WWDC26 Special: Apple Presents New AI-Siri and Despairs Over Europe

WWDC26 (1): Apple Shows Off Its Siri AI Upgrade

At WWDC26, after years of being a running gag, Apple has given Siri a complete overhaul: richer answers, natural conversations, and a standalone app. Additionally, there's smarter photo editing and a refreshed Image Playground. Meanwhile, the rest of the industry is showing where things are headed: NotebookLM now runs on Gemini 3.5 and writes its own code with over 100+ skills, every note gets a secure cloud computer that returns PDFs, spreadsheets, and slide decks. Amazon's Alexa for Shopping turns an idea into printed hoodies and mugs via merch-on-demand, shipped with Prime eligibility. Anthropic preaches Zero Trust for agents and treats every agent as already compromised. ChatGPT is getting a lockdown mode that restricts browsing, downloads, and agent actions to block prompt-injection data theft. And Mira Murati's Thinking Machines is betting that the real frontier isn't pure autonomy, but AI that makes people think better. → TAAFT - There's An AI For That

Synthszr Take: It's the third or fourth season of the same trick: Apple takes something invented elsewhere and makes it radically more convenient. Back in June 2024, I wrote that Cupertino's polished interfaces would once again be the toughest door in the world, with Google and OpenAI jostling for entry. That's exactly what we're seeing now: ChatGPT is happening behind the scenes at Apple, interchangeable, replaceable. The moat isn't the best model, but the one billion users who have their Apple ID welded to the assistant. Who's going to part with a free, 24/7 assistant that already knows everything personal? The interesting part will be whether Apple can avoid making beta software feel like alpha again, because that was precisely the weakness in 2024. If Siri delivers, this season will print money again.

WWDC26 (2): Apple Despairs Over Europe

Apple will not ship the new Siri AI features with iOS 27 and iPadOS 27 in the EU later this year. Craig Federighi, the company's software chief, was 'deeply disappointed' and no longer provided a timeline for availability. Affected are the new app for rediscovering conversations, the enhanced Visual Intelligence, the integrated writing tools, and the Siri mode in the Camera app. Apple had proposed a Trusted System Agent, an intermediary that would securely connect third-party assistants to the same functions as Siri, as well as an 18-month plan for a gradual rollout. The EU Commission rejected every single proposal. EU developers are also not allowed to test or integrate the new Siri AI features. However, Siri AI is coming to macOS 27, visionOS 27, and watchOS 27, while the launch in China is also delayed for regulatory reasons. → www.macrumors.com

Synthszr Take: In June 2024, I already described this as a real dilemma, and two years later, it remains just that. The DMA essentially requires that every AI system gets almost unlimited device access, meaning reading and sending messages, making purchases, opening files, and acting across apps. This very access is the attack vector: security researchers have long shown how such agents can be hijacked to steal passwords and photos. Apple is right on this matter, and the Trusted System Agent sounds like an honest attempt to unify interoperability and data security. The fact that Brussels didn't agree to a single proposal is the truly unsettling signal, because in the end, 450 million EU citizens lose out on features that work without issue on Apple's Mac and Watch. We need more Europe, not less, but then the negotiators must actually negotiate instead of just saying no. Anyone who legislates for the tightest possible data protection while simultaneously demanding forced full access for all creates Kafkaesque conditions in which no one wins in the end.

WWDC26 (3): Apple Builds Its AI Foundation on Google's Gemini

Apple today announced a fundamental overhaul of Apple Intelligence: at its core are new Apple Foundation Models, developed jointly with Google based on Gemini technologies. The models run both on-device and through Apple's existing Private Cloud Compute infrastructure, and Apple explicitly calls the collaboration 'deep'. New additions include reasoning, multimodality with image understanding and generation, realistic image creation, advanced photo editing, and visual question answering. Certain devices will get a more powerful variant with language generation, better dictation accuracy, and stronger language understanding—which ones exactly, Apple isn't saying yet. A new System Orchestrator coordinates features across all platforms and adapts responses to the active app and the specific task. Apple distinguishes itself from competitors who 'rush ahead' without regard for users, emphasizing that data is only used for the immediate request and is not accessible to Apple or third parties, verifiable by external experts. → Hacker News

Synthszr Take: Just under two years ago, Tim Cook paused Apple Intelligence in the EU, pitting privacy against openness. Now, Apple is building its entire AI foundation on Gemini technology and selling the dependency as a 'deep collaboration'. This is the most honest self-assessment Apple has provided in years: they've fallen behind in their own model development and are now buying in from Google, the same company that already pays them 20 billion a year for the default search. The System Orchestrator is the interesting part, as it's Apple's real moat. Whoever knows the context of the active app and the current task controls the intent, and intent beats raw model size. The privacy narrative remains the strongest selling point, it just sounds thinner when the brain behind it comes from Mountain View. Apple has decided it's better to deliver a good product with third-party models than a mediocre one with its own, and that's the right decision.

WWDC26 (4): Apple Lets AI Change Your Password if You've Been Hacked

At WWDC 2026, Apple introduced an extension to its password feature: Apple Intelligence not only detects weak or compromised passwords, the AI can now change them itself. To do this, the agent navigates the website, logs in, and creates a new password that is saved directly in the Passwords app. Previously, you had to access each account individually and change the credentials manually, which quickly became an ordeal with two-factor authentication enabled. While Google's Password Manager offers weak password detection, Gemini cannot change them on its own. Gizmodo calls the feature an outlier because 'passwords' and 'AI agent' in the same sentence would otherwise set off every security alarm bell. Tools like OpenClaw have been repeatedly described as a security nightmare. → gizmodo.com

Synthszr Take: This is a small feature with a big message: Apple dares to go where others won't. An agent that autonomously logs into third-party accounts and changes passwords is exactly the scenario where Replit deleted its production database in July 2025 (explicitly forbidden, yet it happened). The difference is the guardrails: Apple links the whole thing to Passkey logic, local storage, and a narrowly defined intent, instead of opening up the entire world to the agent. This is what useful agentic AI looks like: small in scope, high in everyday value, comparable to the moment iOS started auto-filling confirmation codes. Those with security as a core competency can afford such autonomy because the foundation of trust already exists. Most users have terrible password hygiene, and a tool that closes the gap without asking them is worth more than any security tutorial that no one reads anyway. This is precisely the difference between hacked and not hacked, and it's something you can turn on tomorrow morning.

WWDC26 (5): Safari Gets AI Features to Organize Tabs and Monitor Websites

Apple is bringing Apple Intelligence to its most-used apps: Safari, Passwords, Messages, Mail, Calendar, and Phone. The biggest chunk is in the browser. Safari analyzes open tabs, groups them by topic, and automatically adds new relevant tabs; at the end, you can close an entire topic or save it as a group. With 'Notify Me,' you describe in natural language what change you're waiting for, such as product availability or event registrations, and Safari will notify you when it happens. 'Describe an Extension' generates a custom browser extension from a text description that modifies websites as you wish. In the Passwords manager, Apple Intelligence autonomously navigates to supported websites, logs in, and replaces weak passwords with stronger ones. Apple emphasizes that everything runs on-device or via Private Cloud Compute, including 'Call Context' in the Phone app, which pulls up your flight confirmation from Mail before a call to an airline.

Synthszr Take: Apple has always had an ambivalent relationship with the open web; Steve Jobs preferred native apps. Now Cupertino is sending the browser itself out to navigate websites, log in, and change passwords in the background. This is precisely the step we described in May as the beginning of the Dark Web: the web is becoming content logistics, through which fewer and fewer people actually navigate themselves. 'Notify Me' and automatically sorted tabs sound like small conveniences, but the direction is clear. The browser is ceasing to be a window you look through and is becoming an agent that clicks for you. Anyone building websites should start thinking about a machine audience tomorrow morning instead of a human one, because the day after tomorrow's traffic no longer has eyes. It's convenient. And that's exactly why it will prevail.

OpenAI is Rebuilding ChatGPT into a Superapp and Declares Chat is Dead

According to the Financial Times, OpenAI is preparing the biggest overhaul of ChatGPT since its launch, transforming the chatbot into a so-called superapp that will be rolled out via the website and mobile app in the coming weeks. The new version relies heavily on the Codex assistant and pushes users towards autonomous agents, coding, image generation, and third-party apps. Thibault Sottiaux, who leads the core product and platform teams at OpenAI, says every user will get a personal agent, accessible via mobile, desktop, web, or by voice in the car. Internally, an employee puts it more drastically: 'Chat is dead.' The message behind this: the text dialogue was just the introduction; the product is the acting agent. In March, OpenAI fumbled its e-commerce launch with the checkout feature; now the company is betting everything on the agent card. → Techpresso

Synthszr Take: The chat was never the product; it was the onboarding trick that brought OpenAI to hundreds of millions of users in three years. The text box handled the distribution; now the agent is supposed to provide the retention. Strategically, the superapp is an attempt to move away from the commoditized chatbot before Gemini and open-source models drive the price of pure answers to zero. It's interesting that Codex is at the center, not the next smart small talk: whoever owns the coding and agent layer sits on the user's operational process, not just their curiosity. The risk is overextension. WeChat works as a superapp because it's a platform with payments and identity, and that's precisely where OpenAI, with its March checkout disaster, showed that merchant acceptance can't just be willed into existence. Anyone building an AI strategy today should ask themselves tomorrow morning whether they are still integrating a chat UI or betting directly on agentic workflows with clear cost telemetry. The bet is right; the execution is everything.

Notion Custom Agents

Notion has published a use-case guide for its Custom Agents, tackling the issue every team knows: repetitive work that doesn't just go away on its own. The guide shows three use cases. First, instantly answering standard questions based on the knowledge already in the workspace (sales reps waiting hours for answers about product launches, IT resetting passwords for the hundredth time). Second, automatically routing tasks according to self-defined rules, instead of smart people spending their mornings sorting bug reports and tickets. Third, automatically compiling status reports so that team leads no longer have to gather information from Slack, trackers, and docs. Notion is positioning Custom Agents as a replacement for the manual distribution work that a system should have been handling long ago. → Tech Brew

Synthszr Take: In February, I wrote here that Notion is the new SAP. This guide shows why. The three use cases are not gimmicks, but the most boring, most profitable layer of agents imaginable: triage, routing, reporting. Exactly what the triage agent does in my customer service example, except Notion builds it into the software where the knowledge base already resides, without any MCP tinkering. The trick is the data context, not the model. Anyone who maps their processes in Notion gives the agent the context for free, and the rest qualifies itself. You can set up the first routing agent this week, not after the next AI strategy offsite. The only question is whether you'll continue to expect your smartest people to spend their mornings acting as a distribution hub.

Anthropic Commits to $200 Billion in Cloud Spending with Google

Anthropic plans to spend around $200 billion over five years with Google for cloud capacity and chips. The deal begins next year and specifies an agreement Google announced last month: the delivery of five gigawatts of server capacity to Anthropic, which at the time had no price tag. Now the number is out. Anthropic is thus betting heavily on Google's TPU infrastructure, while Amazon, its previous main investor and cloud partner via AWS, remains in the background. Researched by Sri Muppidi, Erin Woo, and Amir Efrati. Anthropic's valuation was still at $965 billion at the beginning of June, surpassing OpenAI. → The Information

Synthszr Take: $200 billion over five years—that's $40 billion per year for computing capacity from a single provider. Anthropic preaches vendor neutrality and open-source migration paths, but is now shackling itself to Google's TPUs on a scale that makes any re-architecture in the risk register purely theoretical. This is industrial-scale lock-in, signed voluntarily. Anyone sourcing five gigawatts from a single supplier has no backup capable of ever absorbing that number. The role of Amazon is interesting: the largest investor watches as its protégé carries billions in compute spending to a direct AWS competitor. Google is building the perfect moat here, because every token Claude generates will run through Mountain View in the future. Compute discipline looks different, and the question of who collects the margins in the end has now been answered.

Gemini 3.5 Turns NotebookLM into a Research Computer

Google is upgrading NotebookLM to the new Gemini 3.5 and, for the first time, adding Antigravity, the coding environment that now gives NotebookLM its own 'Cloud Computer'. This allows the tool to write and execute code, with over 100 software skills that map out workflows that previously required jumping between multiple apps. In Google's own tests against the old 3.1 version, the new variant wins across five evaluation dimensions (accuracy, multilingualism, analysis of large documents, document creation, advanced research) with an average win rate of 65 percent. NotebookLM also goes beyond plain text and now generates PDFs, slides, Excel files, CSV/JSON, and images via Nano Banana, all of which are editable afterward. Also new is that the app finds sources on its own: you ask for relevant websites in the chat, get a 'Research Report,' and import the sources directly. Gemini 3.5 Flash was already announced at Google I/O as being significantly cheaper in token costs, with the same or better quality. The rollout starts today, first for AI Ultra subscribers and Workspace Business customers. → Ars Technica

Synthszr Take: NotebookLM was one of Google's first GenAI experiments in 2023, and in a very un-Googly manner, it's still alive. Now it's clear why it wasn't shut down: it's the product where Google marries source grounding with code execution. Its old strength was that NotebookLM only works with your documents, no free-form hallucination. With Antigravity and 100+ skills, it becomes a work environment that translates your research into finished slides, spreadsheets, and graphics without you having to switch apps. This is exactly what I meant in May when multimodality turned apps into true everyday companions. The fact that NotebookLM now finds and imports the sources for you is the point at which the open web finally becomes the backend: you no longer read ten tabs, you read the Research Report. Anyone who works with documents today should open this thing up this week and run a real workflow through it, not wait for the wide rollout.

China's Unitree Will Dominate the Global Robotics Market

SemiAnalysis paints a picture of a new Chinese hardware giant. Three years ago, Unitree was building four-legged robots; today, the company leads the humanoid market and will soon ship its 10,000th G1, while Tesla continues to present its Optimus, introduced in 2022, as a work-in-progress. Unitree triples its revenue year after year with a 60 percent gross margin, plans around $300 million in AI research spending, and is heading towards a highly anticipated IPO. The price of its flagship G1 has dropped from over $50,000 to $27,300 within 12 to 18 months, with some deals already going for under $20,000, and SemiAnalysis estimates a gross margin of 67 percent even at that price. The analysts see Unitree copying the BYD and DJI strategy: master the most expensive component themselves (in this case, the QDD actuator), build an insurmountable cost advantage from it, and then devour the markets they themselves created. Over 250 units are already in real-world work environments, and it's precisely at this threshold of economic viability that the real acceleration begins, according to SemiAnalysis. → SemiAnalysis

Synthszr Take: The crucial sentence is in the BYD comparison, not in the robot specs. BYD started in 1994 with the battery cell, the most expensive part of a car, and over two decades built a cost advantage from it that no Western manufacturer can catch up to today. Unitree is doing the same with the actuator, and a G1 under $20,000 with a 67 percent margin isn't a toy price—it's a statement to anyone who thinks the West still has a lead in humanoids. In early May, we wrote about China's cheap tokens, and in the same month, about how AI and robotics are reaching the job market; this is the hardware side of the same movement, and it's scaling faster than the software side suspects. Anyone in Europe still waiting for the perfect, certified prototype is losing the market before it even begins to exist, because Unitree iterates faster and cheaper than all competitors combined. The response to this can be decided tomorrow morning: order 50 units from a Chinese supplier and find real use cases in your own operations, instead of waiting three years for a domestic champion that may never materialize.