Apple AI Extensions, Meta Brain Scans & Wikipedia AI Ban

Samantha: Welcome to The State of Tech — The European Edition. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Apple blows Siri wide open to rival AI assistants including Google Gemini and Anthropic Claude. Meta publishes a brain-scanning AI foundation model that can predict how your neurons fire. Wikipedia officially bans AI-generated article content to stop a poisonous feedback loop. SoftBank takes out a jaw-dropping forty billion dollar loan to pour into OpenAI. Huawei's latest AI chip wins over Alibaba and ByteDance in a major break from Nvidia dependence. And a new cybersecurity report reveals nation-state attacks on critical infrastructure are doubling while ransomware headlines fade. Let's get into it.

Apple admits it lost the AI race — and turns Siri into a switchboard.

Samantha: So let's start with Apple, because this is genuinely a strategic reversal. Bloomberg is reporting that Apple is building what it calls an Extensions system for Siri in iOS 27. The idea is that you, the user, go into your settings and pick whichever AI backend you want — ChatGPT, Google Gemini, or Anthropic Claude. Siri essentially becomes the routing layer, not the brain.

Bob: And that's the key shift here. For years Apple insisted it could build competitive AI in-house. Siri was supposed to be the answer. But behind the scenes, the gap just kept widening. ChatGPT, Gemini, Claude — they all leapfrogged Siri in reasoning, in conversation quality, in just about everything users actually care about.

Samantha: So instead of pretending Siri can compete head-to-head, Apple is doing what Apple does best — controlling the platform layer. You still talk to Siri, but the intelligence behind the curtain could be Google, could be Anthropic, could be OpenAI.

Bob: It's a very Apple move. They lose the model war but win the distribution war. Every AI company now wants to be the default option on a billion iPhones. That's enormous leverage.

Samantha: And here's what I find fascinating. Apple gets to observe which models users actually prefer. That data is incredibly valuable — for negotiations, for product decisions, for everything.

Bob: It's like being the landlord of the busiest shopping street. You don't need to run the best shop — you just collect rent from everyone who does.

Samantha: Now the European angle here is really interesting. The EU Digital Markets Act already requires gatekeepers like Apple to allow interoperability. So Apple may actually be getting ahead of enforcement rather than being dragged there kicking and screaming.

Bob: Which is smart. Brussels has shown it will enforce. Better to frame this as innovation than as compliance. But European regulators will still want to scrutinize how the default selection works, whether there's genuine neutrality, and how user data flows between Apple and these third-party AI providers. That's where the real friction will be.

Samantha: And for European AI companies — this could be an opening. If Apple is building a plug-in architecture, maybe Mistral or Aleph Alpha could eventually slot in too.

Bob: That's the optimistic reading. The pessimistic one is that the three American giants lock up those default slots and European alternatives never get a look in.

---

Meta trains an AI model on a thousand hours of brain scans — and it can now predict what your neurons will do.

Bob: Alright, next up — Meta has published something genuinely remarkable. It's called TRIBE v2, and it's a foundation model for neuroscience.

Samantha: Okay, unpack that for me, because "foundation model for neuroscience" sounds like science fiction.

Bob: So here's what they've done. They trained a trimodal AI model — meaning it processes video, audio, and language all at once — on over a thousand hours of fMRI brain scan data from seven hundred and twenty human subjects. The model learns the relationship between stimuli and brain responses. And then — this is the breakthrough — it can predict how a brain would respond to completely new stimuli it's never seen before.

Samantha: So you don't need a human in the scanner anymore?

Bob: For certain experiments, no. They're calling them "in silico experiments." You simulate the brain response computationally. It significantly outperforms classical linear encoding models that neuroscientists have relied on for years.

Samantha: The implications are huge. Think about drug development, think about understanding neurological disorders, think about how we study perception itself. If you can run thousands of simulated experiments instead of putting people in expensive MRI machines for hours — that changes the speed of discovery.

Bob: And Meta published this as open research. The model is open-source. So any university, any lab can build on it.

Samantha: Which is great news for European research institutions. Europe invested massively in the Human Brain Project — over a billion euros. The infrastructure and expertise from that initiative could combine beautifully with a tool like TRIBE v2.

Bob: The open-source angle also sidesteps some of the concerns about American tech companies monopolizing AI research. If anyone can use it, European labs are on equal footing. But there will be ethical questions — using brain data at scale, even for research, raises privacy issues that European regulators under GDPR will want to examine carefully.

---

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.

---

Wikipedia draws a hard line — no AI-generated text in articles. Period.

Samantha: So Wikipedia has officially banned AI-generated content in its articles. And the vote wasn't even close — forty-four in favour, two against.

Bob: This happened on March 20th on English Wikipedia after a formal Request for Comments process. The community decided that large language model output has no place in encyclopedia articles, with only two narrow exceptions — basic copyediting of your own text, and first-pass translation that still requires human review.

Samantha: And the reasoning is fascinating. It's not just about quality — it's about a feedback loop. AI models are trained on Wikipedia. If Wikipedia articles contain AI-generated hallucinations, those errors get baked into the next generation of AI training data. Which then produces more hallucinations. Which end up back on Wikipedia.

Bob: It's a snake eating its own tail. And Wikipedia, to its credit, recognized the systemic risk before most people even thought about it.

Samantha: The thing is, this only covers English Wikipedia so far. Other language editions haven't followed yet. And enforcement is going to be tricky — how do you reliably detect AI-generated text when the models keep getting better at mimicking human writing?

Bob: That's the practical challenge. The policy is clear, but policing it at scale across millions of edits requires tools that arguably don't exist yet with sufficient reliability.

Samantha: Now here's where Europe comes in. The EU AI Act includes requirements around training data transparency. If AI companies have to disclose what data they trained on, and that data includes Wikipedia, then the quality of Wikipedia directly impacts regulatory compliance.

Bob: It creates an interesting chain of accountability. European regulators could argue that if you're training on a source that's been contaminated by AI-generated misinformation, you haven't done proper due diligence on your training data. Wikipedia's ban actually helps the entire AI ecosystem maintain data integrity. It's a volunteer community doing what regulators are still trying to figure out.

---

SoftBank bets forty billion dollars that OpenAI is the future — with borrowed money.

Bob: Alright, let's talk about a number that's hard to wrap your head around. Forty billion dollars. That's the size of the bridge loan SoftBank just secured, primarily to fuel its investments in OpenAI.

Samantha: And this is an unsecured loan, Bob. Maturing in March 2027. Arranged by JPMorgan Chase, Goldman Sachs, and a consortium of other lenders. Masayoshi Son is essentially betting the house — again — on AI.

Bob: Son has a history of enormous, concentrated bets. Some worked brilliantly — Alibaba being the obvious example. Others were catastrophic — WeWork comes to mind. This OpenAI bet is somewhere in between so far, but the sheer scale of forty billion in borrowed money shows how high the stakes are in the generative AI race.

Samantha: And it tells you something about the capital intensity of leading AI development. OpenAI isn't cheap to run. The compute costs, the talent, the infrastructure — it all requires money at a scale that would make most industries blink.

Bob: The concern is what happens if generative AI doesn't monetize fast enough to justify these valuations. SoftBank has a year to refinance or repay. That's not a long runway.

Samantha: For Europe, this highlights the funding gap that everyone talks about but nobody seems to fix. European AI startups are raising tens of millions, maybe low hundreds of millions. Meanwhile, a single Asian conglomerate is borrowing forty billion for one investment. The playing field isn't level — it's not even the same sport.

Bob: And that has real consequences for talent retention. When OpenAI can offer compensation packages backed by this kind of capital, European AI labs struggle to compete for top researchers.

---

Huawei's AI chip cracks the private sector — Alibaba and ByteDance are placing orders.

Samantha: Let's move to Huawei, because this is a significant development in the chip wars. Huawei's latest AI processor — the 950PR — has reportedly been successfully tested by Alibaba and ByteDance, and both are preparing to place orders.

Bob: Until now, Huawei's AI chips were mainly adopted by state-backed entities. Getting Alibaba and ByteDance on board is a completely different signal. These are commercially driven companies that care about performance and cost. If the 950PR meets their standards, it means Huawei has genuinely closed part of the gap with Nvidia.

Samantha: And the key detail is CUDA compatibility. Nvidia's dominance isn't just about hardware — it's about the software ecosystem. Developers build on CUDA. If Huawei can offer reasonable compatibility with that ecosystem, the switching cost drops dramatically.

Bob: This is the scenario that US export controls were designed to prevent. The sanctions pushed Huawei to accelerate its own chip development, and now we're seeing the results. It's not parity with Nvidia yet, but it's progress that matters.

Samantha: For Europe, this is a wake-up call wrapped in a reminder. The EU Chips Act committed forty-three billion euros to boost European semiconductor capacity. But that money is mostly going to traditional chips, not cutting-edge AI accelerators. If both the US and China are developing independent AI hardware ecosystems, Europe risks being dependent on both — or neither — for the chips that actually power AI.

Bob: And that dependency question becomes very concrete when European companies need to choose which AI infrastructure to build on.

---

Ransomware is fading from the headlines — but something more dangerous is taking its place.

Samantha: Last story today — and it's a sobering one. The Waterfall Threat Report for 2026 shows that cyber breaches with physical consequences in critical infrastructure dropped twenty-five percent last year. Sounds like good news, right?

Bob: On the surface, yes. But dig deeper and the picture is alarming. Nation-state and hacktivist attacks on critical infrastructure doubled. The ransomware decline is largely temporary — driven by law enforcement disruptions and gang restructuring. The underlying threat is shifting from financially motivated criminals to politically motivated state actors.

Samantha: And these aren't just data breaches. We're talking about attacks that cause physical damage — production shutdowns in aerospace, disruptions in maritime operations. Real-world consequences from digital attacks.

Bob: The report also flags that incident reporting is becoming less detailed. Companies and governments are sharing less information about how attacks translate into physical damage. That's dangerous because it means the broader security community can't learn from these incidents.

Samantha: For Europe, this is directly relevant. Energy grids, transport networks, hospitals — all of these are targets. The NIS2 Directive is supposed to strengthen cybersecurity across critical sectors, but implementation is uneven across member states. And the operational technology — the actual systems running factories and power plants — often lags years behind in security updates.

Bob: The doubling of nation-state attacks should be a board-level conversation at every European utility, transport operator, and healthcare provider. This isn't theoretical risk anymore. It's the new normal.

---

Samantha: Alright, that's our show. Today we covered: Apple opening Siri to rival AI models, Meta's brain-scanning foundation model TRIBE v2, Wikipedia banning AI-generated content, SoftBank's forty billion dollar bet on OpenAI, Huawei's AI chip winning Chinese tech giants, and the shift to nation-state cyber attacks on critical infrastructure.

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.