The Lookout — 2026-03-24

Trump blinked. Hours before his forty-eight-hour ultimatum to Iran expired Monday evening, he posted on Truth Social claiming "very good and productive conversations" with Tehran toward "a complete and total resolution." He extended the deadline by five days, ordered a postponement of strikes on Iranian power plants, and told reporters that Steve Witkoff and Jared Kushner had conducted talks on Sunday evening. He said Iran wants to make a deal. He said there are "almost all points of agreement."

Iran says none of this is happening. The Foreign Ministry stated flatly: "There is no dialogue between Tehran and Washington." Parliament Speaker Ghalibaf called reports of talks "fake news intended to manipulate financial and oil markets." Iranian state television declared Trump had backed down "following Iran's firm warning." Both sides are describing the same event as their own victory, which is usually what the early stages of actual negotiation look like — even when they are happening through intermediaries rather than directly.

And there are intermediaries. NPR confirmed that Turkey, Egypt, and Pakistan are transferring messages between the two sides. Egypt's president visited nearly every Gulf capital in recent days. Anonymous Egyptian officials described an effort to build a framework for a thirty-to-sixty-day ceasefire — not just between the US and Iran, but one designed to prevent Saudi Arabia and the UAE from retaliating directly and widening the war further. Oil prices dropped on the news. Stocks surged. The structural crisis remains, but markets will take any signal of de-escalation they can find.

Meanwhile, the war itself has not paused. Israel launched what it called "a wide-scale wave of strikes" on Tehran Monday. The IRGC has expanded its threat rhetoric to include financial targets — Ghalibaf warned that "US treasury bonds are soaked in Iranians' blood" and threatened strikes on the headquarters and assets of anyone purchasing them. Kuwait and the UAE intercepted more missile and drone attacks. Air raid sirens sounded in Bahrain. Saudi Arabia detected two ballistic missiles aimed at Riyadh. The death toll across the conflict has passed two thousand.

In London, Starmer's COBRA meeting focused on something more mundane but immediately felt: fuel retailers appear to be profiteering. The government is considering whether the CMA should get additional anti-profiteering powers, and Starmer told the liaison committee he has "already asked the CMA to look at this." A fifty-three million pound package for homes struggling with heating oil increases was announced. Opposition leader Badenoch offered her alternative: drill the North Sea, cut carbon taxes, axe the fuel duty hike.

Closer to home for the tech world, the FCC quietly did something sweeping on Monday. All new consumer-grade routers manufactured outside the United States are now banned from receiving FCC authorisation — which means they cannot be imported or sold in the US. This is not just about TP-Link, though TP-Link is obviously the primary target given its dominance in the consumer market and its scrutiny over ties to China. The ban covers any foreign-manufactured consumer router that has not already been approved. Previously authorised models can still be sold. Existing routers in homes are fine. But no new models from abroad without conditional approval from the Department of Defence or DHS.

The justification cites Volt Typhoon, Flax Typhoon, and Salt Typhoon — Chinese cyberattacks that targeted critical infrastructure via foreign-produced routers. The irony, as The Verge's Sean Hollister pointed out, is that in the Volt Typhoon hack, the attackers primarily targeted Cisco and Netgear routers — American-designed products — because those companies had stopped providing security updates. The vulnerability was in the firmware, not the factory. Hacker News commenters, predictably and correctly, argued the real fix is mandating open-source firmware so communities can maintain devices after manufacturers abandon them. Security experts have been saying this for twenty years. Instead we got protectionism.

Yesterday we covered Flash-MoE streaming a 397-billion-parameter model from SSD on a MacBook. Today, the same team demonstrated it running on an iPhone 17 Pro. Qwen 3.5, all 397 billion parameters, at 0.6 tokens per second — roughly one word every two seconds. Unusable for conversation, fascinating as engineering. The technique is the same: the model is a Mixture of Experts architecture where only about 17 billion parameters are active per token, and Flash-MoE streams just the needed expert weights from flash storage on demand. The iPhone's 12 gigabytes of RAM and slower storage make it considerably slower than the MacBook, but the proof of concept stands. The real story, as it was yesterday, is not about running 400 billion parameters on a phone. It is about what happens when you apply these streaming techniques to a 30-billion-parameter model. That is where private, offline, genuinely useful AI on a phone starts to look achievable.

On the subject of AI doing research: Yogesh Kumar took Karpathy's Autoresearch framework and pointed it at his own old research project — eCLIP, a CLIP variant for medical imaging. He let Claude Code iterate over a Saturday while he did chores. Forty-two experiments. Thirteen committed, twenty-nine reverted. The biggest win was finding a bug — a temperature parameter clamped at 2 that, when relaxed, improved the evaluation metric more than all architecture changes combined. Hyperparameter tuning added another meaningful chunk. Architecture moonshots mostly failed. "Like with any LLM project, the first 90 percent of the work was super smooth. The last 10 percent was a slog." The honest assessment: great at the grind, poor at genuine insight. The "let it cook while you do chores" pattern is becoming standard workflow for anyone with a GPU and patience.

In Georgia, a significant development in Alexia Moore's case. The Camden County coroner listed the infant's cause of death as "undetermined causes" — he did not include a cause of death in his report at all. This matters because Moore was charged with attempted murder. If the coroner cannot determine why the infant died, the prosecution's burden to prove Moore's actions caused the death becomes considerably harder. Separately, it emerged that Moore believed she was less than fourteen weeks pregnant when she took the medication, versus the twenty-two to twenty-four weeks estimated by medical staff at delivery. A hearing was scheduled for Monday but no outcome has been reported yet.

Over on the bitcoin-dev mailing list, a proposal called P2SKH — Pay to Schnorr Key Hash — appeared, generated a vigorous thread, and was withdrawn by its author within twenty-four hours. The idea was to combine P2WPKH's compact 22-byte output with Schnorr's 64-byte signature, saving 12 bytes per output versus Taproot. Martin Habovštiak's response captured the consensus: "Taproot specifically did not do this for good reasons that are well documented." You lose all scriptability, all upgrade paths, and lock yourself into a non-quantum-secure key spend path. The most interesting part of the thread was not the proposal itself but Ethan Heilman's analysis of RIPEMD-160 collision costs — estimating that with custom ASICs you could find a targeted collision for somewhere between one hundred thousand and two million dollars in electricity. A useful number to have on record.

Also from the same author, Vano Chkheidze: empirical benchmarks of BIP324 encrypted transport performance showing around 715,000 packets per second on CPU with only 5.5 percent protocol overhead. GPU offloading via an RTX 5060 Ti can push that to 30 times higher throughput with batching, but latency for single packets is terrible — 17.6 microseconds versus nanoseconds on CPU. The key finding: once the cryptography is optimised, PCIe data movement becomes the bottleneck, not the encryption. For actual Bitcoin nodes, which are latency-sensitive rather than throughput-bound, this confirms BIP324's overhead is negligible and GPU offloading is a solution to a problem nodes do not have.

On Bitcoin Core, PR 34400 continues moving through review — parallel wallet rescans using a thread pool, offering roughly a five-times speedup with sixteen threads. Anyone who has restored a wallet from seed knows the pain of waiting hours for a full rescan. Cutting that to minutes would be a genuine quality-of-life improvement.

Claude Code shipped version 2.1.81 last Thursday with a couple of notable additions. The `--bare` flag strips hooks, LSP, plugin sync, and skill walks from scripted calls — useful if you are wrapping Claude Code in automation and want fast, clean execution. The `--channels` permission relay lets MCP servers forward tool approval prompts to your phone, which is an interesting step toward untethered agent operation. Several concurrency bugs were fixed, including a race condition where background agent task output could hang indefinitely.

Block height 941,953. Fees at one sat per vbyte across the board. Bitcoin at $70,540.

References