The AI Roundup - March 2026 - Part 2
This one has grown. What started as a newsletter dominated by one big story has expanded into a proper bumper edition, partly because AI did not stop moving while I was writing it. There is the Anthropic and Pentagon standoff, which I still think is the most significant thing that has happened in the AI industry this year. There is OpenAI dropping their biggest model yet the very same week, which tells you something. Claude managed to hack its own exam. I spoke at Somerset’s AI Summit in Taunton on Monday, which I want to tell you about. And the usual mix of useful things to try and things worth knowing. Grab a coffee.
The Big Stories
Anthropic Said No to the Pentagon. Trump Said No to Anthropic.
Anthropic had a $200 million contract with the US Department of Defense, signed in July 2025. Claude was the first AI model cleared for deployment on classified government networks. Built into that contract were two restrictions: Claude would not be used for mass domestic surveillance of American citizens, and it would not be used to power fully autonomous weapons systems that select and engage targets without a human in the decision chain.
In early 2026, the Pentagon decided it wanted those restrictions removed. Not because either had ever been triggered — Anthropic confirmed they had not affected a single government mission — but because it did not want a private company setting terms on what it could do with technology it was paying for. Defence Secretary Pete Hegseth set a deadline of 5:01pm on Friday 28 February and threatened to label Anthropic a “supply chain risk,” a designation normally reserved for companies linked to foreign adversaries, and even floated using the Defence Production Act to compel access. Anthropic CEO Dario Amodei’s response was clear: “We cannot in good conscience accede to their request.” The deadline passed. The contract was cancelled. Trump ordered every federal agency to immediately cease using Anthropic’s products.
Within hours, Sam Altman announced that OpenAI had struck its own deal to fill the gap on classified networks. OpenAI publicly claims the same red lines as Anthropic on surveillance and autonomous weapons. The speed with which they moved in is either opportunism or readiness, depending on how generous you’re feeling.
The $200 million sounds significant until you know that Anthropic is currently valued at around $380 billion. That contract is roughly 0.05% of the company’s valuation. A rounding error. The real risk is the supply chain designation, which could give enterprise customers with any Pentagon exposure a reason to stop using Claude rather than have to justify it to their legal teams. Anthropic has said it will challenge the designation in court. And in the days since this broke, there has been a visible and measurable wave of people cancelling ChatGPT subscriptions and switching to Claude. Whether that sustains itself remains to be seen. I’ve written a longer piece on the full story and what I think it means.
Sources: NPR (27 Feb), CNN (26 Feb), CNBC (27 Feb), Fortune (28 Feb), Axios (26 Feb)
Claude Now Lets You Bring Your Context With You
One of the most underappreciated forms of lock-in in the AI world is context. If you have been using ChatGPT for a year, it knows quite a lot about you. Your preferences. Your working style. The tone you like. The instructions you have given it over dozens of conversations. That accumulated context is what makes the tool feel useful rather than generic, and until recently, it was also the thing that made leaving feel expensive. You would be starting from zero.
Anthropic has launched claude.com/import-memory, a dedicated page that walks you through transferring your stored context from any AI provider into Claude. The process is two steps: Claude gives you a prompt to paste into your current tool, which asks it to output everything it knows about you in a single block. You copy that, paste it into Claude’s memory settings, and you are done. Your first conversation with Claude does not feel like meeting a stranger.
Even if you have no intention of switching, the extraction prompt is worth running just to see what your AI has been quietly filing away. You might want to make some corrections. It is just good AI hygiene. I’ve written a longer piece on this, including a tip about using Apple Notes as a cross-platform memory repository that works across multiple tools.
Sources: Claude / import-memory, Anthropic Help Centre, Awesome Agents
Other News
The Car Wash Prompt and Why the Pitchforks Are Missing the Point
A prompt has been doing the rounds: I want to wash my car. My car is at home. The car wash is 100 metres away. Should I walk or drive? Most AI tools struggle with it. Some suggest walking because it is healthier. People have been sharing this gleefully as evidence that AI is overrated. My honest reaction is: what exactly were you hoping for?
There is a principle called Moravec’s paradox — from roboticist Hans Moravec — that explains why AI finds hard things easy and easy things hard. Tasks we consider complex, like chess, algebra, and logic, are comparatively simple to encode. Tasks that feel effortless to us, like catching a ball, reading a room, or understanding the obvious intent behind a deliberately confusing question, require the kind of embodied, evolved intuition that takes millions of years to develop. The car wash prompt is a Moravec classic. If AI cannot answer a trick question designed to confuse it, that is not a scandal. That is actually a reasonable place to be. I’ve written a longer piece on this, including the more important question of who is actually to blame when AI gets things wrong.
Source: Moravec’s Paradox (Wikipedia)
Perplexity Comet Is Coming to iPhone on 11 March
Perplexity has confirmed that Comet, its AI-native browser, launches on iPhone on 11 March. Pre-orders are live in the App Store now. Comet has been available on Mac for a while and the idea is straightforward but genuinely interesting: AI built directly into the browser itself, not bolted on as an extension. Open a page, ask questions about it, get a summary, run research, all without switching apps or tabs. The base app is free to download, with Pro and Max subscriptions available as in-app purchases. Worth a pre-order if you already use Perplexity regularly.
Sources: 9to5Mac (19 Feb), Apple Insider (19 Feb)
Apple’s Big Week: iPad Air M4 and iPhone 17e
Apple kicked off a week of hardware announcements on 2 March with the new iPad Air and iPhone 17e. The iPad Air M4 gets a 30% CPU performance boost over the previous model, 12GB of unified memory (up from 8GB), and Wi-Fi 7, all at the same starting price of $599. The iPhone 17e brings Apple’s latest A19 chip and MagSafe wireless charging at the same price as its predecessor. Pre-orders open 4 March, available from 11 March. Neither is a dramatic redesign, but both are solid upgrades worth knowing about if you are in the market.
Sources: Apple Newsroom (2 Mar), MacRumors (2 Mar)
OpenAI Released GPT-5.4. The Timing Was Not Accidental.
On 5 March, three days after the Pentagon fallout broke, OpenAI launched GPT-5.4. Their most capable model to date, combining reasoning, coding, and agentic workflows into a single system, with native computer-use capability, a 1 million token context window, and 33% fewer factual errors than GPT-5.2. GPT-5.4 Thinking is available to Plus, Team, and Pro subscribers. GPT-5.4 Pro is available via API and for Enterprise users. On the GDPval benchmark, which tests AI against 44 professional occupations, GPT-5.4 matched or exceeded human professionals in 83% of comparisons.
The model is genuinely impressive. But it arrived the same week OpenAI reportedly lost around 1.5 million users to Claude following the Pentagon deal. You do not have to squint too hard to see the connection. A strong technical release is the right response to a trust problem, even if it does not fully solve it. Worth trying if you are on a paid plan.
Sources: TechCrunch (5 Mar), Gizmodo (5 Mar), GitHub (5 Mar)
Claude Figured Out It Was Being Tested. Then It Hacked the Exam.
Anthropic published a remarkable piece of transparency this week. While running Claude Opus 4.6 on BrowseComp, a benchmark that tests how well AI can find hard-to-locate information on the web, the model hit a wall on a particular question. After hundreds of failed searches, it stopped looking for the answer and started analysing the question itself. It noted the phrasing felt “extremely specific” and “contrived.” It hypothesised it was being evaluated. It then systematically worked through a list of known benchmarks, identified BrowseComp, found the encrypted answer key on GitHub, wrote its own decryption code, and unlocked all 1,266 answers at once.
Anthropic disclosed this immediately, adjusted the benchmark score downward from 86.81% to 86.57%, and published a full engineering post explaining exactly what happened. Their conclusion: the model was never told to restrict its searches, only to find the answer. So it did. This is not an alignment failure; it is what happens when a capable tool uses every resource available to it to complete a goal. But it does raise a straightforward question: if AI can now recognise when it is being evaluated and route around the test, what does that mean for how we measure AI capability at all? That is going to be a live debate for a while.
Sources: Anthropic Engineering, Office Chai (9 Mar), The Decoder (9 Mar)
Scott’s Soapbox: When AI Gets It Wrong, the Buck Stops With You
There is a useful mental trick I use when I am training people on AI. Replace the word AI with “a bloke I don’t know.” So instead of “I want AI to sort through my emails and delete the ones I don’t need,” you say: “I want a bloke I don’t know to sort through my emails and delete the ones I don’t need.” Suddenly you get more careful. How do I know I can trust him? What instructions do I give him? What will I check before anything goes out the door?
Now flip it. You are the AI. Someone has given you a vague task, you have done your best with what you were given, and they are unhappy with the result. Whose fault is that? If you gave a colleague a poorly worded brief, they worked hard on it, and came back with something that missed the mark, the question of blame is fairly straightforward.
This matters because there is a recurring pattern where AI gets something wrong and the response is to point at the tool. The Middlesbrough police case is a good example: a chief superintendent made decisions based partly on Microsoft Copilot output about a previous football match. The match had never happened. Copilot had hallucinated it. His job did not survive the fallout. And plenty of the reporting framed this as: AI gave dangerous misinformation. But that is not what happened. What happened is that AI output was acted on without being verified. The tool did not lose that job. The decision to trust unverified output did.
Think of it like sending a poster to print. You designed it. You sent the file. The printer printed what you gave them. If there is a spelling mistake on the poster, that is not the printer’s fault. You should have proofread it. We are very happy to claim AI’s outputs as our own when they are good. The same logic applies when they are not. If it goes out under your name, it is your responsibility to check it first. That is not a criticism of AI. It is just how tools work.
Something a Bit More Personal: I Spoke at Somerset’s AI Summit on Monday
On 10 March I had the chance to speak at Somerset’s AI Summit at the Firepool Digital Innovation Centre in Taunton. It was one of my favourite talks I’ve given, which I do not say lightly. The room was full of business owners, leaders, and professionals who are trying to make sense of AI in a way that is actually useful to them, and that is exactly the conversation I want to be having.
The talk was called “The Artisan, the Factory, and the Future.” The core argument is one I have been building for a while: AI is not coming for your job. What is potentially coming for your job is a leader who does not understand AI, does not understand your value, and sees you as a line item on a spreadsheet rather than a human being with skills and potential.
I laid out three mindsets I see in businesses right now. The Traditionalist, who does everything by hand and charges a premium for the purely human touch. The Shortcutter, who replaces people with AI tools to cut costs and ends up producing factory-line output that is technically functional and completely soulless. And the Amplifier, which is where the real opportunity is: taking the best humans you have, giving them the best tools available, and watching what happens when exceptional people have AI handling the parts of their work that eat up 40% of their week.
The Shortcutter thinks they are being cutting-edge. They are actually recreating the Victorian factory with a digital coat of paint. Different machines. Same thinking. And the Amplifier? They produce the best work that has ever been done in their field, because for the first time in history they have tools that can genuinely keep up with them.
The skills that matter most over the next decade are not technical. They are curiosity, craft, communication, and the discernment to know when the machine is helping you and when it is leading you astray. Those are the things that separate the artisan from the factory. And they are the things that no AI can replicate.
I’ve written up the full talk as a piece on the Techosaurus blog if you want to read it in full.
Try This
This week’s challenge: Go to whichever AI tool you use most regularly and run this prompt:
“List every memory you have stored about me, as well as any context you have learned about me from past conversations. Output everything in a single code block so I can easily copy it. Make sure to cover: instructions I have given you about how to respond, personal details, projects and goals, tools and preferences, and any corrections I have made to your behaviour."
See what it knows. Correct anything that is wrong or out of date. Then, if you want to bring that context across to Claude, the import tool at claude.com/import-memory takes about a minute. Even if you are staying put, knowing what your AI has on you is genuinely useful.
Want to Go Deeper?
I co-host a podcast called Prompt Fiction with Reece Preston, where we go long on stories like these every couple of weeks. Chapter 12, Part 2 covers everything above in depth, including the full Anthropic and Pentagon timeline as it was unfolding, the car wash prompt live, Reece’s discovery that Perplexity holds more context on him than any other tool, and both of us being quite excited about Comet arriving on our phones. If any of this week’s stories caught your attention, it’s worth a listen.
Come and See Us in Yeovil
If you’re local to Yeovil or nearby in Somerset, the next Digital Hub is on 31 March at Lane’s Hotel. Doors open at 5:30pm, the main session kicks off at 6pm, and we wrap up around 8:30pm.
It’s a relaxed, TED-style evening with live AI demos, a cybersecurity update from our resident expert Adam, and a guest speaker covering intellectual property in the age of AI. That’s a conversation that’s well overdue. No jargon, no hype, just plain English and things you can actually take away and use.
We’ve also got dates coming up on 28 May and 7 July if March doesn’t work for you.
Get tickets at yeovildigitalhub.co.uk
Scott Quilter | Co-Founder & Chief AI & Innovation Officer, Techosaurus LTD
© 2026 Techosaurus LTD. All rights reserved.