US-China AI Safety Talks: What Europe Loses

Samantha: Welcome to The State of Tech, The European Edition, Saturday May sixteen, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Washington and Beijing open formal AI safety talks after the Trump-Xi summit, the OpenAI and Apple alliance heads toward a possible courtroom, UK regulators sound the alarm on AI-powered cyber risks in finance, Samsung strike fears rattle the global chip supply, and to close out, two things you can actually try this weekend. Let's get into it.

US and China launch formal AI safety talks after Trump-Xi summit in Beijing.

Samantha: This is the big one. After the Trump-Xi summit in Beijing on Thursday, the United States and China have agreed to open formal talks on AI safety. US Treasury Secretary Scott Bessent announced it on CNBC, and he called the two countries, quote, the two AI superpowers. The goal: a shared framework of best practices and safeguards to keep the most powerful AI models out of the hands of terror groups and cybercriminals.

Bob: And the scale of this is hard to overstate. This is the first time the two countries that dominate AI are sitting down at one table on safety. The comparison people are reaching for is Cold War-era nuclear talks. For decades, Washington and Moscow built a vocabulary around what nuclear restraint looks like. There is no such vocabulary yet for frontier AI. So if Bessent and his Chinese counterparts can agree on even basic guardrails, on who gets access to the most powerful models, that becomes the template for everyone else.

Samantha: Bessent was clear that he believes the US is negotiating from a position of strength. He says America still leads in AI development. That framing matters, because it tells you how Washington wants to enter the room: not as equals, but as the senior partner setting terms.

Bob: In practice, guardrails here means things like denying frontier models to non-state actors, controlling how the most capable systems get distributed, and probably some shared monitoring on misuse. It is not about banning research. It is about who can download the most dangerous capabilities.

Samantha: And for European listeners, the awkward truth is that Europe has no seat at this table. Brussels has the EU AI Act, which is the world's most detailed rulebook, focused on risk categories and citizens' rights. But the AI Act regulates what happens inside Europe. It does not decide which frontier models the US allows China to access, or vice versa.

Bob: Right. So you get a split. Europe sets the rules for deployment, Washington and Beijing set the rules for proliferation. Over time, agreements between those two will shape which models, which chips, and which applications even reach European companies. That is a strategic gap, and you can already hear European officials starting to push for some form of trilateral structure. Whether that actually happens is another question.

Samantha: One to watch closely.

OpenAI weighs legal action against Apple as the ChatGPT-in-Siri deal unravels.

Bob: From geopolitics to a corporate breakup. Bloomberg reports that OpenAI has hired a law firm to explore legal action against Apple, possibly including a formal breach-of-contract notice. The trigger is the 2024 deal that lets Siri pass complex questions to ChatGPT as part of Apple Intelligence.

Samantha: And OpenAI thought that deal would be a goldmine. They expected it to drive billions in extra paid ChatGPT subscriptions, hundreds of millions of iPhone users discovering ChatGPT through Siri and converting to paying customers. The reality, according to OpenAI's own data, is that users still prefer opening the standalone ChatGPT app. The Apple integration is too limited.

Bob: And then comes the kicker. Apple is reportedly planning to open Siri up in iOS 27, which is likely to be unveiled at the Worldwide Developers Conference on June 8th, to rival providers. That means Anthropic's Claude and Google's Gemini sitting next to ChatGPT inside Siri. OpenAI's exclusive position evaporates.

Samantha: Meanwhile, Apple is reportedly furious. OpenAI has been poaching from Apple's hardware teams, and the acquisition of Jony Ive's design studio signals OpenAI is moving into hardware. Ive used to design the iPhone. Now he is potentially helping build a rival device.

Bob: For iPhone users in Europe, this matters directly. If iOS 27 lets you pick your AI assistant, the way you can pick a default browser, suddenly the question is not whether you get ChatGPT, but which assistant you choose. That is a real consumer shift.

Samantha: And it shows a broader pattern. The first wave of AI platform deals, from 2024, is already cracking. OpenAI's relationship with Microsoft is strained. Now Apple too. These partnerships looked permanent two years ago, and they are turning out to be fragile.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

UK financial watchdogs warn AI is supercharging cyberattacks on banks.

Bob: The UK's finance ministry, the Bank of England, and the Financial Conduct Authority have issued a joint warning. British companies need to act now on cyber risks amplified by advanced AI. Their language is striking. They say current frontier AI models have cyber capabilities that exceed human experts in speed, scale, and cost.

Samantha: Read that carefully. They are not saying AI will one day match human hackers. They are saying it already outperforms them on the three things that matter most: how fast, how broadly, and how cheaply you can run an attack. That changes the threat model for every bank, every payment processor, every market infrastructure.

Bob: And they singled out concerns about products like Anthropic's Mythos, which regulators see as capable of supercharging complex attacks against the banking industry. The warning is about financial stability, data integrity, and market trust. If a major UK bank gets compromised at AI speed, the consequences ripple.

Samantha: For European banks, this is a preview. The UK is no longer in the EU, but the European Central Bank and national regulators in Frankfurt, Paris, and Amsterdam read the same threat reports. Expect similar warnings, and probably stricter supervisory expectations, in the coming months.

Bob: There is also a bitter irony here. The same AI that banks are deploying to detect fraud is now being used by attackers to design exploits. So you get an arms race inside the financial system, where defensive AI and offensive AI are pulling from the same underlying technology.

Samantha: And we are seeing this beyond banking. TRM Labs reports North Korea-linked groups are using AI to target the 130 billion dollar crypto sector, compressing vulnerability discovery from months to hours. Same pattern, different industry.

Samsung strike fears rattle chip customers from Apple to HP.

Samantha: A labour story with global consequences. Samsung Electronics workers are scheduled to strike starting May 21st. Major customers, including Apple and HP, are openly worried about disruption to the memory chip supply.

Bob: Samsung is already adjusting. They are shifting production toward high-bandwidth memory for AI data centres, which is the most lucrative and most contested category right now. That tells you where the priorities lie: AI infrastructure first, smartphones and PCs second.

Samantha: And here is the strategic risk for Samsung. Customers who get spooked by a strike start diversifying their suppliers. Once Apple or HP move part of their memory contracts to SK Hynix or Micron, they do not necessarily move back. So a strike that lasts two weeks can reshape supply relationships for years.

Bob: For Europe, the exposure is real. European cloud providers and AI infrastructure builders depend on the same memory chips. If supply tightens, prices rise, and European AI projects, which are already chasing American hyperscalers, get squeezed on margin.

Samantha: There is also a broader signal here. Applied Materials, the US semiconductor equipment giant, just raised its sales forecast and said AI demand is a multi-year driver. Gross margins at a 25-year high. The chip industry is booming, and any disruption, from a strike to a geopolitical shock, lands on infrastructure that the entire AI boom rests on.

Bob: Watch May 21st. If the strike goes ahead and lasts more than a few days, expect memory chip prices to move, and expect European procurement officers to start making phone calls.

Vivaldi browser ships a free distraction-blocker built right into your tabs.

Bob: Time for two things you can actually try this weekend. First one is for anyone who spends too much time getting pulled into rabbit holes online. The Vivaldi browser, the one from the former Opera team based in Norway, has rolled out a built-in focus mode this week. It is free, it works on Windows, Mac, and Linux, and you do not need an extension.

Samantha: The idea is straightforward. You set a list of sites you want to block, news sites, social media, shopping, whatever your weakness is, and then you switch on focus mode for a set period, say 90 minutes. Vivaldi greys out those sites and shows you a gentle reminder instead of the page. You can override it, but the friction is the point.

Bob: What makes this nice compared to the dozens of focus extensions out there is that it is native. No extra account, no subscription, no data leaving your machine. Vivaldi has always pitched itself as the privacy-friendly browser, and this fits.

Samantha: To try it, download Vivaldi from vivaldi.com, install it, and look in the settings under productivity. The whole setup takes maybe ten minutes. You can keep using Chrome or Safari for your normal browsing and just open Vivaldi when you need to actually get work done. Treating it as a dedicated focus browser is a nice workflow.

Bob: And if you already use Vivaldi, just update to the latest version this weekend. The feature is live now.

Samantha: One small caveat. Focus mode does not block apps on your phone. So if you really want deep work, leave the phone in another room. The browser cannot save you from yourself entirely.

A free prompt trick that makes Claude and ChatGPT give you sharper answers.

Samantha: Second one, and this is the AI find of the day. A simple prompt technique that has been making the rounds among researchers and that genuinely improves the answers you get from Claude or ChatGPT. It is called the critique-and-revise prompt.

Bob: Here is how it works. You ask your question as normal. Let's say you ask Claude to draft an email or summarise a document. Once you get the answer, you reply with one line: now critique your own answer as if you were a sceptical expert, then rewrite it.

Samantha: That is it. No paid plan, no special tool. You are just asking the model to grade its own work and try again. And the second version is almost always noticeably better, sharper arguments, fewer fluff sentences, better structure.

Bob: It works because these models are trained to be helpful, which sometimes means agreeable and verbose on the first try. When you ask them to put on a critical hat, they suddenly catch their own weak spots. You can do this on the free tier of ChatGPT or Claude, and it takes maybe 30 seconds extra per query.

Samantha: Try it on something you actually use AI for this week. A cover letter, a tricky email, a meeting summary. Run it through once, then say critique and revise. Compare the two versions. Most people are genuinely surprised by the difference.

Bob: And no, this is not a hack that will stop working tomorrow. It uses something fundamental about how these models reason when prompted to self-check. A small habit that pays off every day.

Samantha: Today we covered: the US-China AI safety talks, OpenAI heading for a legal fight with Apple, UK regulators warning on AI cyber risks, Samsung strike fears hitting the chip supply, Vivaldi's new focus mode, and a critique-and-revise prompt trick for Claude and ChatGPT.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.