Tensorial

@charliemktplace @WillManidis You are retarded mate

@WillManidis "He proposes" - he did not propose it!!! He is using theory of mind to say what he expects trump to do based on what trump has done in the past. Literacy crisis! Reopen the schools!

My friend Dean Ball has advanced an argument for the de facto protection of American frontier intelligence providers. Dean does not propose banning Chinese open-weight models. Banning things requires Congress. He proposes something more characteristic of the modern administrative state: every agency issues enough warnings, bulletins, and speculative security notices that no regulated company will risk touching them. Even a reader sympathetic to Dean would call this protectionism, and protectionism has a long history in America. More precisely, it's a proposal to use the informal, coercive power of the terminal, late-stage bureaucratic state to clear the American market of a cheaper frontier competitor to OpenAI or Anthropic. But throughout the history of American industrial protectionism, it has always had two features. First, it's done in the daylight, and two, it comes with a bill. In the spring of 1952, the United States was fighting a war in Korea. Truman concluded that a shutdown would endanger soldiers abroad and ordered the Secretary of Commerce to seize and operate most of the nation's steel mills. The Supreme Court sent him straight back to Congress in the Youngstown Steel case. Justice Black, writing the majority's opinion, begins with the rule that Dean's proposal is seemingly designed to evade: that presidential power "must stem either from an act of Congress or from the Constitution itself." It's easy to flatten the Youngstown decision into the proposition that the president could not seize a steel mill. Its actual lesson is subtler: that an emergency does not dissolve the difference between making a law and executing one, that the importance of the object does not create the authority, that the inconvenience of the regulatory process is not inherently a source of presidential power. Truman's approach failed not because steel was unimportant, but because it was so important that the constitutional bargain had to be made and the policy had to be carried through the front door. Much like policy proposals from the rest of the AI agenda, Dean is proposing a smaller action in formal appearance and a much larger one in practical effect. We will not ban Kimi, we will not prohibit it from use, and we will certainly not publish a rule declaring Chinese weights unlawful. But we will whisper about it. A regulator may even ask management whether it has considered the reputational consequences of relying on the Chinese model, but the agency certainly will never be coherent enough to ask anyone to stop. It merely ensures that continuing becomes professionally indefensible. This is how we grow the administrative state, with bureaucrats that we placed in these roles, without accepting responsibility for the actual process of governing. America has tried this experiment before. Operation Chokepoint didn't make payday lending, firearm sales, or any of the other seemingly distasteful businesses caught in its net illegal, but it encouraged banks to understand that serving legally disfavored customers would invite regulatory interest. We didn't pass a law, we simply just asked, "Are you sure you really want to be doing this?" Reputational risk was powerful precisely because it's not law. It has no limiting content. A regulator did not need to identify a violation or even a material financial risk. He only needed to make the bank afraid of being asked what was actually going on here. The analogy is almost embarrassingly exact to Dean's policy proposal. Dean need not prove that a Chinese model contains a backdoor, nor prove that it uses any more distillation than American models do. He simply needs to announce that there may be one. The agency does not need to order a company to stop using it, but simply ask whether management has considered the risk. The absence of formal policy is by design. The Supreme Court dealt with this technique in NRA v. Vullo. New York's financial regulator could not directly punish the NRA's speech, so she allegedly pressured the insurers and banks she regulated to sever their relationships with it. The Court's rule was unanimous: government officials may not use their offices to "coerce private parties" into suppressing what the government disfavors. The communication must be understood in the context of the regulator's power, including the regulated party's knowledge that the person offering advice can also investigate, prosecute, fine, and settle. The current administration has gone even further. In April 2026 the FDIC and OCC issued a final rule to prohibit regulators from criticizing institutions, formally or informally, on the basis of reputational risk, and from encouraging banks to deny services to lawful but politically disfavored businesses. In June, the federal banking agencies removed the remaining references to reputational risk from their supervisory materials. Dean is proposing that this administration recreate for AI the same machinery that all of us argued against when we were widely debanked. A government that can quietly remove Kimi from the market can also quietly remove gun makers, crypto companies, churches, newspapers, or American open-weight models from it. The bureaucracy does not remain attached to the intentions of those who staff it at the current moment. You don't get to build this machine just because your friends happen to be in office right now and keep it pointed at where you left it. Protectionism through a whisper is not a more modest protectionism than by law. Protectionism also has always come with a bill. OpenAI and Anthropic increasingly speak of themselves as national institutions. Their compute is "strategic infrastructure," their losses are "national security losses." Their competitors are not just competitors, but instruments of hostile states, and their access to power, chips, capital, copyrighted material, and public customers is a matter of national survival and great power competition. When Washington decided that the atom was too dangerous and too important to remain an ordinary private business, Congress created the Atomic Energy Commission and transferred the Manhattan Project assets and responsibilities to it. Production facilities and reactors were government-owned, and technical information sat under federal control, and private participation only returned later through a statutory licensing regime. The existential framing of the atom by its greatest proponents produced public control. When national security concerns helped to preserve AT&T's integrated position, that is, a monopoly, in 1956, Bell did not receive this protection for nothing. The consent decree required compulsory licensing of roughly 9,000 patents and restricted Western Electric's commercial activity outside the telephone system. The settlement diffused the inventions accumulated inside the protected monopoly into the broader economy before breaking it up just a few decades later. The pattern is really simple. It's not that every tariff necessarily demands nationalization. It's that the bigger the shield you are asking for, the bigger the bill you owe to the American taxpayer. And OpenAI and Anthropic have been unambiguous about asking for the biggest shields of all time. Listen to what they are asking for: public infrastructure, privileged energy, federal preemption of state law, favorable copyright treatment, government contracts, export controls, and a domestic market swept clear of their strongest price competitor, all filed under national security interests. And what do they want to pay? Almost nothing. OpenAI has floated giving 5% of the company to the American taxpayer. They would like the benefits of nationalization at the price of being an ordinary public company. There is also a profound moral hazard buried in Dean's proposal, as well as adjacent commentary on this. The labs say the Chinese companies distilled their models. Perhaps they did. Perhaps distillation matters. And perhaps the Chinese labs are running distillation attacks on scales that the Western labs are. I can't be sure of this. But if the reward for failing to secure an API is that the government removes the resulting competitor, the taxpayer is paying the lab to be careless. We know how to secure an API. Know-your-customer laws exist. Access controls exist. Extraction detection exists. If you spend some fraction of the hundreds of billions being raised to defend the asset whose theft is said to threaten the republic, you might be able to stop some of this. Theft remains theft when the lock is bad, but the owner of a badly secured store does not receive ownership of the street for his failure to protect it. Dean's fourth point is that open-weight AI ends in communism: the state builds the training runs and subsidizes the product of intelligence and gives the models away. But, at least for me, this is not a particularly Chinese idea, but one of the most American ones imaginable. The roads we build are public. Our radio spectrum is publicly allocated. The government funded the early internet and much of the research base behind modern computing. The state is welcome to build a platform, and American businesses are welcome to be built on top. Just because they're bad for our market position doesn't mean we get to call them Chinese in some fundamental way. There will be inference companies and application companies and security companies and fine-tuning companies and data companies and chip companies and 10,000 businesses we don't even have names for yet. A public road existing does not abolish the trucking industry, nor does it nationalize it. Sure, this may reduce the value of a couple trillion dollars of equity in the first generation of model companies, but it's certainly not communism. This technology may be civilizational without its present owners being permanent. And that is the thing that I feel like none of you will say out loud: that AI is welcome to be a civilizational technology when we ask for support, and an ordinary private product when anyone asks what the public receives in return. The United States has two honest options. First, treat AI as a competitive industry. Then the answer to Kimi is a better model, run cheaper and exported harder, with written rules excluding Chinese systems from defense, intelligence, and critical infrastructure when a concrete security case can be made. Or two, decide frontier AI is too important for ordinary competition. Protect the labs through pseudo-nationalization, guarantee there's a market for them, and exclude the rivals. But in that second case, the American taxpayer must be paid, likely through a majority of equity in these companies, if not full nationalization. What no one gets is that private upside, public infrastructure, government-mandated scarcity, and immunity from cheaper competition delivered through a late bureaucratic state issuing warnings is a disgusting ask for something that is easy to name: regulatory capture. There is a serious American argument for protecting industries that we can't afford to lose. But there has never been a serious argument for doing it invisibly, for free, through a bureaucracy instructed to manufacture fear, even if we can do it because our friends happen to be in office right now. If the labs want to be protected, they should ask for it in the way that Americans have always asked for it. In public. With a price.

@__paleologo @Google When are you joining Anthropic?

Mini-rant. Every time I open gmail on the browser, I get this message: "Oops, something went wrong". To fix it, I try everything suggested by OAI. Then I contact @google customer service. After an hour of chat, of sending all possible screenshots (in multiple browsers also incognito mode), the exchange with OAI and the remedies, and then even a video of the event (!), the representative says "oh sorry, we are google one, not workspace." me: "can you at least forward the exchange?". Of course not. You have to spend another hour. It seems completely absurd. I am 90% curious about fixing a google problem *for them*. I have been thinking about what makes tech companies good. A sign of organizational rot is the "I don't care about what's good for the firm, this is the process" and it close cousin "This is very hard to do", and the "Why make trouble? This works well". Nothing works quite as expected, both at great companies and mediocre ones. It's just that these expectations are different.

@0xBADB01E Incredible work chief!

First you have to understand that modern LLM inference already disaggregates weights as models outgrew single chips years ago. You shard either by layer (pipeline parallelism) or by slicing every layer (tensor parallelism), and the two do very different things. As an example, let’s look at Llama 3.3. It has 70B of weights and at FP8 that’s 70 GB of memory which is enough to fit on a single H100. Now that H100 has 3.35 TB/s of HBM, so the fastest it can ever decode for one user is 70/3.35 ≈ 21 ms/token or ~48 tok/s while using under 1% of its FLOPs. Now if we pipeline it across 8 chips: each chip holds ~8.75 GB, which means it only needs 1/8th the bandwidth and 1/8th the FLOPs to sustain the same aggregate throughput. Now crucially the token/sec a user gets is limited by the amount of data that crosses the link. In current LLMs all that is a small amount of activations for LLama 3.3 it’s ~8 KB per token…. Yes, you read that right it’s 8 KILOBYTES we are sending over a <900 GB/s link. That’s only 9 ns of serialization time but the overhead of 224G PAM4 SerDes adds ~100 ns per link traversal with RS-FEC which is 11x longer than the payload itself. And then you have the NVSwitch adding ~300 ns per hop and you need to pay twice. That’s ~600 ns of just hardware latency wrapped around 9 ns of data making a 98% tax before software even shows up. Then NCCL’s collective stack turns 600 ns into 10-20+ us… all to move 8 kilobytes lol. For comparison 8 KB serializes over 10 Gigabit Ethernet NRZ, in just 6.6 us. Pipeline parallelism however doesn’t make a single user faster as the token still needs to visit every layer in the sequence, so per-user speed is still weights / per-chip bandwidth. To get more speed per user token you need to use tensor parallelism and have all the chips work on the same layer simultaneously. TP costs you 2 all reduce OPs per layer, 160 per token on llama 3, that’s still kilobytes of traffic but with NVLink overhead it’s a massive tax and why pipeline parallelism on most models still gives more interactivity per user. However, this gives you a huge latency lever to pull that scales tokens per second with interconnect speed instead of memory BW. The clever amongst you might have also realized that sharding doesn’t just cut memory bandwidth per chip it also cuts FLOPs per chip and is why we have such bad MFU on decode. So once you’ve sized the link for the memory, you need to size the compute for it too. This is called “balancing the pipeline”, and currently no shipping chip does it because they were all designed as standalone monsters. Remember Tokens/sec = ~aggregate memory BW / bytes touched per token. At batch 64 in FP4 you need ~250 FLOPs per byte, and Blackwell ships 1,250. Provisioned 5x more than the narrow pipe of HBM. Nobody saturates shit cause they are all building around HBM. So now it all comes full circle. Parallelism reduces memory bw pressure and thus FLOPs but increases interconnect latency pressure. Despite having HBM and GigaSERDES we aren’t actually doing more work lol. But if you really wanted to balance the pipeline you need to match the memory bandwidth, the flops, and most importantly the interconnect. So what does that look like ? Well if you build around LPDDR’s lower bandwidth, lower your interconnect latency, you actually can beat Nvidia on decode with a fraction of the silicon.

I agree with this post whole heartedly but I’d push it even further. The interconnect IS the binding constraint for AI even more so than memory. If we want faster inference & training with better economics we are best served by designing our interconnect first and then working backwards towards the optimal chip architecture. Today’s chips weren’t really designed with this principle in mind. There is no better example than running autoregressive decode on a GPU. Despite all those reticle sized logic dies & CoWoS integration decode runs at under 20% of peak FLOPs on Blackwell, wasting silicon and burning power while waiting for memory. The naive solution has always been to increase memory bandwidth whether that’s adding more HBM or using SRAM. However, that is a vast simplification of the problem which I’ll explain later. But if you were clever you’d have realized while reading that you could feed those idle FLOPs by streaming weights over the interconnect itself. Wallah 🪄 you just discovered the idea that forms the basis behind disaggregated memory from first principles. But sadly this currently doesn’t work on Nvidia’s hardware. NVLink5 carries 1.8 TB/s against 8 TB/s of local HBM, and scale out is 80x behind that. The “pipe” is smaller than the memory at the other end and thus leads to worst token/sec if its relied upon. But we get an interesting lemma out of this which is that remote memory is only as fast as the interconnect. Therefore you must balance the pipe for the memory it attaches to. SRAM needs an 80 TB/s link, HBM needs 2+ TB/s, and LPDDR gets away with a couple hundred GB/s. So Nvidia selling a rack of 72 GPUs, each GPU’s memory is pretty segregated. The core idea is still sound though but this raises a question, why would Nvidia build a fabric that’s high bandwidth and high latency leading to memory access being segregated per GPU? It’s because they were optimizing for training over inference. Training is dominated by collectives on huge tensors, and a couple microseconds of latency on a huge all reduce operation is just noise so the bandwidth gains justify the latency tradeoff. But more importantly, this also works because it matches what the chip is good at. GPUs are great at hiding latency with occupancy (also what allows them to be OK for training) but bandwidth is the only thing warp switching can’t create. You can justify a 224G PAM4 + FEC with overhead when you have a chip that’s designed to be latency tolerant as well. It’s a latency tolerant fabric for a latency tolerant chip. Maybe a good design for training but inference inverts this completely. Now everyone knows decode is bandwidth bound so you might assume again that more/higher BW memory and thus higher BW interconnects are necessary. However, it’s the exact opposite and the name of the game is actually lower latency and that’s why despite having high bandwidth memory MFU on decode is still so low and also why I made the point earlier that the interconnect is MORE important than the memory itself as well as the chip architecture. In part two I will explain why lower latency interconnects are not just ideal for inference but also allow you to get away with a smaller cheaper memory and a simpler chip architecture.

@bscholl Just $50k? That is nothing lol.

Boom will be offering a prize—$50k cash and $50k in Boom stock—for the first amateur-built RC airplane to exceed Mach 1.

Blessing ya all ears😊.

Dear brothers and sisters in Islam, the reciter below has passed away. Please help spread this beautiful Qur’an recitation he left behind.

@MrsOWalker @cirnosad Metal undergoes elastic and plastic deformation. Plastic deformation is when it can’t go back to its previous shape. Plane impact caused enough plastic deformation + heat weakening strength. But look, I do believe 9/11 was inside job. It still didn’t look normal

@sjdu32142 @cirnosad How would momentum and energy bring the building down an hour later? That momentum and energy were already absorbed and dissipated.

Plane hits truck: Truck rolls over. Plane disintegrated. Keep this in mind for later. And no, speed wouldn't change things, it would only make things even worse for the plane.

@Philly____3 @cirnosad Exactly the same thing if the ice cream had the same amount of fuel inside it. Mass and volume same asw

@sjdu32142 @cirnosad What would happen if the plane was made out of ice cream mr physicist?

@cirnosad I am a physicist. I’m not denying it being an inside job btw. Plane + demolition. Either way, it wasn’t Muslims

@sjdu32142 You simply have no idea what you're talking about lol.

@WillManidis Mate, you come off as more than anything honest in your pursuits. I would encourage you learn properly about the Lord God in Islam - Allah. Not about Islam, but about God in the way he presents himself in Islam. You will find the level of profundity you’re after in your writings

@Adriandhy I don’t like that last sentence. It sounds really Christian. Allah is free of wants and needs. It’s a very Christian phrasing to say Allah wants to spend time with you. 50 prayers was simply the weight of the initial obligation on us

@__paleologo Gappy this movie fails in the second half. Don’t praise so much. Eggers struggles with overall cohesion in his stories

PSA: "The Northman" is streaming on Netflix. Every movie by Eggers is a masterpiece of sorts and a retelling of some myth. Maybe "The Iliad" would be more suited to be directed by him than Nolan. But ok.

@ElifYilmaz78087 @MK_VOXX @liammendes Stop with the nonsense. She was saying as far as she knew he was just a normal quite well behaved boy until this event. And that from what she knew of him Anton would want a son like that. Shes expressing shock and her level of ignorance

@MK_VOXX @liammendes He is better ( for her) since he kills jews

The close relatives of the alleged Bondi Beach terror attackers Naveed and Sajid Akram have themselves called police, complaining about media knocking on their door on two occasions since they returned to their home just after midday. Full coverage: l.taus.au/0dE

@VictorTaelin Let’s go. I feel we will soon reach access to that level of leisure Karl Marx was talking about.

seriously, look at these commits and tell me, with a straight face, that I'd be able to be that productive without AI I still remember that famous guy who blocked me for daring to say AI boosts my productivity btw I've also been playing lol, and just finished hunter x hunter

@F1 Visor cams are just is beautiful man

And here's Max Verstappen completing his fastest lap on Friday ⏱️ #F1 #AbuDhabiGP

TRACK ⚔️ DOMINATION Lando with the edge heading into Saturday 👀 #F1 #AbuDhabiGP

@llallawg @Rohanburdened @skdh Yes it is precisely a ball. How else would you define a ball?

@Rohanburdened @skdh I mean you can just arbitrarily define any word in that way The question is it really a ball

another fun math fact: a ball of radius 1 has maximal volume in 5 dimensions. (normalized, ie prefactor only.) x.com/rob_mcrobberso…

@ieatcrayons_3 There has to be a catch right?

nearly 200 years of the smartest people who’ve ever lived are chasing 0.5% gains (pic related) because turbomachinery for power generation might be the single most financially incentivized optimization problem ever, and SCO2 has a free 10% thermal efficiency gain. like cmon

@noor_supernova7 .

My Name is NOOR . I don't need likes, I don't need views. I just need food 🙏🥺 If you're browsing, please leave a dot. It's just a dot