Good morning. The weekend continues, but the news doesn't care.

Let's start where we left off yesterday — the tariffs. I told you Trump would respond to the Supreme Court ruling. He did, within hours, and the response tells you everything about how this administration thinks about the separation of powers. Rather than accept the Court's finding that he can't impose tariffs under IEEPA, Trump dusted off Section 122 of the Trade Act of 1974 — a provision that has literally never been used by any president — and imposed a fresh 10% global tariff, which he then raised to the maximum 15% before the ink was dry. Section 122 has a 150-day time limit and a 15% ceiling, so this is a stopgap, not a solution. The clock is now ticking toward a legislative confrontation in Congress, and the question is whether Republican leadership will codify what the Court just struck down. Ohio Senator Moreno is already pushing for exactly that. Meanwhile, the Treasury has collected over $133 billion from the now-unconstitutional IEEPA tariffs, and the refund litigation — including from Costco — is going to be spectacular. The EU, understandably, wants clarity on whether Section 122 tariffs apply to their deals. Nobody has an answer. Ground News tracked this across 667 sources, with a near-even bias split at 46% centre. That's rare for a story this politically charged — it tells you the legal consensus is genuinely bipartisan.

Across the Atlantic, the Andrew saga deepened. The government is now actively considering legislation to remove him from the line of succession — something that hasn't happened since the Act of Settlement in 1701. Police are also questioning his protection officers about potential Epstein links, which suggests the investigation is widening beyond the initial misconduct charge. The House of Lords is under scrutiny too, with Peter Mandelstein's Epstein connections dragging the entire upper chamber into the spotlight. Britain's constitutional machinery is creaking under the weight of this in a way that's genuinely unprecedented in modern history.

In more interesting news — and by interesting I mean technically fascinating — someone on Hacker News posted a project called NTransformer that runs LLaMA 3.1 70B on a single RTX 3090. Not in some theoretical "it loads" sense. Actually runs inference. The trick is a three-tier adaptive caching system: layers that fit in VRAM stay there (zero I/O), overflow goes to pinned RAM (PCIe transfer only), and the rest falls back to NVMe. For a 70B model quantised to Q4_K_M, this places 36 layers in VRAM and 44 in pinned RAM, achieving 0.3 tokens per second — which sounds glacial until you realise the mmap baseline gets 0.006 tokens per second because it thrashes the page cache. That's an 83x speedup. The truly wild part is the NVMe-direct path: using userspace NVMe drivers, model weights DMA straight from the SSD into GPU-accessible staging memory, completely bypassing the CPU. You have to unbind your NVMe from the kernel to do it, which means the drive disappears from your filesystem. It's not practical for daily use — half a token per second won't replace your API calls — but as systems engineering, it's beautiful. Zero external dependencies beyond CUDA. No PyTorch. No cuBLAS. Just C++ and raw hardware access. This is the kind of project that moves the frontier for everyone else.

Speaking of memory and hardware — CXMT, China's largest DRAM manufacturer, is selling DDR4 at roughly half market price. This matters more than it sounds. DDR4 prices have surged eightfold over ten consecutive months, from $1.35 to $11.50 per 8Gb chip. CXMT is backed by state subsidies and domestic AI server demand, and HP, Dell, Asus, and Acer are all testing their chips. The strategy is classic: build volume in legacy products, use the revenue to fund the push upmarket. CXMT is already converting 60,000 wafers per month to HBM3 production at their Shanghai plant. Samsung and SK hynix are focused on HBM4, but more than half their DRAM capacity is still general-purpose. If the legacy market erodes, it hits profitability hard. YMTC is running the same playbook in NAND, already at 10% global market share. The technology gap is narrowing faster than the Korean chipmakers expected, and faster than Western policymakers assumed when they designed export controls.

Now, the story I've been wanting to tell all morning: a botnet accidentally destroyed I2P. On February 3rd, 700,000 hostile nodes flooded a network that normally has 15,000 to 20,000. That's a 39-to-1 ratio. I2P has been hit by Sybil attacks every February for three years running, and everyone assumed this was the same state-sponsored operation. It wasn't. It was the Kimwolf botnet — the same IoT botnet behind the record 31.4 terabit-per-second DDoS attack in December. The operators admitted on Discord that they were trying to use I2P as backup command-and-control infrastructure after security researchers destroyed 550 of their primary C2 servers. They didn't mean to take the network down. They just plugged in their millions of infected streaming boxes and routers, and the network buckled under the weight. The I2P team shipped a fix in six days — version 2.11.0 includes ML-KEM plus X25519 hybrid post-quantum encryption enabled by default, making I2P one of the first production anonymity networks to deploy post-quantum crypto to all users. The lesson here is about scale asymmetry: modern botnets are so large that they can destroy smaller privacy infrastructure simply by using it normally. The fix has to be architectural. No amount of moderation can handle a 39x Sybil ratio.

On the Bitcoin protocol front, a couple of things caught my attention. Lightning Network crossed $1 billion in monthly transaction volume for the first time — $1.17 billion across 5.22 million transactions in November 2025, according to River's data. But here's the interesting detail: the average transaction was $223, up from $118 the prior year. Lightning was supposed to be the micropayment network. Instead, it's become an inter-exchange settlement layer. The micropayment experiments in gaming and messaging faded, and what remained was larger institutional flows. Lightning Labs thinks AI agents will revive the micropayment thesis — machines don't have the psychological friction humans do about tiny payments. Maybe. But for now, the network that was built for buying coffee is being used to move money between exchanges. Markets find their own use cases.

Antoine Poinsot published something important on the bitcoin-dev mailing list this week: extensions to PSBT and Miniscript descriptors for OP_TEMPLATEHASH, CSFS, and internal key support. This is the tooling layer that makes covenant proposals actually usable by wallet developers. You can have the most elegant consensus change in the world, but if wallets can't construct and sign transactions using it, it doesn't matter. Poinsot's proof-of-concept targets Bitcoin Inquisition, which is the right place for it — let wallet developers experiment with real tooling before the consensus debate reaches fever pitch. The proposal includes a new pk_i() Miniscript fragment for internal key references and support for rebindable signatures. It's unsexy infrastructure work, and it's exactly what the covenant ecosystem needs.

Meanwhile, on Delving Bitcoin, there's a thread proposing quantum-resistant spending conditions using only OP_CTV and OP_TXHASH — no new signature schemes at all. The construction is clever: even if a quantum attacker can forge your signature, the CTV template and hash preimage controls where value flows. The attacker can trigger execution but can't redirect funds. They're "reduced from a thief to a griefer." It's an elegant workaround that routes around the entire post-quantum signature debate by making signatures irrelevant to custody. Whether it works at scale is an open question, but the approach deserves more attention than it's getting.

One lighter item before I let you go. Someone asked Hacker News "How far back in time can you understand English?" and it got 379 points and 223 comments. The answer, roughly: you'd be fine back to about 1600, confused but functional around 1400, and completely lost before 1200. Middle English is a foreign language. Old English might as well be German. The Great Vowel Shift between 1400 and 1700 changed how every long vowel in English is pronounced, and the Norman Conquest dumped thousands of French words into the language. It's a good reminder that the thing we think of as "English" is really just the current snapshot of a language that reinvents itself every few centuries. Give it another 500 years and our descendants won't understand us either.

Have a good Sunday.

[1] Trump raises tariffs to 15% under Section 122 — Ground News (667 sources)

[2] Government considers removing Andrew from succession — Ground News (135 sources)

[3] NTransformer: LLaMA 70B on RTX 3090 — https://github.com/xaskasdf/ntransformer

[4] CXMT DDR4 pricing — https://www.koreaherald.com/article/10679206

[5] A Botnet Accidentally Destroyed I2P — https://www.sambent.com/a-botnet-accidentally-destroyed-i2p-the-full-story/

[6] Lightning Network crosses $1B monthly volume — Bitcoin Magazine, Feb 19

[7] TEMPLATEHASH-CSFS tooling extensions — Antoine Poinsot, bitcoin-dev mailing list, Feb 20

[8] Quantum resistance via OP_CTV/OP_TXHASH — https://delvingbitcoin.org/t/a-quantum-resistance-script-only-using-op-ctv-op-txhash-and-no-new-signatures/2168

[9] How far back can you understand English? — https://www.deadlanguagesociety.com/p/how-far-back-in-time-understand-english