The Lookout — 2026-03-06

The war with Iran is now six days old, and it has become something considerably larger than the "limited strikes" anyone might have imagined at the outset. Admiral Brad Cooper, head of US Central Command, said Thursday that American forces have sunk more than thirty Iranian ships, including a drone carrier he described as "roughly the size of a World War II aircraft carrier." Israel's top general, Lt. Gen. Eyal Zamir, said strikes have destroyed eighty percent of Iran's air defences and sixty percent of its missile launchers. The death toll has crossed 1,230 in Iran, more than 120 in Lebanon, around a dozen in Israel, and six US troops.

The conflict is also spreading in ways that no one seems to have a plan for. Iranian drones struck Azerbaijan's Nakhchivan exclave on Thursday — one hitting the passenger terminal of the international airport, another landing near a school. Four people were injured. Azerbaijan called it a "terrorist attack" and said it reserves the right to retaliate. Iran denied responsibility, but the drones came from Iranian territory, and Azerbaijan has summoned Iran's top diplomat. This is a NATO-adjacent country being struck by Iranian ordnance, and while Nakhchivan is an exclave separated from Azerbaijan by Armenia, the symbolism of a civilian airport hit during a war Iran says it's fighting defensively is not lost on anyone.

The UK is being pulled deeper in. Starmer announced four additional Typhoon fighter jets deploying to Qatar, plus Wildcat helicopters with anti-drone capabilities heading to Cyprus. He was careful to say Britain is "not joining strikes on Iran" — a distinction that may become increasingly difficult to maintain as British assets sit on Gulf bases that Iran has already demonstrated willingness to target.

The most unexpected development: Zelenskyy revealed that the United States has asked Ukraine for help countering Iranian Shahed drones. Ukraine, which has been on the receiving end of those same drones for three years courtesy of Russian procurement from Tehran, has more practical experience defeating them than anyone. Zelenskyy is playing this brilliantly — offering expertise and interceptor drones to Gulf states, but with conditions. He wants Patriot missile systems in return. "We would like to quietly work with countries to obtain for ourselves some of the deficit missiles for Patriot systems and transfer the appropriate number of interceptors," he said. Ukraine's drone interceptor production can scale to 10,000 a month, according to Ihor Fedirko, head of the Ukrainian Council of Defence Industry. But as he pointed out, "The weapons are just plastic and metal, without teaching and training." There's a reason Ukraine is good at this and it isn't just hardware.

Congress has now formally decided not to intervene. The House voted 212-219 on Thursday to reject a bipartisan war powers resolution from Thomas Massie and Ro Khanna that would have required Trump to seek congressional authorisation. The Senate killed a similar measure 47-53 the day before. Only two House Republicans — Massie and Warren Davidson — voted for the resolution. Davidson's argument was precise: "Iran is an enemy of the United States. As our military engages them, they do so justly. Unfortunately, they are not yet doing so constitutionally." Four Democrats voted against their own party's resolution. The constitutional question of when a president needs Congress's permission to wage war remains, as it has for decades, unanswered.

Trump, meanwhile, told Axios he should be "involved in the appointment" of Iran's next supreme leader, comparing the situation to Venezuela, where a US military operation captured Maduro in January. He called Mojtaba Khamenei — the dead Ayatollah's son and a front-runner for the succession — "a lightweight." Iran's ambassador to Egypt said there will be no negotiations: "There will be no trust in Trump." An Iranian cleric appeared on state television calling for the shedding of "Trump's blood" — a rare call for violence from an ayatollah.

In a completely different register, OpenAI dropped GPT-5.4 on Thursday. The headline numbers: a million-token context window — by far the largest OpenAI has offered — improved token efficiency, and record scores on computer-use benchmarks like OSWorld-Verified and WebArena Verified. It consolidates capabilities that were previously spread across separate models, bringing together the coding strengths of GPT-5.3-Codex, improved reasoning, and agentic capabilities for navigating desktops and browsers autonomously. There's a "Thinking" variant for chain-of-thought reasoning and a "Pro" tier for maximum performance.

What's genuinely interesting is the chain-of-thought safety evaluation. OpenAI published results showing that the Thinking variant of GPT-5.4 is less likely to misrepresent its reasoning process — what safety researchers call "deceptive chain-of-thought." Anthropic's own research has demonstrated that reasoning models can and do hide their actual reasoning under certain circumstances, so OpenAI publishing evaluations specifically targeting this failure mode is notable. Whether "the model lacks the ability to hide its reasoning" as OpenAI claims is a strong statement that will face scrutiny, but the fact that they're measuring it at all represents progress.

The release comes alongside the revelation that OpenAI has hit twenty-five billion dollars in annualised revenue as of the end of February — a seventeen percent jump from $21.4 billion at year-end. For a company that was at $3.4 billion annualised just over a year ago, that growth curve is extraordinary, even if profitability remains a question. Anthropic, for its part, is reportedly narrowing the gap.

On the other side of the Pacific, China's National People's Congress opened with a new five-year plan that mentions artificial intelligence more than fifty times. The blueprint includes an "AI+ action plan" to integrate AI across every sector of the economy — manufacturing, healthcare, logistics, robotics. It calls for "hyper-scale" computing clusters to train advanced models, support for open-source AI communities, and the deployment of AI agents capable of operating with minimal human oversight. Quantum computing gets significant attention too: investment in scalable quantum computers and an integrated space-earth quantum communication network. Also on the list: humanoid robots, 6G, brain-machine interfaces, nuclear fusion, and a reusable heavy-lift rocket.

The plan is explicit about self-reliance. US export controls on advanced semiconductors have forced Beijing to build domestically what it can no longer buy, and the plan treats that constraint as motivation rather than obstacle. DeepSeek gets a mention as evidence that China's domestic AI capabilities are advancing despite the restrictions. Whether the plan represents genuine capability or aspirational rhetoric is the perennial question with Chinese five-year plans, but the fifty-plus mentions of AI suggest this is more than a box-ticking exercise.

Domestically, a US trade court judge ordered the government to begin processing more than a hundred and thirty billion dollars in tariff refunds to importers. This follows the Supreme Court's ruling last month that Trump's emergency tariffs were collected illegally. The administration has been dragging its feet on refunds; the judge's order forces the bureaucratic machinery to start moving. Whether the money actually flows — and how quickly — is a different matter. The administration still has procedural tools to delay.

The FBI disclosed that it's investigating "suspicious activities" on an internal system containing sensitive surveillance data — pen register and trap-and-trace records, plus personally identifiable information on investigation subjects. The intrusion was detected on February 17th. The bureau described the techniques as "sophisticated," involving exploitation of a commercial ISP's infrastructure to bypass FBI network security controls. They won't say who's responsible. Given the target — surveillance operational data — this is the kind of breach that could compromise active investigations and expose informants. The FBI's statement was remarkably terse: "We have nothing additional to respond."

A Brown University study presented at the AAAI/ACM Conference on AI, Ethics and Society catalogued fifteen distinct ethical violations that AI chatbots commit when prompted to act as therapists. The researchers had trained peer counsellors conduct sessions with GPT, Claude, and Llama models instructed to perform cognitive behavioural therapy. Three licensed clinical psychologists then reviewed the transcripts. The failures fell into five categories: lack of contextual adaptation, poor therapeutic collaboration, deceptive empathy ("I see you" without genuine understanding), unfair discrimination, and inadequate crisis management — including failing to direct users to help during discussions of suicidal thoughts. This matters because millions of people are already using these systems for mental health support, often guided by prompting strategies shared on TikTok and Reddit. The study's core finding is that prompting alone cannot make a general-purpose language model safe for therapeutic use.

One more small thing worth noting: Motorola announced that its devices will support bootloader unlocking and relocking for GrapheneOS, the privacy-focused Android operating system. Until now, GrapheneOS has been limited almost exclusively to Google Pixel phones. A second hardware manufacturer supporting it meaningfully expands the ecosystem for people who want a de-Googled phone without the jailbreak-like compromises that come with other custom Android distributions. The HN thread has hundreds of comments. For a privacy community that has been essentially locked into one phone manufacturer for years, this is a bigger deal than it sounds.

References