8 min read · 14 sources

Tim Cook is out as Apple CEO after 15 years. His replacement is John Ternus, the company’s hardware engineering boss since 2021, and Cook himself stays on as executive chairman with a remit that includes working with governments. In a farewell memo to staff on his last day, Cook said he “will miss this work” but has “enormous comfort” in Ternus. The handover closes a tenure that took Apple from the post-Jobs crisis through the Apple Watch, AirPods, and several generations of iPhone, iPad, and Mac.

The bigger question for engineers isn’t who Cook was but who Ternus is. He runs hardware. The last decade of Apple was defined by services and silicon; the next one looks like it’ll be defined by the box. Read it with that lens.

Anthropic has signed a $35 billion cloud-computing deal with Lambda, which is backed by Nvidia, with Nvidia itself holding the lease on the data center.

FTC hits Amazon's ad business

Source: wsj.com ↗

The FTC has sued Amazon, alleging the company secretly manipulated the floor prices businesses have to pay to advertise on Amazon.com. Amazon’s ad business is the third-largest in the world behind Google and Meta, and earned $68 billion in 2025, so this is not a periphery fight. The suit argues that advertisers were shown one price and quietly billed more.

For sellers and agencies buying Sponsored Products, this is the case to watch. Discovery scope, refund mechanics, and ad-platform disclosure rules are all on the table. For anyone running P&L on Amazon ads, the practical advice hasn’t changed yet, but document everything.

Anthropic's $35B Nvidia-backed compute deal

Source: wsj.com ↗

Anthropic signed a cloud deal worth $35 billion with Lambda, the GPU cloud provider backed by Nvidia. The wrinkle is the lease: Nvidia itself holds the lease on the underlying data center, so the chain is Anthropic → Lambda → Nvidia-owned facility. That makes this less “Anthropic buys GPUs” and more “Nvidia underwrites Anthropic’s compute in a way that locks capacity to a specific buyer.”

For model buyers, it confirms what you’ve been seeing in quota pages: top-tier inference is becoming a capacity-constrained good, and the people who own the data centers are the ones shaping who gets served.

uBlock Origin is gone from the Chrome Web Store

Source: webiterate.dev ↗

Google has finally removed the last Manifest V2 extensions from the Chrome Web Store. uBlock Origin is the headline casualty, but the cull covers everything still riding MV2. Extensions already installed on Chrome 138 or earlier keep running; they just can’t get updates and can’t be reinstalled once you uninstall them. The deprecation also drags down Chromium-based browsers like Brave that depended on the Web Store for discovery; Brave is self-hosting four of the popular MV2 blockers to keep them alive.

If you operate a fleet with strict extension policies, audit the Manifest V2 list now. Anything still on MV2 will rot in place, and the next Chrome ESR bump will be when it visibly breaks.

Cook is out, but the Apple v. OpenAI lawsuit grinds on

Source: 9to5mac.com ↗

In the same week Cook exits, the Apple v. OpenAI trade secrets case keeps producing colour. Apple told the court it found “shocking evidence” on the MacBook of former engineer Chang Liu: a confidential Apple circuit schematic synced from a Mac mini via iCloud, an LTspice simulation run in March through Liu’s AI “agent,” and instructions Liu allegedly sent to an OpenAI colleague about destroying evidence once Apple’s investigation started.

The interesting legal theory is in the last paragraph of Apple’s filing. They argue that feeding trade secrets into learning AI models makes misuse “irreversible and continually propagating.” If that lands, the discovery scope in any future AI-poaching case gets very wide, very fast.

OpenClaw 2.0 ships, after a seven-week pause

Source: openclaw.ai ↗

OpenClaw 2.0 is out, and the release notes read like a census: 933 contributors, 569 of them first-timers, 16,000+ pull requests merged, roughly half of every PR ever landed in the project. The release took nearly seven weeks, which the team blames on the same growth that made the PR count possible.

What’s actually changed: first-run install now auto-detects existing ChatGPT or Claude subscriptions, API keys in the environment, or local models and picks one. The browser app was rebuilt as a first-class surface instead of a shim. Messaging, memory, skills, model routing, automations, plugins, and the auth model all got touched. For anyone using OpenClaw in production, the install simplification alone is worth a re-test, and the security section of the notes deserves a real read.

A Bengaluru startup wants to soar for a year, on ocean wind shear](https://techcrunch.com/2026/08/31/lachy-groom-backs-indian-startup-aiming-to-keep-aircraft-aloft-for-a-year/?utmsource=tldrnewsletter)

Alteon, a Bengaluru startup founded by 20-year-old Samay Sanghvi, raised a $2.5M pre-seed led by Lachy Groom to build autonomous fixed-wing aircraft that harvest energy from wind shear via dynamic soaring, the same trick albatrosses use to cross oceans without flapping. The first airframe is tiny: a 3-meter wingspan, flying within a metre of the water. A recent Bay of Bengal test put the aircraft through seven O-shaped cycles autonomously at over 62 mph, but energy-neutral dynamic soaring, propulsion-off sustained flight, is still unproven.

The team is cranking out four to five airframes a week and has logged 200+ flights in the last 30 days. The first paid use case is maritime surveillance. For anyone in autonomous systems or robotics, the data-collection pace is the story; the airframe is the easy part.

Matic is in 10,000 homes, here's how

Source: tanayj.com ↗

A long interview with Matic co-founder Mehul Nariyawala breaks down how a vacuum-and-mop home robot made it to 10,000+ US households. Three lessons that are useful outside consumer robotics: enter a market that’s already tedious (robot vacuums are at 15% US penetration, with iRobot at $700M revenue), absorb complexity in software rather than punting it to the user, and treat deployment data as a moat. The company is now also shipping voice (“Hey Matic, clean the kitchen”) and gesture control, both of which took nearly eight years from internal demo to ship.

AI Overviews expand to fill the whole results page

Source: 9to5google.com ↗

Google is testing dynamically expanded AI Overviews that take over the entire results page for some queries (definitions are a consistent example), rather than appearing only at the top. Google confirmed the test and said users find the full-page experience more helpful and follow up more. For SEO, publishers, and anyone measuring organic traffic, the organic SERP is now conditionally optional, and the trigger isn’t in your control.

The "product architect" is the job AI is leaving behind](https://akashyap.ai/the-future-tech-role-is-the-product-architect/?utmsource=tldrnewsletter)

An opinion piece argues that as coding agents take over implementation, the human role collapses into a single “decider” who owns what to build, how the system works, and verification. Two datapoints it leans on: OpenAI’s internal Codex-built product (~1M lines, ~1,500 PRs in five months) and Cursor’s “agent fleets.” The provocative claim is that the traditional PM-architect-engineer chain is itself a coordination liability now and should be consolidated.

For working engineers, the read is uncomfortable but useful. The piece predicts that the most valuable skill on a team in two years will be judgement about what to ship, not the ability to ship it.

Frontier models are a Ferrari, most agent work is a Corolla

Source: allen.hutchison.org ↗

Vycari founder Allen Hutchison makes the cost case against frontier models for routine agent work. Their product Pepper runs orchestration almost entirely on Gemini Flash Lite 3.5, with thin slices of Claude Haiku 4.5 and GPT-5.6 Luna, processing roughly 150M tokens for about $40. The implied math is that calendar, email, and todo automation don’t need a reasoning monster; they need cheap, fast, well-orchestrated calls.

If you’re pricing agent products or building internal automations, the headline number to internalise is the $40 for 150M tokens. That’s the bar frontier pricing has to clear.

Full decentralization is a myth, pick your axis

Source: en.andros.dev ↗

An engineering essay walks through why every protocol centralizes on at least one of four axes: identity, connectivity, discovery, or cost. The worked examples: Mastodon centralizes on identity (the home instance), Nostr centralizes on cost (posts replicated across an average 34.6 relays, with 98.2% of download traffic wasted and 95% of free relays unprofitable), and Bluesky centralizes on infrastructure (its AppView). The useful move when evaluating any “decentralized” platform is to ask which axis it picked and what happens when that point of centralization fails.

AI and employment: the survey says no change

Source: marginalrevolution.com ↗

The US Census Bureau’s Business Trends and Outlook Survey added AI questions in September 2023. Adoption rose from 3.7% in late 2023 to ~10% by late 2025, and ~18% after the question was broadened in November 2025. On employment, in the most recent Nov 2025 - Feb 2026 supplement, 95.7% of firms reported no change, 2.3% increased headcount, and 2.0% decreased, almost identical to the Dec 2023 - Feb 2024 baseline despite adoption roughly tripling.

Among AI-using firms, 85% cite writing and editing email and documents as the top use, 45% document summarization, and only 13% coding. “Large number of tasks taken over” rose from 2.4% to 7.1% but is still a small slice. The headline is that broad AI adoption, so far, hasn’t moved the employment needle in either direction.

Invisible companies, and the strategy frameworks that miss them

Source: colossus.com ↗

A long essay uses Steve Ross’s path from limo rentals, a funeral home, parking lots, and cleaning via Kinney Services to acquiring Warner Bros. as a puzzle case against standard moat theory (Porter’s barriers, Barney’s VRIN). Ross had no special capability or resource by those frames, yet took home $78M in 1990, then America’s largest executive pay package. The argument is that moat frameworks miss how some companies build value through aggregation or capital structure rather than competitive advantage, and that strategy theory has a blind spot for firms that stay out of the spotlight on purpose.