell bent on chasing #1 they just want to keep iterating over time, they know they just need it to be "good enough" and they'll keep having repeat customers.
If Elon REALLY wanted to do anything like that he would be better off poaching talent from competitors, less legal hell to go through.
No Grok is not 'up there'. Not by any stretch of the imagination is it anywhere close to anthropic or openai in any single domain whatsoever. Not even close to deepseek. It really isnt a good model. Their research team's talent is just not good, unfortunately.
If I was training a top-tier model, having a competitors’ weights feels like it would be an excellent refining tool. As a user I find cross-model iteration to be a huge power-tool, and presumably the boffins can zero in on areas of relative strength and weakness and work out which area they want massive amounts of synthetic data from.
Bear in mind these data centers were built for X.ai to use themselves, so there would have been no reason to backdoor them. Also, this is all off-the shelf equipment: Dell & Supermicro servers with NVidia compute modules and ethernet switches.
xAI had a lot of negotiating power here because Anthropic had ~0 comparable options and ultimately desperately needed the compute now. So, it wouldn't surprise me if data sharing was an explicit part of the agreement
There's accusations that the Chinese labs have done essentially that to OpenAI and Anthropic and exfiltrated their models without having DC access, so if you had DC access, yes, you could do that. If you had DC access though you could just copy the model onto an SSD.
The Chinese labs have been accused of training on elicited chat logs on a massive scale in violation of ToS. That's possibly a real concern, but it's nowhere close to "exfiltrating" the model or even roughly matching its behavior.
xAI gets the cashflow and makes spaceX bottom line more appealing.
But that guy Musk hardly makes deals that favor other companies over his, so what am i missing?
xAI is bleeding tons of cash, and they don’t have the ability to use the compute for robotaxis, which have been a disaster. Grok monetization is also low.
I think that's the datacenter with the gas turbine generators that operate without permits because they're "portable." Data centers have tremendous externalities but colossus is a particularly nasty offender, and not just due its size.
The newer location is about 3 miles southwest of the Memphis Airport (MEM), one of the world’s largest cargo airports and the center of FedEx operations (500 take-off and landings per day most concentrated in a 6 h FedEx window with lots of engines running on ramps and that produces about 2000 tons of NOx per year).
I live about 18 miles downwind of the new Colossus sites, the airport, and lots of truck logistics sites, and a large refinery.
I definitely will be getting 2x exposure to ozone and particulates from both Colossi when they are running full bore. Plus an extra dose of ultrafine particulate with my morning fresh air.
Yes, wouldn’t it be nice to be in Nashville instead with HCA, Oracle, many insurance and financial institutions, and the joy of country music.
As an avid Opus user I am in an ethical Nimby bind. We do need almost any investments we can get in Shelby County TN. I’ll take Anthropic in preference to Grok NOx. And it will be my NOx.
The biggest issue with the interim onsite generation is the lack of meaningful stack height on the generating units.
Airplanes by virtue of their mode of operation stay out of the unhappy regime most of the time. Also, engines at/near idle produde orders of magnitude less emissions. Those aeroderivative generators are running at full capacity 24/7.
Dumping exhaust at ground level continuously is probably much worse than the airport. Even if it's a FedEx world hub.
Solar and batteries are a bad choice for a constant 24/7 load.
That's the exact reason we will never go fully solar (or wind) unless an insanely impressive battery breakthrough makes storage effectively free while using only common, renewable components rather than rare earths.
Solar, wind, etc are excellent parts of an energy system, but its nearly impossible to cover base load at scale with generation that may only run for 0-5 hours a day.
Sodium is an interesting chemistry, though it has different voltage curves than lithium ion once hardware is built to match it they may scale well for industrial use.
That still doesn't avoid the more fundamental math of having to store such massive amounts of energy though, even if you skip batteries and pump water to retention ponds uphill.
Even for small scale residential, the recommendation is to store 3 days worth of average usage to handle stretches of cloudy days and to have a generator for the times when that still doesn't cut it. You also need enough solar that 4-6 hours of generation can fill those batteries back up after a stretch of cloudy days.
You also have to contend with frequency issues. This is what took down Spain's grid, they had turned on a ton of solar at the time - with many gas plants offline a seemingly small dip below 60Hz really wrecks much of the system when it wasn't designed to handle those swings and triggers multiple safety mechanisms.
> Jordan is suspiciously lurching to the extremely high energy end of the light spectrum when we know that the low energy end (comparable to infrasound) doesn’t have negative impacts on us if we can’t detect its presence.
Masley spouts the falsifiable propaganda that any photon (light/emf) below ionizing radiation energy can only cause “heat” and can have no other possible harm effects.
Many science-minded people (though more accurately in this case, billiard-world materialists) have become quite militant defenders of this idea (ostensibly fighting off the hoardes of tinfoil hatters and quantum aquarians sensitive to 5g).
This point is very plausible military industrial propaganda. There were numerous studies with evidence (starting from the 60s) such as - non ionizing radiation (eg hv powerlines) might cause lots of cancer, and it’s weaponized (sub-thermal) usage can microwave the brains of enemy spies. THOSE studies have come out in declassified and leaked docs.
Now we have several plausible and serious theories of mechanism for low-f light disrupting biology that lean into quantum biology. While the iceberg of quantum biology understanding is still in its early decades, the mounting downstream evidence of health and medical issues are established public knowledge.
Interesting, I read this series of posts and as someone who does not have a dog in this fight but does have a more than passing background in both audio engineering and datacenter engineering, the response Masley gives here to the very first criticism is fundamentally incorrect. I haven't read the rest of this, but his claim about sound intensity what it would imply about energy is on its face untrue.
When you take a measurement of a sound, you are measuring both its pressure and its intensity, that is what is implied by a measurement in decibels. The measurement is taken from the point of the measurement device/listener in relation to the source/generator. If the measured value is potentially harmful, there is no such implication about needing to redirect additional energy to make it harmful, it's already been measured as potentially harmful at the point of measurement.
It's basically nonsense. My most charitable interpretation of his very first responsive argument is he's saying that a datacenter would need to intentionally direct energy towards increasing the intensity of its sound output to make Jordan's original measurements meaningful. That's neither how measurement works, nor how sound works, nor even how datacenters work. Things like sound and heat are BYPRODUCTS and not the point of the datacenter, both have an intensity, and that intensity is measurable, and any energy which is expended towards that intensity is energy that was wasted away from doing computations.
I stopped reading after this. I don't know if Masley is out of his element or just practicing motivated reasoning and thinks his readers are stupid. Either way, his rebuttal already failed on the first point.
I dont know, Id say the enraging thing is that the government is so incompetent and unable to expand electricity supply that datacenters are forced into using loopholes to get power the only way they can.
> Id say the enraging thing is that the government is so incompetent and unable to expand electricity supply
So let's say you're a homebuilder, if I tell you I want a new home and I want to live there tomorrow, you can all of a sudden build it in a day, right?
Electricity use is skyrocketing for various reasons, these datacenters being one of them. There are a lot of countries struggling to keep up with demand. So incompetence? No, probably more like supply lagging demand.
Or ASML and Nvidia and all also are incompetent, because they didn't see demand coming....
ASML and Nvidia did see the demand coming. Intel didnt, theyre incompetent.
> There are a lot of countries struggling to keep up with demand. So incompetence?
Mostly, yea. If the government gets out of the way there are plenty of people who would be happy to supply power in exchange for money. The homebuilding analogy is good because homebuilding in the US often has the same issues that plague the electricity market. There's no reason it should take years to build homes in many parts of this country, but it does.
In the early days of this the AI companies were asking for massive new energy supplies but also refusing to sign contracts to pay for it over the decades of its life.
They're basically attempting to game the system, politically and economically, to put as much of the cost on taxpayers and ratepayers as they can. This naturally slows things down.
I'd be curious to hear from the silent down votes. Do you disagree that power companies are effectively regional monopolies? Or do you disagree that government oversight and regulation isn't the reason for poor maintenance and lack of capacity?
These are the same companies and individuals who are actively working to destroy functional government, and are happily looting the US treasury rather than let it be spent on things such as encouraging more energy production.
Seems like selc's time would be better spent trying to close the loophole that allows for unpermitted turbine generators instead of going after one company for doing what they were allowed to do when they did it.
Have you considered that the march of progress requires human blood to grease the gears and mulched skulls to pave the (highly efficient) road? Really, when you take into account all of the future lives this will improve and save it's difficult to claim any cost now is too high. Would you stand in their way and delay the day that Mythos cures cancer?
I wonder what percentage of GDP expenditure will give us SkyNet.
Undoubtedly, it will find cures to all cancers… The ARR and stock appreciation will be amazing. Except the cures will be found long after it has wiped out all humans.
I'm not saying human blood and mulched skulls are a renewable source of power, I'm just saying. Or maybe they can partner with SoulCycle to power computation with 24/7 spin classes?
I'm going to bet it's simpler than that. I'm betting they changed the script for the Duracell product placement that made an acquaintance of mine a ton of money for pulling it off. Always follow the money.
If it was indeed the original script, the reason they changed to batteries is maybe not because "some exec" is an idiot, but because it worked better from a storytelling perspective.
Even if treating people as batteries doesn't make much sense as we are pretty terrible power plants, the message is clear and impactful. It is common for movies to oversimplify things, because they want to avoid having the viewer from being distracted from the main plot. It is tricky, as being too obviously wrong can breaks the immersion. I think the people = batteries analogy is a good compromise. Brains = computers, while technically more plausible would add a layer of complexity that could be a bit too much for a 2h action movie.
I don't think it'd have much of an effect on the story, outside of background stuff like the Animatrix, it's just an interesting little fact about why the world is the way it is. Shooting and hitting things in slo mo is still the core either way.
Sam Altman and Dario Amodei are already billionaires. That doesn't stop those greedy fucks from stealing everything in their path in the pursuit of money.
xAI’s turbines produce meaningful local/regional pollution (especially NOx in a vulnerable area) but represent a rounding error nationally and globally.
> xAI’s turbines produce meaningful local/regional pollution (especially NOx in a vulnerable area) but represent a rounding error nationally and globally.
If you shoot someone in the face, it will produce a meaningful increase in local/regional murder, but represent a rounding error nationally and globally.
It doesn’t matter if people have to suddenly live by gas turbines that run 24/7 because why again? Can you repeat that last part back to me but say it a little dumber for me?
What's the tremendous externality of gas generators? People heat their own homes with natural gas and it's no big deal. How can a datacenter that is miles away be worse than that?
Its totally inefficient - burning the same gas in a co-generation plant, ideally combined with district heating, would produce the same amount of pollution and basically make use of all the energy.
Why is xAI giving up their advantage? Is this a signal that their frontier model improvements are plateauing and decided there is no value in hoarding all their compute?
xAI have (with some questionable ethics) managed to actually build a data centre, so they have a ton of compute but not much inference demand for their model which is second-tier.
Everyone else trying to build data centres is really struggling (turns out building physical things is not as easy as writing code, who knew).
So Anthropic have the model but they're compute starved because everyone else they've signed agreements with still mostly have piles of dirt and the GPUs are still in Nvidia's warehouse somewhere.
It's a bit of a win-win: xAI's financial numbers will be massively improved by the revenue using otherwise useless data centre capacity, Anthropic get the compute they desperately need albeit probably paying through the nose for it
Musk's goal with X.ai from the start seems to have been primarily to compete with OpenAI, and given that Grok itself isn't doing that (not even mentioned in competitive benchmark tables), and his data centers are therefore mostly sitting idle, this deal with Anthropic helps him achieve that goal by helping OpenAI's #1 competitor.
Note how the initial deal was only for the older Colossus #1 data center, and now after losing his lawsuit against OpenAI the Anthropic deal has been expanded to Colossus #2 also. Coincidence?
Elon got rid of most of the core xAI researchers recently. It looked like firings and he said something like Grok needs to be rewritten because it was fundamentally built in a bad way. My guess - Grok was not one of the top two models and had no traction beyond users on Twitter - so investing in it did not make sense. Or maybe he needs time to build a new team and rebuild Grok, and while he’s doing that, he needs revenue to make up for all the GPUs they bought.
Either way, I doubt he has any ‘real’ advantage in this game. It’s OpenAI and Anthropic, and then everyone else including the open weight models.
Boy I’d love to have the sort of ‘no real advantage’ that got me $1.9bn in revenue per month from Anthropic.
Make no mistake - Elon’s charging premium pricing because he has something Anthropic realllly needs. In exchange, he will get significant revenues added to the balance sheet along with the cursor team right around IPO. It’s very good business for Elon and SpaceX.
What advantage? Has there ever been any indication they’re leading in any segment? Sure Elon has thrown a bunch of money at hardware, but to what end?
And frankly as bad as Altman is from a: if AI is really going to disrupt humanity do I want this guy in charge? Elon is 10x worse. So why would the best and the brightest ever work for him?
yes, but it's only advantage if one is compute-strained and the other isn't. if they both have lots then there's no advantage. if one doesn't fully utilise their compute then it's not an advantage either
Well, other than your ability to turn that into cash by renting it out for the highest price to someone who needs it, you can promise prospective employees that are supposed to use that infra to train models that they won’t be compute starved.
You can kick off more model training runs and experiments than your competitors.
You can kick off a $1-2t IPO claiming you are going to capture a large portion of the largest TAM the world has ever seen.
Could have (past tense). They now have rented the capacity out, presumably with a multi-year contract.
A rental deal, even at a premium price, actually makes great sense for Anthropic. Rather than building themselves, they are de-risked by renting instead, and they don't own the depreciating asset. They've also compute-starved X.ai as a potential competitor as a marginal added bonus.
They have neither the most resources nor the best models. They are mediocre at everything except the CSAM generation market, they've got that one cornered
I see this as a huge warning sign. If a frontier ai lab is in this position of renting their own capacity, imagine how much overcapacity there is in the system.
Outside of VC money, and circular financing, the only external money coming into ai are into open-ai, and anthropic via their subscriptions and APIs.
I think any warning this provides is more about the ability to forecast demand. Obviously it was a giant leap of faith, that didn't pan out, for Must to build these data centers without having a product with a growth trajectory to more rationally justify it.
OpenAI and Anthropic (and their investors!) are obviously also out on a limb here, both with deals in place to 10x their available compute (roughly 1GW -> 10GW) over next few years. Maybe these growth estimates are a bit more grounded, but they have all sorts of assumptions baked in. I still would not be surprised to see OpenAI self-implode, and to see Google and Microsoft as the eventual winners here.
For sure. There’s also this kind of interesting meta strategy question: if you think you can float the huge capex with some internally valuable processes as a base case then you might just go really big. Dario and Elon’s temperaments come in to play here
They were briefly SOTA on some benchmarks, although there were suggestions there might have been some massaging of the results since real world usage showed lackluster performance compared to the benchmarks.
More signs that xAI might be giving up on the AGI race. xAI let Cursor train a model on Colossus 2, gave the entire Colossus1 to Anthropic, and is now giving compute in Colossus2 to Anthropic as well.
Bad read on the situation. xAI has too much compute and not enough customers using it. They have around half a million GPUs, some of which are stolen from Tesla, running at 11% utilization. xAI predicted more people would be using Grok, but Grok is not a SOTA model & users primarily want to use SOTA models. They have excess capacity and it makes sense to rent out GPUs to other customers while they improve their models.
I keep hearing this statement, and it always makes me wonder if people have actually used Grok…
I have a Claude Max plan I use for coding, but I also have a Grok Lite plan I use for web search type tasks (similar to Perplexity) because I like how the Grok harness handles searches and I don’t need a SOTA model for that use case. I’d never pay $30/mo for a full SuperGrok account but to me it’s worth the $10/mo for Lite as I was hitting limits on the free tier.
I’ve never noticed it to be particularly biased at least for anything I’ve been searching for on it. And on the other side, I’ve never noticed it to be particularly less censored or anything compared to other models either (also a claim I’ve heard a lot about Grok but I think because it is/was part of their marketing).
Opposed to all other models being the bastion of objectivity? Must be truly vindicating to have to hear other peoles opinions after decades in the silicon valley bubble.
There is a difference between when somebody openly instructs their model to infer disproven lies vs who doesn’t do this. And it’s quite tiring that this is even a question because of politics.
As somebody from Hungary: the biggest impact of my mood was that this kind of thinking went back with the collapse of far right there to where it belongs: to a deep hole which is not in front of normal people. Average people suddenly don’t ask illogical questions or answer stupid things because there is nobody who would tell them that they need to think stupidly, there is nobody who tell them what stupid thing they should think that week. It’s marvelous when you get the proof that the whole “stupid thinking” is completely controlled from above.
In end user perspective, it’s the same. The difference is probably volume. I have no clue in which direction. Both in trained lies by models and in number of people defending the indefensible. But one for sure, there are probably 10s or 100s of millions of Chinese who try to do the same. I encountered with it quite frequently. Sometimes with flat out doublespeak.
As a non-US AI user I do not particularly like using a US model following the recent political events, but I specifically do not want to use a model made by an ex-member of the current administration.
It is always great fun using a model via API without search/web access and quoting a recent development, being told that it must be hamfisted satire, then providing access. The reasoning traces are a delight, Opus 4.6 and GPT-5.4 during the administrations war with Anthropic were prime grade A kobe beef.
This comment is very similar to what russian propaganda does.
It's not aimed at convincing you to support them, but to convince you everyone is lying and there is no meaningful difference between each position, so you stay apathetic.
They tried and failed. xAi made a mistake building Colossus 1 and ended up with heterogenous cluster of H100/H200/GB200 GPUs. This is a nightmare to train huge models on because each card has different specs, features, and hardware requirements. During gradient synchronization, a heterogeneous cluster would bottleneck on the slowest GPU (H100) so the faster GPUs would end up idling. They also probably ran into unexpected compatibility issues, which are difficult to resolve.
It makes more sense to use this cluster for inference, since they can segment the cluster by GPU type and avoid GPU mixing. xAI doesn't have enough inference customers so it makes sense to monetize this to companies that need inference compute such as Anthropic or Cursor.
Apparently xAI will try building SOTA models on Colossus 2, which will be built on Blackwell GPUs only.
How can something so obvious be overlooked by team building the data centre? Can't the sharding be uneven so that weaker GPUs still finish fast by taking on a smaller workload?
It's not like they had much of an option, when everybody was hoarding every GPU they could. For the second Colossus they could book future production, but the first one had to be built ASAP so xAI looked as a serious competitor in the AI space.
More people should try Grok. I don't use it for coding but it's replaced a lot of my ChatGPT usage. Definitely more perferred model for quick questions or easy answers.
One thing I do like about Grok is that it makes it stupid easy to see what its referencing, and gives you the links to those resources. Which most models sometimes either don't bother, or don't do much of a good job of doing. It's not the top model, but it is definitely high up there, people's blind rage for anything Elon Musk is the only reason most people don't realize how capable it is unfortunately. Grok is not exclusively made by Elon Musk, there's definitely other engineers working day and night on it.
For conversational or general knowledge questions I also much prefer Grok. Musk's vanity aside, it is much less censored than the other frontier models.
Far more mundane that whatever bogeyman you imagine, it will discuss things like COVID origins or not treat favored classes as a special case e.g. when asked to make a joke.
Elon lost his lawsuit with openAI and knows xAI isn't on the same trajectory. Might as well try to win the bet and flip off Sam by supporting the best competition. Also they are getting a head start on AI as a commodity. I'm sure there's plenty of money to be made for those that can leverage their capital to essentially rent capacity right now. If he's not making enough off of grok, might as well cover their expenses.
It’s a very fractured and heterogeneous landscape where your own perspective will be warped by your personal experience.
Anthropic has a lot of the market share and dominates the mind share, but each of Codex, Devin, Cursor, Claude, et al have significantly more market usage than they had 6 months ago and each are likely still growing very quickly based on publicly-reported information.
I don't see a scenario where it really makes sense to be a frontier lab long term. Eventually model quality will plateau then you distill and get 90% for 10% or less cost.
General AI is that scenario. The investor dream is that their horse hits general AI first, patents it (or otherwise somehow stops the competition from hitting it), and then reaps the massive benefits.
I'm not saying it's a likely scenario, but I genuinely believe a big percentage of AI investment revolves around that (or similar) scenarios.
Opus and GPT are so different there’s room for both and I wouldn’t be ignoring Gemini even if it isn’t ahem great at coding, since it’s quite obviously very good otherwise.
In the times before you also would rather have very smart people working together instead of one very smart dude alone even if he had an identical twin.
I don't like youtube gemini summaries. whenever I have asked gemini about detailed comprehensive report on a long hour podcasts, it always refuses. hence claude is my go report generation AI.
This might come down to when you expect plateau to happen. You could have said similar about transistor density many decades ago.
(I'm not denying it could happen next year for all we know. But we simply don't know, and from what I hear from researchers the breadth of ideas we've tried are still small)
They make good models, at times SotA (at least if you don't need coding, their last good coding model was six months ago), with lower safeguards than either Anthropic or OpenAI, and they still fail to capture meaningful market share or mind share. The name Grok is tainted by the twitter bot of the same name operated by xAI/X. Being owned by Musk lets the company appear unstable and untrustworthy in the minds of many. Their marketing game is just bad all around. They struggle to retain top talent.
Maybe their next model will be great. I doubt it will matter. I doubt xAI siphoning off Anthropic models and distilling that would matter. Model performance is not the main factor dragging down xAI
Matmuls need access to decrypted weights to do their work.
Which means that getting the full weights out isn't even an "if" - it's "how much effort". The encryption wouldn't do much more than a gentleman's agreement would.
The only real move for Anthropic there is to outline contract penalties for letting weights get leaked, and never give less trusted external inference providers access to cutting edge system weights.
Exposure is limited either way. Opus 4.7 weights are a deprecating asset - it's bleeding edge today, very valuable now, but it'll lose a lot of its value the moment Opus 5.0 drops.
That would require hacking Nvidia's GPUs/racks to extract the weights. The weights are encrypted, sent to the GPU/rack encrypted. When it does inference, it will use decrypted weights but there is no way to get those weights unless you find a way to exploit Nvidia's GPU security.
Do you think OpenAI would send CoreWeave their GPT 5.5 Pro weights if an admin employee at CoreWeave can access the full weights unencrypted? Of course not.
It would require exactly that. A bit more involved than "scp that big file", yes. But you make a mistake by treating it as a hard blocker.
Like I said: it's a gentleman's agreement. If Musk said "I want Opus 4.7 weights", and those weights were on Colossus 1 hardware, he'd have those weights on his desktop, unencrypted, within a couple of weeks.
There's also the side channel line, because having inference on your hardware typically allows you to do things like snoop into KV cache and peek at per-layer, or even per-expert, residuals. Which allows for some very advanced distillation attacks. Might be easier/more deniable to pull that off than dumping full weights, in some circumstances.
Dude, Chinese labs distil attack via the APIs, if Musk wanted to do something like that, technically he could. Legally it would be a giant slam dunk liability though
Well, knowledge distillation requires a teacher model and a student model and the student model attempts to learn and extract and (preferrably) compress the information of the teacher model, so it is possible for model collapse due to high SNR in between [1].
What I suggested is to steal the (possibly intermediate) weight in between by sniffing the network communication bus, which means MITM for getting the exact values. Or unless it turns out OpenAI or Anthropic leveraged homomorphic encryption, or I'm not certain how is Anthropic would safely allow Mythos to run on AWS without their control.
Distilling is different from "siphoning the model weights". I would think that Anthropic has a system for this. After all, they deploy to different clouds already. Their weights are worth billions, I presume that they take security very seriously and have done a lot of homework to trust no one.
xAI cannot train models. Anthropic cannot do inference. The roles of these two companies have already been decided; one as a pure inference provider, the other as a pure model manufacturer.
Ok, maybe this question is a bit silly, but could it possibly be that Elon is doing this to steal Anthropic models secrets and using them to improve Grok?
Love the curiosity. I recommend you spend that curiosity on figuring out why both companies are in the situation they are that this deal is a win-win: Elon has provided a masterclass in public company launch and management over the last 15 years. (And in recent years done it amidst an incredibly divisive reputational environment).
Put another way - they could maybe try to do this. But it would be worth literally nothing compared to the actual benefits to SX of this deal. And add an immense amount of risk. There’s much more clever and interesting things to do.
No. Model weights are encrypted. Elon is doing this because Anthropic is desperate for compute, SpaceX needs to juice up its balance sheet before IPO, and xAI model team isn't competitive.
You won’t see “resell electricity” on the IPO brochures. They’ll say something like multi billion dollar ARR hyperscaler business.
To the extent both are true, it’s non trivial to have large grid connections in the US these days or even gas pipeline connections hooked up to generators around a datacenter. Those assets are valuable.
How come a company like Anthropic has invested in photonic computing? If it's good enough for Boeing, I'd assume they would at least invest in it. qc-LPU100 seems worth it for some matrix mult, if it can be proven its O(n) (at least less than O(n^3).
Grok’s whole thing is being irresponsible and edgy. Can’t s ee Anthropic going for that, can’t see Grok’s customer base accepting an AI that refuses to do nonconsensual porn.
I suppose you're also using Chinese models, who all have to be at the behest of the Communist party. China also have "horribly damaging" coal power plants. Do they acquire training data ethically? Do they censor? Be honest with yourself. Your concerns have nothing to do with ethic.
Yes, quite a bad reputation hit for Anthropic to get in bed with Musk. Tesla lost half its marketshare in Europe because Elon is meddling in our democracy, is promoting right wing populist parties and supporting Trump in general.
Europe has a hugely shrinking demographic, and is actively decoupling from the United States. There's almost no innovation and you are making the mistake of getting in bed with China and constantly suing our tech companies.
Europe needs to focus more on Europe and less US politics.
Denmark has a freakin' King. You guys are still a monarchy... Greenland is closer to the US than Europe, and we defended and gave back demark during WW2.
Demark provides almost no financial support to Greenland, and can't defend it. It will prove to be very strategically important as the ice caps melt.
Russia invaded Ukraine, and Demark lets russia use the Danish straights to transport oil despite the US asking them to not allow them entry. Demark has no way to protect it's own straits, how is it suppose to stop an invasion of Greenland from russia? How is it supposed to impose order and control the north artic if it can't even protect it's own straights?
These are very real security concerns from then US that a monarch can't address.
No, they’ve lost market share because they haven’t released a new product in years. Meanwhile China is releasing hundreds of excellent models into the European market which are better than Tesla’s.
Fake __AI__ revenue. Maybe Hetzner should build Colossus4, rent it out and book it as AI revenue instead of hardware rental revenue and get a P/E of 100.
Anthropic this quarter will have revenue of $10.9B, up from $4.8B last quarter[0]. They're paying SpaceX $1.25B per month for compute[1] - which is more than what SpaceX earn on space. SpaceX spent about $30-40B in capex on Colossus 1 & 2.
This is all real revenue, real spend, real usage.
Hetzner just aren't at this scale. Not even close. If they wanted to get into this business - first, they're late. Second, it's at a scale of ~10x of their total lifetime datacenter buildout. Third, they'd need to change their business to being one that is debt fronted.
xAI have proven out that being able to deploy compute is a very viable business (and difficult to pull off)
At some point AI cynicism clashes with reality, it must be exhausting maintaining it.
"I'm an investor, the founder of new alchemy, co-founder Lamina1, former CEO of coinlab.com, Chairman of the Bitcoin Foundation, Ethereum security researcher, holder of patent 9298806, general nerd. Currently working at Capital6, my private equity fund."
You poor misunderstood Bitcoin booster! Now you have to deal with AI critics! How about getting a real job?
Anthropic’s revenue is computed different from OpenAI. As I recall, they inflate it by including money that they end up just paying forward to some of the companies they depend on. OpenAI doesn’t count that component. And none of these companies - including SpaceX - have trustworthy accounting.
Jesus Christ, the Hetzner example is obviously an example of booking revenue as AI revenue (where investors assume it is generated by Grok subscriptions) vs. hardware rental revenue, which traditionally not valued as highly.
Nowhere does the hypothetical state that Hetzner, an example for hardware rental, has the funding or the capabilities to execute the sarcastic example.
But ok, now hardware rentals have a P/E of 100 or more.
Investors aren’t dumb. These numbers are being reported and the fact that the data centers are being rented out is publicly disclosed everywhere. Investors know full well that the revenue is from the data center rental. No (non-retail) investor is going to see the jump in revenue and think “I better buy up because grok must be kicking ass!”
And yes, if hetzner built a massive AI hyper scale datacenter and rented it out for billions, with the expectation that they would keep building more, they would also see massive PE ratios because it’s expected that their revenue would be going up.
You’re falling for the Musk shell game. He just says whatever is convenient at any given moment. FSD. Funding secured. I talked to the Claude guys.
He didn’t change his mind about anything. It’s just that he wants to prop up SpaceX’s IPO as much as he can. Plus if he hits various targets, he unlocks more shares for himself, I think.
Very possible Elon is doing this to make give Anthropic better chances against OAI while he attempts to reshape xAI.
Also possible he sees infra as the future of xAI if he really believes in the value of space compute.
Hard to see this any of this as anything other than a bearish sign for Grok though.
I think its a mix of both of your first two points. He has hinted that he will be producing chips for Tesla and SpaceX here in the US in the future, so there will be a time where Elon will have every critical component we need for Tesla and SpaceX here in the US and not overseas, and fully in-house. Once he does this, nothing stops him from selling components. He's also insanely efficient at standing up data centers, so I would not be surprised if he gets enlisted by someone like Anthropic to setup infrastructure for him.
That's very doubtful given recent news that Colossus is running at 11% capacity and has hundreds of thousands of idle GPUs. xAI acquired too many GPUs and currently doesn't have enough customers to use them. That's why they are making compute deals with Anthropic and Cursor.
xAI is bleeding money and this compute deal with Anthropic will pay for all of xAI's capex ($25 billion) in 2 years.
11% MFU does not mean 89% of GPUs are idle, it means that they're using the GPUs ineffectively.
11% capacity is not 11% MFU. The first is about actually using the hardware for something in the first place, the latter is about how efficiently you compute. Different things.
The post you’re referring to doesn’t even include “89%” or “idle”. I’m not sure I understand the point you’re making then rebutting.
The original post of 11% is referring to leaked numbers about _MFU_, which has erroneously been re-reported as fraction of GPUs being used at all. The parent post is trying to correct this misconception.
I see. Thanks
This might actually be worse from a power costs standpoint.
> will pay for all of xAI's capex ($25 billion) in 2 years
This is a confusing framing; you pay down capex with profit not revenue, and there is presumably a high opex cost.
Nevertheless, you are still talking about 15B / year - when the initial investment was < 30B. That investment also probably covers things like power, salaries, etc.
So you are still covering all of that.
> That investment also probably covers things like power, salaries, etc.
That does not seem probable to me, other than for the initial build-out.
free cash flow pays down capex
> space compute
Heard anybody remotely competent about space talk the topic? It's pretty much a literal laugh every time.
Some of the proposals are laughable, some of the example calculations are. The idea of running AI training sounds extremely challenging. But the idea of inference in space doesn't seem absurd
The power budget of a starlink v2 mini satellite is estimated at around 20kW based on the known solar panel size. That also matches what a satellite of that size would roughly dissipate without dedicated radiators, using just some heat pipes to spread the heat evenly over the satellite's surface. There is nothing fundamentally preventing you from taking the same satellite design, remove most of the comms payload and instead put 15kW of GPUs there. Or about 10 GB200 including the CPUs, networking, etc that you need along with the GPUs
Now, do the economics work out for $300k worth of compute for each satellite, in an environment where maintenance is impossible and degradation will be higher than on the ground? Probably not right now, but in a couple years they might
The starlink satellites have hall effect thrusters, and presumably beefy laser and radio comms systems that will account for a lot of the energy budget. Also, they are sometimes in the earth's shadow, a naive calculation says they have 2x the solar budget they need, in order to charge up batteries for shade-time operations. There is no where near 20kw of compute in a single satellite, and thus no where near 20kw of heat to get rid of.
Furthermore, xAI's colossus supercomputer is specced at 250MW. And this seems to be a number that'll just increase over the coming years with new bigger DCs.
To match this level of performance they will have to launch what, ~15k satellites _per_ equivalent datacentre?
Regarding cooling: you can't just cover the outer surface with pipes. You cant't dissipate the heat, you need to _radiate_ it away. You need to point that surface to the deep dark cold of space. If you point it to the sun, you will heat your satellite. Think a massive "reverse solar panel" that works with infrared. You need surface area, and loads of it.
I'm not saying this is impossible. Obviously elon will prove us all wrong because he's stubborn like that. But there is no way this will ever be economically viable when competing with terrestrial based systems.
The assumption behind "spread the heat over the size of the satellite" is radiative cooling. Admittedly my numbers were a bit off, you need to make the satellite a bit bigger, or use some of your solar panels. A starlink v2 mini has 8m² of area per side, so 16m² total. To dissipate 20kW to space, at a surface temperature of 80°C and emissivity of 0.85, you need about 28m² of space. So you need to increase size a bit, or add 12m² of dedicated radiators. A bit more to deal with the real live complications (the sun exists, the earth is actually warmer than space and covers a significant portion of the sky if you are in leo, etc.).
The actual Starlink V2 Mini is has estimates for solar generation that range up to 35kW. We need a bit more, but not much more.
My numbers are a bit optimistic, but they are in the right order of magnitude. Power and cooling are very achievable in the area of putting one server on one satellite. It's the "putting one DC on one mega satellite" ideas that run into feasability issues, and for inference those aren't needed. The economics are the bigger issue, and launch costs are only moving in one direction
The terrestrial based systems getting protested, blocked, and demonized everywhere? Those ones?
How is this not absurd? What is the benefit? Space is a harsh environment, with issues due to solar radiation etc, etc. And it's permanently 100ms away from any user.
Tl;dr — but it’s too heavy
You’re making some decent points here, but you’re either forgetting or ignoring the major thing that people usually neglect to mention when they want to make a case for this (crazy) idea — weight. Unless SpaceX is going to completely redesign hardware such that it is optimized entirely for its mass, it requires many (many) launches to even get a small set of racks into space. I don’t normally get up in arms about the CO2 emissions of data centers, I think there is offsetting value created by their use, but I would absolutely protest trying to put data centers in space and do my best to shut down the hundreds, if not thousands of launches it would take to achieve even a tiny fraction of an AI data center.
Have you calculated the CO2 output of Terran datacenters run on natural gas vs. space datacenters run on solar? The launch CO2 usage is one time, the datacenter energy is for the life of the equipment.
While I agree with you on the merit of the idea, rockets that can take off and then land vertically without damage were also laughable pre-SpaceX.
NASA had literally done it. It was never laughable, just thought to be an incredibly difficult engineering challenge.
I think compute in space suffers less from being "impossible" and more from being "impractical". It is plenty easy to put compute in space. It is just still silly expensive and by the time your equipment makes up the cost of putting it in space, it will be well out of date.
The non-deterministic nature of LLMs will get a whole lot more fun when we add randomly flipping bits to the situation.
It is absolutely baffling to me how frequently I hear people talk about this around me. There is no way this is happening anytime in the near future.
But if the US isn’t the first to put an AI on the moon, we’ll lose the AI space race!
Mr. President, we must not allow a mine-shaft gap!
To this day still my favorite line in the movie. It really just punctuates the whole thing
You don't have to actually have an orbiting datacenter for the idea to work. You just have to convince enough people. Once you've done that, you can claim that regional regulations don't apply to your data because the data is in orbit. Its not like somebody is gonna go up there and catch you in the lie.
Out-of-regulatory-reach is what they'll actually be selling. It can be on earth, it just has to be sufficiently hidden such that you can claim that it's in space.
I don't think that's the purpose and if it is, it's obviously trivially solvable. The US can and does simply assert jurisdiction all over the globe. There's no reason it can't or wouldn't extend that to orbit.
That conversation goes like this:
US: Stop doing that thing with your data center
X: It's in space, you can't tell us what to do
US: Yes we can
X: <Say OK> or <Go to jail>
The US already claims jurisdiction on all launches from US and NZ (Rocket Lab). Every one of these had to get approval from the US to launch.
Space compute isn't real though, it's just a scheme to pump the value of SpaceX before IPO by associating it with AI. It's really hard to cool things in space, because there's no matter to transfer the heat away. All you've got is radiative cooling, and that's really really slow.
xAI is a dead company; you don't sell compute if you're growing.
More promising is that cursor is training a model using it.
Right, also Anthropic has been having difficult time getting more GPUs
That would be absolutly ludicris.
His own product is competing against anthropic.
It's a money maker. Maybe not a number one spot, but they don't have to stand up any more datacenters.
His product should be the money maker.
He believes in the value of the idea of “space compute” for attracting investors to SpaceX. But the existence of the idea of “space compute” as a better way to deploy datacenters (along with everything else Musk has claimed in the past decade) should give everyone pause as to the plausibility of literally everything else he says.
> along with everything else Musk has claimed in the past decade
* mass produced EVs * Neural links to brain * reusable rockets * more efficient tunnel digging
Dude has a pretty good record of taking stuff in R&D and making pretty real products and/or companies. Can you name someone better?
> He believes in [...]
...the value of having others buy the idea of space computing. I don't think he himself believes in what he says.
> should give everyone pause as to the plausibility of literally everything else he says.
That ship has sailed, we're in the Age of Cults. If you're a believer you're probably also invested in his companies, so your mind is doubly-clouded.
> so your mind is doubly-clouded
Further I think external agents are controlling their mind.
We could see the first company vertically integrated from etching to chip to data center
That's Intel. Probably IBM too, though they've been doing mass manufacturing with TSMC (and GloFo before) for years instead of their fabs. I wouldn't be surprised if HP did similar things back in the 80s.
We wont for anytime soon.
Intel and TSMC are not what they are today just because they buy a very expensive EUV Machine from ASML but because they have the knowledge and infrastructure to even use these machines.
What do you mean by etching? Google does also it's own chip design with TPUs, data centers, and models but afaik only TSMC Intel and Samsung do the actual semiconductor fabrication
Assume he's referring to SpaceX's https://en.wikipedia.org/wiki/Terafab
> Analysts estimate the costs for the full-scale facility at between US$5−13 trillion.
https://en.wikipedia.org/wiki/The_Line,_Saudi_Arabia
Musk and MBS should hold hands.
Turning The Line into one giant ~~paperclip~~ chip assembly line!
We can call it the Peta(Principle)Line!
IBM was doing this since the 50s.
Here's a probably stupid question - if someone were unbounded by ethics and conceivably had enough power and connections to power to shield themselves from many consequences of their actions - and that person owned these DCs, could they in theory observe all the streams of tokens coming in and out of these models, and even exfiltrate copies of these models wholesale to have their own teams do what they will with them in the pursuit of building their own competitive models?
Or is there something fundamental in the way these models get deployed (encryption or something or than legal contracts?) at this scale that prohibits the owners of the infra from gaining this level of insight / access?
1) The situation you described would be covered under the contract between Anthropic and xAI, and that any violation of that would be subject to financial penalties and legal proceedings. The US has a robust corporate legal system, and disputes do get resolved through the court system, although in a slow and costly manner.
The contract can stipulate a penalty at a high enough amount to discourage this behavior.
2) Output from models & intra-datacenter communications can be encrypted if customers truly cared.
3) There is no reason do this, because there are far better ways to exfiltrate data from Anthropic models. Chinese companies are already doing this at an industrial scale where they are reselling Claude tokens for 10-20% of the cost while retaining the data to train their own models. https://www.chinatalk.media/p/how-to-buy-cheap-claude-tokens...
If we look at Deepseek V4-pro, created by Deepseek who Anthropic formally accused of harvesting Claude tokens at scale, it performs the same as Claude did 6 months prior.
> The US has a robust corporate legal system
Thanks for the chuckle. ;)
It's an accurate statement. The US (specifically Delaware) has a world-class corporate law system. Delaware has a court dedicated to only corporate and commercial disputes, with 200+ years of case law.
It does, though.
It does until you're embedded enough with the surveillance system [1]. If a company was able to get immunity in the wat AT&T got it, no contract would protect the other side.
[1]: https://en.wikipedia.org/wiki/Hepting_v._AT%26T
That case has nothing to do with contract law.
> There is no reason do this, because there are far better ways to exfiltrate data from Anthropic models. Chinese companies are already doing this at an industrial scale where they are reselling Claude tokens for 10-20% of the cost while retaining the data to train their own models.
I think you missed the part where Anthropic stopped displaying their thinking tokens over the past few months, and instead now provides “summarized thinking”, letting Haiku summarize Opus’ thoughts.
So it is now much more difficult (impossible?) to distill the models.
I also think you over-estimate how well the legal systems works in the US nowadays, and under-estimate just how much power Elon has in the government.
At this scale thinking tokens don't matter anymore.
In Feb Anthropic called out three Chinese labs for "distillation attacks", but a lab missing in their post actually had most Claude generated tokens among all Chinese labs in their midtrain data :p
Incidentally, Claude started the "Summarized Thinking" bullshit around a year ago.
Deepseek kept its pace of improvement nonetheless.
I guess there's a reason why those Chinese companies are in china
I hope the Chinese keep harvesting Claude etc. since they stole all their data anyway, who cares?
It's already in the public domain (thanks to the OpenAI trial) that Grok distilled OpenAIs models. Listening to the data going into the models in the data centre would be very similar thing. There's some downsides (you're passively listening, not controlling the queries), and some upsides (way more data). But it only ever gets you to some percentage of the existing production model. It doesn't get you what Musk wants - an AI company capable of designing and deploying leading edge models. It gets you to fast follower status.
You could, but Grok is pretty high up there, it might not be "#1" but its definitely up there with the giants, people seem to overlook it. Gemini has a similar problem, it was #1 once, and it seems like Google isn't hell bent on chasing #1 they just want to keep iterating over time, they know they just need it to be "good enough" and they'll keep having repeat customers.
If Elon REALLY wanted to do anything like that he would be better off poaching talent from competitors, less legal hell to go through.
No Grok is not 'up there'. Not by any stretch of the imagination is it anywhere close to anthropic or openai in any single domain whatsoever. Not even close to deepseek. It really isnt a good model. Their research team's talent is just not good, unfortunately.
If I was training a top-tier model, having a competitors’ weights feels like it would be an excellent refining tool. As a user I find cross-model iteration to be a huge power-tool, and presumably the boffins can zero in on areas of relative strength and weakness and work out which area they want massive amounts of synthetic data from.
Bear in mind these data centers were built for X.ai to use themselves, so there would have been no reason to backdoor them. Also, this is all off-the shelf equipment: Dell & Supermicro servers with NVidia compute modules and ethernet switches.
xAI had a lot of negotiating power here because Anthropic had ~0 comparable options and ultimately desperately needed the compute now. So, it wouldn't surprise me if data sharing was an explicit part of the agreement
What prevents a data center operator from reading your chats? [FEATURE] Provide a way to select your data center #56916 https://github.com/anthropics/claude-code/issues/56916
Yes he could do that and I thought about this too. Very weird tbh
There's accusations that the Chinese labs have done essentially that to OpenAI and Anthropic and exfiltrated their models without having DC access, so if you had DC access, yes, you could do that. If you had DC access though you could just copy the model onto an SSD.
The Chinese labs have been accused of training on elicited chat logs on a massive scale in violation of ToS. That's possibly a real concern, but it's nowhere close to "exfiltrating" the model or even roughly matching its behavior.
This isn't pointing towards a merger is it?
xAI gets the cashflow and makes spaceX bottom line more appealing. But that guy Musk hardly makes deals that favor other companies over his, so what am i missing?
xAI is bleeding tons of cash, and they don’t have the ability to use the compute for robotaxis, which have been a disaster. Grok monetization is also low.
I think that's the datacenter with the gas turbine generators that operate without permits because they're "portable." Data centers have tremendous externalities but colossus is a particularly nasty offender, and not just due its size.
Edit: They did it with Colossus and now they're doing the exact same thing with Colossus2. https://www.selc.org/news/xai-built-an-illegal-power-plant-t...
The newer location is about 3 miles southwest of the Memphis Airport (MEM), one of the world’s largest cargo airports and the center of FedEx operations (500 take-off and landings per day most concentrated in a 6 h FedEx window with lots of engines running on ramps and that produces about 2000 tons of NOx per year).
I live about 18 miles downwind of the new Colossus sites, the airport, and lots of truck logistics sites, and a large refinery.
I definitely will be getting 2x exposure to ozone and particulates from both Colossi when they are running full bore. Plus an extra dose of ultrafine particulate with my morning fresh air.
Yes, wouldn’t it be nice to be in Nashville instead with HCA, Oracle, many insurance and financial institutions, and the joy of country music.
As an avid Opus user I am in an ethical Nimby bind. We do need almost any investments we can get in Shelby County TN. I’ll take Anthropic in preference to Grok NOx. And it will be my NOx.
The biggest issue with the interim onsite generation is the lack of meaningful stack height on the generating units.
Airplanes by virtue of their mode of operation stay out of the unhappy regime most of the time. Also, engines at/near idle produde orders of magnitude less emissions. Those aeroderivative generators are running at full capacity 24/7.
Dumping exhaust at ground level continuously is probably much worse than the airport. Even if it's a FedEx world hub.
Cattle can't choose how to be slaughtered, though.
> I’ll take Anthropic in preference to Grok NOx
It's the same datacenter? Ran by the same people?
I think that was the joke.
Hey Siri - can you please order this guy a spare set of lungs ?
Or upgrade their Tesla to that bio hazard filtration level
Your math is way wrong. The airport is and will be far, far worse. You didn't even mention all of the lead
there is no lead in jet fuel.
i would expect close to no 100LL burning planes use MEM.
What was the specific mathematical or factual error? It is not theoretical for me.
So for a guy that literally has a company that produces batteries and solar panels, choose to use gas turbines. So much for saving the planet.
Solar and batteries are a bad choice for a constant 24/7 load.
That's the exact reason we will never go fully solar (or wind) unless an insanely impressive battery breakthrough makes storage effectively free while using only common, renewable components rather than rare earths.
Solar, wind, etc are excellent parts of an energy system, but its nearly impossible to cover base load at scale with generation that may only run for 0-5 hours a day.
edit: typo
You may want to update your priors [1][2][3]
1. https://www.carbonbrief.org/guest-post-how-solar-panels-and-...
2. https://www.fastcompany.com/91500104/google-minnesota-data-c...
3. https://www.rootsanalysis.com/sodium-ion-battery-market
Sodium is an interesting chemistry, though it has different voltage curves than lithium ion once hardware is built to match it they may scale well for industrial use.
That still doesn't avoid the more fundamental math of having to store such massive amounts of energy though, even if you skip batteries and pump water to retention ponds uphill.
Even for small scale residential, the recommendation is to store 3 days worth of average usage to handle stretches of cloudy days and to have a generator for the times when that still doesn't cut it. You also need enough solar that 4-6 hours of generation can fill those batteries back up after a stretch of cloudy days.
You also have to contend with frequency issues. This is what took down Spain's grid, they had turned on a ton of solar at the time - with many gas plants offline a seemingly small dip below 60Hz really wrecks much of the system when it wasn't designed to handle those swings and triggers multiple safety mechanisms.
He is building the skill tree in such a way that he is prioritising speed rather than environment.
> So for a guy that literally has a company that produces batteries and solar panels, choose to use gas turbines
“Sorry, we have to shut down your database because it’s been overcast for the last two days”
Now, they could have a mix, of course, but just running on solar and batteries at that power is not realistic.
I dug into this topic in some detail on my blog and it's both enraging and depressing.
https://poiesic.com/posts/pattern-recognition
This is a really great watch from Benn Jordan on the topic https://www.youtube.com/watch?v=_bP80DEAbuo
Some wild things happening with those, and infrasound. Colossus is shown 4 mins in
This was debunked: https://blog.andymasley.com/p/contra-benn-jordan-data-center...
The debunk is extremely sketchy on many points.
https://www.bearlythinking.com/p/andy-masley-doesnt-understa...
That post is mentioned at the bottom of what I linked.
He has a full post response here: https://blog.andymasley.com/p/to-be-clear-i-do-understand-ho...
How i know Masley is trash:
> Jordan is suspiciously lurching to the extremely high energy end of the light spectrum when we know that the low energy end (comparable to infrasound) doesn’t have negative impacts on us if we can’t detect its presence.
Masley spouts the falsifiable propaganda that any photon (light/emf) below ionizing radiation energy can only cause “heat” and can have no other possible harm effects.
Many science-minded people (though more accurately in this case, billiard-world materialists) have become quite militant defenders of this idea (ostensibly fighting off the hoardes of tinfoil hatters and quantum aquarians sensitive to 5g).
This point is very plausible military industrial propaganda. There were numerous studies with evidence (starting from the 60s) such as - non ionizing radiation (eg hv powerlines) might cause lots of cancer, and it’s weaponized (sub-thermal) usage can microwave the brains of enemy spies. THOSE studies have come out in declassified and leaked docs.
Now we have several plausible and serious theories of mechanism for low-f light disrupting biology that lean into quantum biology. While the iceberg of quantum biology understanding is still in its early decades, the mounting downstream evidence of health and medical issues are established public knowledge.
Interesting, I read this series of posts and as someone who does not have a dog in this fight but does have a more than passing background in both audio engineering and datacenter engineering, the response Masley gives here to the very first criticism is fundamentally incorrect. I haven't read the rest of this, but his claim about sound intensity what it would imply about energy is on its face untrue.
When you take a measurement of a sound, you are measuring both its pressure and its intensity, that is what is implied by a measurement in decibels. The measurement is taken from the point of the measurement device/listener in relation to the source/generator. If the measured value is potentially harmful, there is no such implication about needing to redirect additional energy to make it harmful, it's already been measured as potentially harmful at the point of measurement.
It's basically nonsense. My most charitable interpretation of his very first responsive argument is he's saying that a datacenter would need to intentionally direct energy towards increasing the intensity of its sound output to make Jordan's original measurements meaningful. That's neither how measurement works, nor how sound works, nor even how datacenters work. Things like sound and heat are BYPRODUCTS and not the point of the datacenter, both have an intensity, and that intensity is measurable, and any energy which is expended towards that intensity is energy that was wasted away from doing computations.
I stopped reading after this. I don't know if Masley is out of his element or just practicing motivated reasoning and thinks his readers are stupid. Either way, his rebuttal already failed on the first point.
I dont know, Id say the enraging thing is that the government is so incompetent and unable to expand electricity supply that datacenters are forced into using loopholes to get power the only way they can.
> Id say the enraging thing is that the government is so incompetent and unable to expand electricity supply
So let's say you're a homebuilder, if I tell you I want a new home and I want to live there tomorrow, you can all of a sudden build it in a day, right?
Electricity use is skyrocketing for various reasons, these datacenters being one of them. There are a lot of countries struggling to keep up with demand. So incompetence? No, probably more like supply lagging demand.
Or ASML and Nvidia and all also are incompetent, because they didn't see demand coming....
ASML and Nvidia did see the demand coming. Intel didnt, theyre incompetent.
> There are a lot of countries struggling to keep up with demand. So incompetence?
Mostly, yea. If the government gets out of the way there are plenty of people who would be happy to supply power in exchange for money. The homebuilding analogy is good because homebuilding in the US often has the same issues that plague the electricity market. There's no reason it should take years to build homes in many parts of this country, but it does.
In the early days of this the AI companies were asking for massive new energy supplies but also refusing to sign contracts to pay for it over the decades of its life.
They're basically attempting to game the system, politically and economically, to put as much of the cost on taxpayers and ratepayers as they can. This naturally slows things down.
If only there was some sort of planning, by a central authority!
That isn't a problem of inept government, its a problem of over regulation and what amount to state-sponsored monopolies in many areas.
We don't need the government to fix it by expanding power grids from the top down, we need free markets allowing competition.
I'd be curious to hear from the silent down votes. Do you disagree that power companies are effectively regional monopolies? Or do you disagree that government oversight and regulation isn't the reason for poor maintenance and lack of capacity?
These are the same companies and individuals who are actively working to destroy functional government, and are happily looting the US treasury rather than let it be spent on things such as encouraging more energy production.
Seems like selc's time would be better spent trying to close the loophole that allows for unpermitted turbine generators instead of going after one company for doing what they were allowed to do when they did it.
And Anthropic can pay money to Musk to absorb all that liability while they keep training models.
Yup, and now Anthropic is complicit in the environmental damage and health problems for local residents that these data centers are causing.
But hey, number must go up, right?
Have you considered that the march of progress requires human blood to grease the gears and mulched skulls to pave the (highly efficient) road? Really, when you take into account all of the future lives this will improve and save it's difficult to claim any cost now is too high. Would you stand in their way and delay the day that Mythos cures cancer?
This is a joke. Read it in a mocking tone.
https://www.newyorker.com/cartoon/a16995
I wonder what percentage of GDP expenditure will give us SkyNet.
Undoubtedly, it will find cures to all cancers… The ARR and stock appreciation will be amazing. Except the cures will be found long after it has wiped out all humans.
Wiping out all humans _is_ the cure.
Not Anthropic, but Sam Altman - AI will solve climate change and cure all diseases.
My favorite claims are that it will solve both aging and death, whatever that means.
AI is the new religion, and one needs to be stupid to believe it.
Already so many people are treating it as a higher being, believing whatever that comes out of it
That implies you don't need to be stupid to believe the other ones.
I'm not saying human blood and mulched skulls are a renewable source of power, I'm just saying. Or maybe they can partner with SoulCycle to power computation with 24/7 spin classes?
And people called the matrix’s human batteries far-fetched.
Always felt like it would've made more sense if it was using part of the peoples' brains to do their computation, as super energy efficient computers.
I believe that was in the original script, and rewritten after some exec didn't understand how brains could be computers.
I'm going to bet it's simpler than that. I'm betting they changed the script for the Duracell product placement that made an acquaintance of mine a ton of money for pulling it off. Always follow the money.
https://alistentertainment.com/marsha-r-levine/
If it was indeed the original script, the reason they changed to batteries is maybe not because "some exec" is an idiot, but because it worked better from a storytelling perspective.
Even if treating people as batteries doesn't make much sense as we are pretty terrible power plants, the message is clear and impactful. It is common for movies to oversimplify things, because they want to avoid having the viewer from being distracted from the main plot. It is tricky, as being too obviously wrong can breaks the immersion. I think the people = batteries analogy is a good compromise. Brains = computers, while technically more plausible would add a layer of complexity that could be a bit too much for a 2h action movie.
I don't think it'd have much of an effect on the story, outside of background stuff like the Animatrix, it's just an interesting little fact about why the world is the way it is. Shooting and hitting things in slo mo is still the core either way.
Matrix franchise is well known for not trying to pack in too much complexity /s
Per sanguinem ad astra
I loled
When do we start building pyramids and doing the Sardaukar blood letting ritual?
When the 10-year yield hits 6%. Basic macroeconomics.
Some of you may die, but it’s a sacrifice that I am willing to make.
A superintelligent AI will be safe though, because it learnt its morality from us.
Doesnt AI learning its morality from humans makes it unsafe, I mean just look at some cases, humans dont exactly always have good morals
Yeah, look at the parent comment I was sarcastically responding to.
Of course top of all that, even if human morals were perfect, it's still a dubious claim.
Do you think Boris cares about people getting cancer and dying from these data centers? No, he cares about becoming rich as fuck.
Isn't he rich already by now?
Sam Altman and Dario Amodei are already billionaires. That doesn't stop those greedy fucks from stealing everything in their path in the pursuit of money.
Greed is a societal illness.
This is simply called disruption. They really don't care.
Given how much our EPA has been gutted by the current administration, I don’t think relief is very likely.
So much for Dario’s ethics. He happily partners with Elon. He seems like just another power hungry monopoly seeking liar.
That’s because he is. None of these people are your friends and all of them will fuck you over if it means getting richer and more powerful.
> So much for Dario’s ethics. He happily partners with Elon. He seems like just another power hungry monopoly seeking liar.
Dario has been glorified unnecessarily. He's just like all the other people in the space: not good, not bad.
And keep in mind that when Dario was opposing AI usage by the US State he wasn't really opposing, he was just saying "not yet".
this is why ai in space will win
The physics of data centers in space will be extremely difficult and expensive to pull off in any meaningful timeline.
I fully expect "space AI" to be about as realistic as the flying cars and hover boards we've been promised since televisions were black and white.
xAI’s turbines produce meaningful local/regional pollution (especially NOx in a vulnerable area) but represent a rounding error nationally and globally.
> xAI’s turbines produce meaningful local/regional pollution (especially NOx in a vulnerable area) but represent a rounding error nationally and globally.
If you shoot someone in the face, it will produce a meaningful increase in local/regional murder, but represent a rounding error nationally and globally.
No one should drive any kind of vehicle because they cause accidents.
> No one should drive any kind of vehicle because they cause accidents.
Come on, stop with the desperate false equivalencies.
Gave me a chuckle on my commute, thanks!
It doesn’t matter if people have to suddenly live by gas turbines that run 24/7 because why again? Can you repeat that last part back to me but say it a little dumber for me?
farting in a crowded elevator because the people outside the elevator won't even notice
multiply rounding errors by thousands and you get somewhat meaningful impact. you are underestimating the scale of independently small pollutions
What's the tremendous externality of gas generators? People heat their own homes with natural gas and it's no big deal. How can a datacenter that is miles away be worse than that?
Its totally inefficient - burning the same gas in a co-generation plant, ideally combined with district heating, would produce the same amount of pollution and basically make use of all the energy.
How can something that is a million times bigger than your home be any different, right?
The gas furnace in my basement don't have a massive jet turbine emiting high frequency noise
It's the low frequencies that's more of an issue apparently. Benn Jordan has a great series of videos about it, including one on Colossus
I wouldn't call noise pollution a "tremendous externality". The gas turbines should just be placed far enough from where people live.
Should, but in fact are too close.
Why is xAI giving up their advantage? Is this a signal that their frontier model improvements are plateauing and decided there is no value in hoarding all their compute?
I would guess it's purely because Grok isn't nearly in-demand enough to produce meaningful revenue. And they want to juice the numbers for IPO
And I'm sure it's a bonus point for Musk that it goes to OpenAI's most relevant competitor
As weird as it seems I think this is Musk’s best shot at winning over Altman. He has personal vendetta.
xAI have (with some questionable ethics) managed to actually build a data centre, so they have a ton of compute but not much inference demand for their model which is second-tier.
Everyone else trying to build data centres is really struggling (turns out building physical things is not as easy as writing code, who knew).
So Anthropic have the model but they're compute starved because everyone else they've signed agreements with still mostly have piles of dirt and the GPUs are still in Nvidia's warehouse somewhere.
It's a bit of a win-win: xAI's financial numbers will be massively improved by the revenue using otherwise useless data centre capacity, Anthropic get the compute they desperately need albeit probably paying through the nose for it
Musk's goal with X.ai from the start seems to have been primarily to compete with OpenAI, and given that Grok itself isn't doing that (not even mentioned in competitive benchmark tables), and his data centers are therefore mostly sitting idle, this deal with Anthropic helps him achieve that goal by helping OpenAI's #1 competitor.
Note how the initial deal was only for the older Colossus #1 data center, and now after losing his lawsuit against OpenAI the Anthropic deal has been expanded to Colossus #2 also. Coincidence?
Elon got rid of most of the core xAI researchers recently. It looked like firings and he said something like Grok needs to be rewritten because it was fundamentally built in a bad way. My guess - Grok was not one of the top two models and had no traction beyond users on Twitter - so investing in it did not make sense. Or maybe he needs time to build a new team and rebuild Grok, and while he’s doing that, he needs revenue to make up for all the GPUs they bought.
Either way, I doubt he has any ‘real’ advantage in this game. It’s OpenAI and Anthropic, and then everyone else including the open weight models.
Boy I’d love to have the sort of ‘no real advantage’ that got me $1.9bn in revenue per month from Anthropic.
Make no mistake - Elon’s charging premium pricing because he has something Anthropic realllly needs. In exchange, he will get significant revenues added to the balance sheet along with the cursor team right around IPO. It’s very good business for Elon and SpaceX.
What advantage? Has there ever been any indication they’re leading in any segment? Sure Elon has thrown a bunch of money at hardware, but to what end?
And frankly as bad as Altman is from a: if AI is really going to disrupt humanity do I want this guy in charge? Elon is 10x worse. So why would the best and the brightest ever work for him?
No other LLM has made as much child porn as grok so there is that...
In a compute starved world, big ass data centers are an advantage.
yes, but it's only advantage if one is compute-strained and the other isn't. if they both have lots then there's no advantage. if one doesn't fully utilise their compute then it's not an advantage either
Well, it appears all their competitors are compute starved so…
can you elucidate what that advantage is, that isnt renting it out for the highest price to somebody that really needs it?
Well, other than your ability to turn that into cash by renting it out for the highest price to someone who needs it, you can promise prospective employees that are supposed to use that infra to train models that they won’t be compute starved.
You can kick off more model training runs and experiments than your competitors.
You can kick off a $1-2t IPO claiming you are going to capture a large portion of the largest TAM the world has ever seen.
Could have (past tense). They now have rented the capacity out, presumably with a multi-year contract.
A rental deal, even at a premium price, actually makes great sense for Anthropic. Rather than building themselves, they are de-risked by renting instead, and they don't own the depreciating asset. They've also compute-starved X.ai as a potential competitor as a marginal added bonus.
They have neither the most resources nor the best models. They are mediocre at everything except the CSAM generation market, they've got that one cornered
Training a model that is larger than your competitor's.
The datacenter advantage, obviously
I think they overestimated the demand for Grok, which is mostly useless, and now they have too much compute on hand.
I see this as a huge warning sign. If a frontier ai lab is in this position of renting their own capacity, imagine how much overcapacity there is in the system.
Outside of VC money, and circular financing, the only external money coming into ai are into open-ai, and anthropic via their subscriptions and APIs.
To the contrary - there’s so much demand that Dario needs Elon’s compute and is willing to pay premium month to month pricing for it.
I think any warning this provides is more about the ability to forecast demand. Obviously it was a giant leap of faith, that didn't pan out, for Must to build these data centers without having a product with a growth trajectory to more rationally justify it.
OpenAI and Anthropic (and their investors!) are obviously also out on a limb here, both with deals in place to 10x their available compute (roughly 1GW -> 10GW) over next few years. Maybe these growth estimates are a bit more grounded, but they have all sorts of assumptions baked in. I still would not be surprised to see OpenAI self-implode, and to see Google and Microsoft as the eventual winners here.
For sure. There’s also this kind of interesting meta strategy question: if you think you can float the huge capex with some internally valuable processes as a base case then you might just go really big. Dario and Elon’s temperaments come in to play here
> a frontier ai lab
Wait, do people consider xAI a frontier AI lab?
They were briefly SOTA on some benchmarks, although there were suggestions there might have been some massaging of the results since real world usage showed lackluster performance compared to the benchmarks.
When did xAI have an advantage?
Advantage in training compute.
xAI advantage is that it's not censored, that's why people use Grok
More signs that xAI might be giving up on the AGI race. xAI let Cursor train a model on Colossus 2, gave the entire Colossus1 to Anthropic, and is now giving compute in Colossus2 to Anthropic as well.
Bad read on the situation. xAI has too much compute and not enough customers using it. They have around half a million GPUs, some of which are stolen from Tesla, running at 11% utilization. xAI predicted more people would be using Grok, but Grok is not a SOTA model & users primarily want to use SOTA models. They have excess capacity and it makes sense to rent out GPUs to other customers while they improve their models.
Grok is also tuned to align with Musk's personal beliefs. I wouldn't touch it with a 10 foot pole.
I keep hearing this statement, and it always makes me wonder if people have actually used Grok…
I have a Claude Max plan I use for coding, but I also have a Grok Lite plan I use for web search type tasks (similar to Perplexity) because I like how the Grok harness handles searches and I don’t need a SOTA model for that use case. I’d never pay $30/mo for a full SuperGrok account but to me it’s worth the $10/mo for Lite as I was hitting limits on the free tier.
I’ve never noticed it to be particularly biased at least for anything I’ve been searching for on it. And on the other side, I’ve never noticed it to be particularly less censored or anything compared to other models either (also a claim I’ve heard a lot about Grok but I think because it is/was part of their marketing).
Opposed to all other models being the bastion of objectivity? Must be truly vindicating to have to hear other peoles opinions after decades in the silicon valley bubble.
There is a difference between when somebody openly instructs their model to infer disproven lies vs who doesn’t do this. And it’s quite tiring that this is even a question because of politics.
As somebody from Hungary: the biggest impact of my mood was that this kind of thinking went back with the collapse of far right there to where it belongs: to a deep hole which is not in front of normal people. Average people suddenly don’t ask illogical questions or answer stupid things because there is nobody who would tell them that they need to think stupidly, there is nobody who tell them what stupid thing they should think that week. It’s marvelous when you get the proof that the whole “stupid thinking” is completely controlled from above.
Qwen has post training to tell you incorrect answers about Taiwan. Seems worse to me
In end user perspective, it’s the same. The difference is probably volume. I have no clue in which direction. Both in trained lies by models and in number of people defending the indefensible. But one for sure, there are probably 10s or 100s of millions of Chinese who try to do the same. I encountered with it quite frequently. Sometimes with flat out doublespeak.
Nobody ever said other models were bastions of objectivity. They only implied they weren't corrupted by Musk. Which is true, and which is good.
As a non-US AI user I do not particularly like using a US model following the recent political events, but I specifically do not want to use a model made by an ex-member of the current administration.
It is always great fun using a model via API without search/web access and quoting a recent development, being told that it must be hamfisted satire, then providing access. The reasoning traces are a delight, Opus 4.6 and GPT-5.4 during the administrations war with Anthropic were prime grade A kobe beef.
This comment is very similar to what russian propaganda does.
It's not aimed at convincing you to support them, but to convince you everyone is lying and there is no meaningful difference between each position, so you stay apathetic.
Yeah yeah every opinion you don't share or like is a russian bot or literally hitler.
I see what you did there. Nice.
Why are they selling compute instead of using it to build that SOTA model?
They tried and failed. xAi made a mistake building Colossus 1 and ended up with heterogenous cluster of H100/H200/GB200 GPUs. This is a nightmare to train huge models on because each card has different specs, features, and hardware requirements. During gradient synchronization, a heterogeneous cluster would bottleneck on the slowest GPU (H100) so the faster GPUs would end up idling. They also probably ran into unexpected compatibility issues, which are difficult to resolve.
It makes more sense to use this cluster for inference, since they can segment the cluster by GPU type and avoid GPU mixing. xAI doesn't have enough inference customers so it makes sense to monetize this to companies that need inference compute such as Anthropic or Cursor.
Apparently xAI will try building SOTA models on Colossus 2, which will be built on Blackwell GPUs only.
How can something so obvious be overlooked by team building the data centre? Can't the sharding be uneven so that weaker GPUs still finish fast by taking on a smaller workload?
It's not like they had much of an option, when everybody was hoarding every GPU they could. For the second Colossus they could book future production, but the first one had to be built ASAP so xAI looked as a serious competitor in the AI space.
I imagine it involved a petulant billionaire screaming "Fucking build it. Build it NOW!" in response to expert feedback.
It is a race that has a flywheel effect.
Once xAI training team “fix” their model, where will Anthropic be then?
More people should try Grok. I don't use it for coding but it's replaced a lot of my ChatGPT usage. Definitely more perferred model for quick questions or easy answers.
One thing I do like about Grok is that it makes it stupid easy to see what its referencing, and gives you the links to those resources. Which most models sometimes either don't bother, or don't do much of a good job of doing. It's not the top model, but it is definitely high up there, people's blind rage for anything Elon Musk is the only reason most people don't realize how capable it is unfortunately. Grok is not exclusively made by Elon Musk, there's definitely other engineers working day and night on it.
For conversational or general knowledge questions I also much prefer Grok. Musk's vanity aside, it is much less censored than the other frontier models.
Where less censored means "censors facts and left leaning claims while actively promoting far-right lies".
Far more mundane that whatever bogeyman you imagine, it will discuss things like COVID origins or not treat favored classes as a special case e.g. when asked to make a joke.
What's the blind rage, he's totally out in the open.
It's not stolen if it was taken from Tesla, investors already agreed that Elon can do anything he pleases with their money.
Elon lost his lawsuit with openAI and knows xAI isn't on the same trajectory. Might as well try to win the bet and flip off Sam by supporting the best competition. Also they are getting a head start on AI as a commodity. I'm sure there's plenty of money to be made for those that can leverage their capital to essentially rent capacity right now. If he's not making enough off of grok, might as well cover their expenses.
It was kinda obvious when SpaceX "acquired" it. Elon rewarded xAI investors/prevented lawsuits by giving them SpaceX equity, and that was that.
FWIW, SpaceX (parent company of xAI) has an option to acquire Cursor for $60B that expires 7 days after their imminent IPO.
Do people still use Cursor? My company’s leadership has been clear that Cursor was cool for a hot minute but you Should Not be using it anymore
It’s a very fractured and heterogeneous landscape where your own perspective will be warped by your personal experience.
Anthropic has a lot of the market share and dominates the mind share, but each of Codex, Devin, Cursor, Claude, et al have significantly more market usage than they had 6 months ago and each are likely still growing very quickly based on publicly-reported information.
xAI might acquire Cursor. They are in the process of training new coding models and probably a new Grok.
Until they finish training, it makes sense to rent the excess capacity.
I don't see a scenario where it really makes sense to be a frontier lab long term. Eventually model quality will plateau then you distill and get 90% for 10% or less cost.
General AI is that scenario. The investor dream is that their horse hits general AI first, patents it (or otherwise somehow stops the competition from hitting it), and then reaps the massive benefits.
I'm not saying it's a likely scenario, but I genuinely believe a big percentage of AI investment revolves around that (or similar) scenarios.
Opus and GPT are so different there’s room for both and I wouldn’t be ignoring Gemini even if it isn’t ahem great at coding, since it’s quite obviously very good otherwise.
In the times before you also would rather have very smart people working together instead of one very smart dude alone even if he had an identical twin.
My go-to usecase for Gemini is summarizing Youtube tech-influencers.
How do you do it?
I basically open up a new conversation, copy paste a link to Theo's latest video and ask it to summarize his yapping :p
I copy paste the transcript because sometimes youtube has blocked AIs from scrapping
Would YouTube really block Gemini? I thought summarizing (or asking questions about) YouTube videos was one of their advertised features.
I don't like youtube gemini summaries. whenever I have asked gemini about detailed comprehensive report on a long hour podcasts, it always refuses. hence claude is my go report generation AI.
Over the past 6 months, Anthropic has made more waves as a product company than a frontier lab.
Cant wait to see their books once they IPO
Plateau on a saturated benchmark where an asymptote to 100% is mathematically baked in?
Or plateau as in it won't solve any Millennium Prize problems within the next decade?
This might come down to when you expect plateau to happen. You could have said similar about transistor density many decades ago.
(I'm not denying it could happen next year for all we know. But we simply don't know, and from what I hear from researchers the breadth of ideas we've tried are still small)
Aren't Anthropic afraid of Elon siphoning the model weights out from the network buses?
Is xAI a competitor worth worrying about?
They make good models, at times SotA (at least if you don't need coding, their last good coding model was six months ago), with lower safeguards than either Anthropic or OpenAI, and they still fail to capture meaningful market share or mind share. The name Grok is tainted by the twitter bot of the same name operated by xAI/X. Being owned by Musk lets the company appear unstable and untrustworthy in the minds of many. Their marketing game is just bad all around. They struggle to retain top talent.
Maybe their next model will be great. I doubt it will matter. I doubt xAI siphoning off Anthropic models and distilling that would matter. Model performance is not the main factor dragging down xAI
Theft of trade secrets. And so many people will have to be involved that evidence of the crime is bound to leak out.
Pretty sure models are encrypted all the way.
Can't run inference on encrypted weights and get any kind of performance out of it.
The overhead shrinks with larger models. It doesn't seem that bad.
https://arxiv.org/pdf/2409.03992v2
The whole system has encryption all the way through.
Otherwise, OpenAI/Anthropic would never use external clouds since the weights are some of the most valuable assets in the world.
Matmuls need access to decrypted weights to do their work.
Which means that getting the full weights out isn't even an "if" - it's "how much effort". The encryption wouldn't do much more than a gentleman's agreement would.
The only real move for Anthropic there is to outline contract penalties for letting weights get leaked, and never give less trusted external inference providers access to cutting edge system weights.
Exposure is limited either way. Opus 4.7 weights are a deprecating asset - it's bleeding edge today, very valuable now, but it'll lose a lot of its value the moment Opus 5.0 drops.
That would require hacking Nvidia's GPUs/racks to extract the weights. The weights are encrypted, sent to the GPU/rack encrypted. When it does inference, it will use decrypted weights but there is no way to get those weights unless you find a way to exploit Nvidia's GPU security.
Do you think OpenAI would send CoreWeave their GPT 5.5 Pro weights if an admin employee at CoreWeave can access the full weights unencrypted? Of course not.
It would require exactly that. A bit more involved than "scp that big file", yes. But you make a mistake by treating it as a hard blocker.
Like I said: it's a gentleman's agreement. If Musk said "I want Opus 4.7 weights", and those weights were on Colossus 1 hardware, he'd have those weights on his desktop, unencrypted, within a couple of weeks.
There's also the side channel line, because having inference on your hardware typically allows you to do things like snoop into KV cache and peek at per-layer, or even per-expert, residuals. Which allows for some very advanced distillation attacks. Might be easier/more deniable to pull that off than dumping full weights, in some circumstances.
Dude, Chinese labs distil attack via the APIs, if Musk wanted to do something like that, technically he could. Legally it would be a giant slam dunk liability though
Well, knowledge distillation requires a teacher model and a student model and the student model attempts to learn and extract and (preferrably) compress the information of the teacher model, so it is possible for model collapse due to high SNR in between [1].
What I suggested is to steal the (possibly intermediate) weight in between by sniffing the network communication bus, which means MITM for getting the exact values. Or unless it turns out OpenAI or Anthropic leveraged homomorphic encryption, or I'm not certain how is Anthropic would safely allow Mythos to run on AWS without their control.
[1]: https://en.wikipedia.org/wiki/Knowledge_distillation
Distilling is different from "siphoning the model weights". I would think that Anthropic has a system for this. After all, they deploy to different clouds already. Their weights are worth billions, I presume that they take security very seriously and have done a lot of homework to trust no one.
https://x.com/nottombrown/status/2057194829986300375?s=20
xAI cannot train models. Anthropic cannot do inference. The roles of these two companies have already been decided; one as a pure inference provider, the other as a pure model manufacturer.
Ok, maybe this question is a bit silly, but could it possibly be that Elon is doing this to steal Anthropic models secrets and using them to improve Grok?
Love the curiosity. I recommend you spend that curiosity on figuring out why both companies are in the situation they are that this deal is a win-win: Elon has provided a masterclass in public company launch and management over the last 15 years. (And in recent years done it amidst an incredibly divisive reputational environment).
Put another way - they could maybe try to do this. But it would be worth literally nothing compared to the actual benefits to SX of this deal. And add an immense amount of risk. There’s much more clever and interesting things to do.
No. Model weights are encrypted. Elon is doing this because Anthropic is desperate for compute, SpaceX needs to juice up its balance sheet before IPO, and xAI model team isn't competitive.
The model weights, sure he can "steal" them. But do what then? Serve it on the side?
Even distillation would be obvious.
And I wouldn't be surprised Anthropic is running the smaller inference models, keeping the base large model in machines they fully control/own.
SpaceX's data centre business is much more valuable than Grok, so this wouldn't make sense.
Valuable? If SpaceX is a majority "resell electricity" business, its valuation will be a tiny fraction of what they are trying to push.
You won’t see “resell electricity” on the IPO brochures. They’ll say something like multi billion dollar ARR hyperscaler business.
To the extent both are true, it’s non trivial to have large grid connections in the US these days or even gas pipeline connections hooked up to generators around a datacenter. Those assets are valuable.
They should set up a Doom server there.
I use Claude daily but I do not want that my spend is going towards Elon.
Might want to check the Anthropic parking lot, this might have already been happening.
I see the programming was effective on you.
Does an expansion of computing power on this scale imply that computing capital is displacing model architecture as the true moat in AI competition?
How come a company like Anthropic has invested in photonic computing? If it's good enough for Boeing, I'd assume they would at least invest in it. qc-LPU100 seems worth it for some matrix mult, if it can be proven its O(n) (at least less than O(n^3).
When they are doing Colossus (The Forbin Project), I wonder who will be building Samaritan (Person of Interest)?
Nobody uses grok except checking shit on X At least the hardware is not idle
Seems like either Grok is being shut down or it will be "powered by anthropic" soon.
Grok’s whole thing is being irresponsible and edgy. Can’t s ee Anthropic going for that, can’t see Grok’s customer base accepting an AI that refuses to do nonconsensual porn.
Model fine-tunes to change the 'edgy' tone are very cheap and easy to do.
Grok could easily be powered by Claude in just a few weeks engineering time.
Technically yes, but my point was that I do not think Anthropic would be receptive to a pitch to fine tune Opus to be an edgelord.
Related:
SpaceX S-1
https://news.ycombinator.com/item?id=48213933
War makes strange bedfellows.
Musk said Anthropic Claude was woke DEI until he said it wasn't. It must be hard for Musk fans to keep up.
I mean … they’re mostly sitting on their SpaceX stock and rubbing their hands gleefully. What do they have to keep up with?
Disclosure: former (small) SX investor.
You just made my sub text the text. Money makes the world go around. Who cares about values when there's money to be made?
I predict SpaceX will subsume Anthropic at some point.
Sigh.
I’d hoped Anthropic would steer clear of blatantly unethical practices but here they go in bed with that guy and his horribly damaging data center.
Anthropic works with Palantir, they are all but "ethical" lol.
I suppose you're also using Chinese models, who all have to be at the behest of the Communist party. China also have "horribly damaging" coal power plants. Do they acquire training data ethically? Do they censor? Be honest with yourself. Your concerns have nothing to do with ethic.
Yes, quite a bad reputation hit for Anthropic to get in bed with Musk. Tesla lost half its marketshare in Europe because Elon is meddling in our democracy, is promoting right wing populist parties and supporting Trump in general.
Europe has a hugely shrinking demographic, and is actively decoupling from the United States. There's almost no innovation and you are making the mistake of getting in bed with China and constantly suing our tech companies.
Europe needs to focus more on Europe and less US politics.
This might have something to do with the US president threatening to attack Greenland? Decoupling from the US is the only sane thing to do.
Denmark has a freakin' King. You guys are still a monarchy... Greenland is closer to the US than Europe, and we defended and gave back demark during WW2.
Demark provides almost no financial support to Greenland, and can't defend it. It will prove to be very strategically important as the ice caps melt.
Russia invaded Ukraine, and Demark lets russia use the Danish straights to transport oil despite the US asking them to not allow them entry. Demark has no way to protect it's own straits, how is it suppose to stop an invasion of Greenland from russia? How is it supposed to impose order and control the north artic if it can't even protect it's own straights?
These are very real security concerns from then US that a monarch can't address.
No, they’ve lost market share because they haven’t released a new product in years. Meanwhile China is releasing hundreds of excellent models into the European market which are better than Tesla’s.
That adds to it. But you are underestimating how much Europeans were/are shocked by Elon's Nazi salute and the DOGE fallout.
Pretty sure Americans hate him too. You can see Tesla new car staging lots full and the cars never move. They try to hide the Cybertruck lots.
American's love Elon... Just saying. Fun fact, most of us love Trump too :)
Anthropic is gross for this. The grandstanding about principles and values is intolerable.
Too bad Enron is still not around. They'd have some real fun with today's electrical markets.
I though Claude is too woke. Musk has posted that at least 50 times in the last year.
But booking outrageous rental fees as fake AI revenue ahead of the SpaceX IPO apparently takes precedence.
He posted that AI will kill everyone, then stopped when he started his own AI company.
He tried to steal the OpenAI charity, then started complaining about someone else stealing the OpenAI charity after he failed.
He attacks European democracies about free speech, but is a compliant little censor for Turkey, India and other non-European countries.
Anthropic is paying real cash, how is it fake revenue?
Fake __AI__ revenue. Maybe Hetzner should build Colossus4, rent it out and book it as AI revenue instead of hardware rental revenue and get a P/E of 100.
Anthropic this quarter will have revenue of $10.9B, up from $4.8B last quarter[0]. They're paying SpaceX $1.25B per month for compute[1] - which is more than what SpaceX earn on space. SpaceX spent about $30-40B in capex on Colossus 1 & 2.
This is all real revenue, real spend, real usage.
Hetzner just aren't at this scale. Not even close. If they wanted to get into this business - first, they're late. Second, it's at a scale of ~10x of their total lifetime datacenter buildout. Third, they'd need to change their business to being one that is debt fronted.
xAI have proven out that being able to deploy compute is a very viable business (and difficult to pull off)
At some point AI cynicism clashes with reality, it must be exhausting maintaining it.
[0] https://www.wsj.com/tech/ai/mind-blowing-growth-is-about-to-...
[1] https://www.wired.com/story/spacex-ipo-anthropic-compute-fin...
Thank you. It’s arguably even more tiring to hunt it down on HN, and often thankless.
"I'm an investor, the founder of new alchemy, co-founder Lamina1, former CEO of coinlab.com, Chairman of the Bitcoin Foundation, Ethereum security researcher, holder of patent 9298806, general nerd. Currently working at Capital6, my private equity fund."
You poor misunderstood Bitcoin booster! Now you have to deal with AI critics! How about getting a real job?
all of that is fake. The bitcoin foundation was sold for SEO purposes after a string of scandals.
Anthropic’s revenue is computed different from OpenAI. As I recall, they inflate it by including money that they end up just paying forward to some of the companies they depend on. OpenAI doesn’t count that component. And none of these companies - including SpaceX - have trustworthy accounting.
Jesus Christ, the Hetzner example is obviously an example of booking revenue as AI revenue (where investors assume it is generated by Grok subscriptions) vs. hardware rental revenue, which traditionally not valued as highly.
Nowhere does the hypothetical state that Hetzner, an example for hardware rental, has the funding or the capabilities to execute the sarcastic example.
But ok, now hardware rentals have a P/E of 100 or more.
Investors aren’t dumb. These numbers are being reported and the fact that the data centers are being rented out is publicly disclosed everywhere. Investors know full well that the revenue is from the data center rental. No (non-retail) investor is going to see the jump in revenue and think “I better buy up because grok must be kicking ass!”
And yes, if hetzner built a massive AI hyper scale datacenter and rented it out for billions, with the expectation that they would keep building more, they would also see massive PE ratios because it’s expected that their revenue would be going up.
I think you're low at 100. TSLA has 370 PE, and SpaceX is targeting 250+.
If Hetzner could build that, they would.
He talked with the Anthropic team, and his concerns lessened. It's actually a good thing to be able to change one's mind.
You’re falling for the Musk shell game. He just says whatever is convenient at any given moment. FSD. Funding secured. I talked to the Claude guys.
He didn’t change his mind about anything. It’s just that he wants to prop up SpaceX’s IPO as much as he can. Plus if he hits various targets, he unlocks more shares for himself, I think.
Keep drinking the kool aid
The eternal truth: money talks, bullshit walks
Musk will buy Anthropic and fix the wokeness. It was done before.
Don’t put it past Dario to buy spaceX
Elon will never sell SpaceX. And he controls 86% of votes.
Cool. Can they make Claude not absolute dogwater then?