7 min read · 14 sources

Apple’s Mac Mini is about to break a two-year silence. A refresh could land within days, ahead of the September 9 event where the company is expected to unveil its first foldable iPhone and the iPhone 18 Pro line. Apple has been testing versions with both its M5 and next-generation M6 chips, though the final silicon is unconfirmed. Demand for the current Mini has surged because the box turned out to be a sweet spot for running AI locally, and the new one is going to have to clear that bar.

The rest of the day is a tour of what AI is actually doing to the stack: how it’s wired in, what it eats, who runs the power, and what happens to the humans left in the loop.

Uber says more than 70% of its pull requests are now written by AI agents, on a custom platform that doubled code shipped per engineer in a year.

Apple's Mac Mini Returns, and So Does Local AI

Source: bloomberg.com ↗

Bloomberg’s Mark Gurman reports the Mini refresh is imminent, alongside an OLED iPad Mini and AirPods 5. The interesting question is the chip. Apple running both M5 and M6 silicon through testing means a launch on M5 is the safe bet, with M6 arriving as a mid-cycle surprise or a quiet step-up SKU later. Either way, the Mini has quietly become the platform engineers reach for when they want to run a local LLM without paying cloud rent, so memory bandwidth and unified memory ceiling are the specs that actually matter.

Uber's Software Factory: 70% of PRs Are Agents Now

Source: newsletter.port.io ↗

Port’s deep dive into Uber’s AI Software Factory is the most concrete reference architecture for agentic coding at org scale that has been published. The numbers are the headline: more than 70% of pull requests at Uber are now written by agents, and code shipped per engineer has doubled in a year. The plumbing underneath is six pieces — a context graph with 40M entries across 150 node and edge types that replaces 20 to 30 source systems, an MCP gateway, an LLM gateway, compute, governance, and workflow orchestration on top.

The MCP gateway is the load-bearing wall. It mediates which tools and resources an agent can see, scopes permissions per task, and is the place where the org’s policy actually gets enforced. If you are trying to figure out what an internal agent platform is supposed to look like when you have a few thousand engineers, this is the blueprint. The lesson the author keeps coming back to: platform infrastructure has to exist before agent workflows do, not after.

Tesla's Cybercab: Same Stack, New Body

Source: electrek.co ↗

Tesla confirmed a Cybercab event in Austin on September 3, invite-only for top Robotaxi riders and winners of a sweepstakes, livestreamed publicly. The Cybercab is a two-seat, no-steering-wheel vehicle that joins the existing Robotaxi app, running on the same FSD software stack already in Model Y robotaxis. The autonomy software is unchanged — the demo is the hardware.

That matters for context. Tesla’s active unsupervised fleet sits at roughly 20 to 30 vehicles, against Waymo’s roughly 3,000. The company claims about 380,000 cumulative unsupervised miles as of its July earnings call. The Cybercab adds a purpose-built body to a small service, and on September 3 you should expect a launch narrative, not a step-change in fleet size.

Solid-State Transformers: The Grid Piece AI Was Missing

Source: arstechnica.com ↗

The AI data center boom is now pulling the physical electrical stack behind it. Ars Technica reports that solid-state power transformers, long a quiet R&D project, are getting serious investment because they slot directly into hyperscaler buildouts. They use high-frequency semiconductor switching, get manufactured from silicon carbide, and do AC-to-DC conversion in one box where the old stack needed separate transformer and rectifier stages.

For an SRE or a data center engineer the relevant differences are size, weight, and modularity. They are physically smaller and lighter than conventional transformers, and the modular design means a failed module gets swapped instead of a transformer getting pulled. The conversion from grid AC to the DC rails the servers actually want happens in a single device, which changes the upstream design assumptions for new builds.

Humanoids Are a Production Business Now

Source: itcanthink.substack.com ↗

The World Humanoid Games in Beijing made one thing obvious: humanoids are past the demo reel. Honor’s Lightning humanoid reportedly beat Usain Bolt’s 100m record, alongside a steady background hum of fires and crashes, which is itself the point — these are devices that get tried, broken, and re-spun fast. Unitree sold over 5,500 humanoid robots in 2025, mostly the G1, and is the benchmark for fast iteration, vertical integration, and durability testing. Agibot is reportedly producing at higher volumes.

The takeaway for anyone outside robotics: a mature, fast-iterating humanoid ecosystem with real production-scale shipments now exists, and the gap between demo and deployable hardware is closing on a quarter-by-quarter basis.

Context Engineering Is Just State Machine Design

Source: vedanshh.com ↗

Vedansh’s piece on context engineering reframes the agent stack with one useful line: the context window is the whole state machine. Every piece of a harness either adds information to the model’s context, gates what gets in, or validates what comes out. The trap is thinking in terms of which components to add — the real question is what is worth keeping when the context is finite.

If you have ever tried to bolt a RAG layer onto a chat agent and watched it get slower and weirder, this is the framing that explains why. The harness is not an assembly of features, it is a state machine with a fixed budget, and the discipline is choosing what to evict.

The Data Center Backlash, Quantified

Source: thezvi.substack.com ↗

The Zvi summarises polling that 75% of Americans oppose local data center development — worse than coal plants — even when the economic benefits are spelled out. The author argues the opposition is mostly a proxy for distrust of AI, big tech, big money, and NIMBY-ism, and is not meaningfully moved by arguments about tangible costs and benefits. Even purely beneficial infrastructure like transmission lines is being pulled into the same backlash.

The practical read for anyone planning a buildout: the political environment around data centers is becoming a real bottleneck, and “but the jobs and the tax base” is not a counter-argument that lands.

SQLite as an Executable Format

Source: fzakaria.com ↗

Farid Zakaria’s prototype replaces ELF with SELF — a Structured Executable & Linkable Format whose on-disk representation is a SQLite database you can chmod +x and run. The post maps ELF internals onto database primitives: string interning instead of strtab, indexes instead of .gnu.hash, schema tables instead of section headers, b-tree pages instead of segment layouts, foreign keys instead of relocations. There is a working sqlelf tool for SQL-based ELF exploration, and the code is on GitHub.

If this proves out, executables become queryable, mutable, and dramatically more inspectable — and a lot of the awkward tooling that grew up around readelf, objdump, and nm gets replaced by SQL.

Starlink Wants to Be a Cell Carrier, From Space

Source: potsandpansbyccg.com ↗

On SpaceX’s first earnings call, Starlink laid out a plan to enter terrestrial cellular by deploying cell sites with satellite backhaul, riding on next-gen V3 satellites and the existing constellation. The obstacles are concrete: the US cellular market is already competitive, Starlink only acquired 65 MHz from EchoStar against an 800 MHz FCC auction mandate, and feeding hundreds of thousands of cell sites from a constellation currently sitting at roughly 15,000 approved satellites is a real bandwidth problem. The requests to grow the constellation to 100,000 are still pending.

The author’s read: the US is the harder market, and developing markets like Nigeria are a more realistic first target. The plan’s feasibility hinges on spectrum acquisition, V3 capacity, and how fast the constellation actually scales.

The Rest: Streaming Bundles, Founder Prep, and the Harness Org Chart

Source: nytimes.com ↗

YouTube and Roku are racing to become the front door for everything on your TV, with a Peacock deal inside YouTube marking the bundling war’s first obvious shot. Paul Graham argues universities should not bother with “entrepreneurship” curricula and should instead teach CS and engineering well, because founders win by being good at building things, not by what they studied. The Harness Is the Company lays out a four-stage trajectory from no harness to harnesses orchestrating individuals, where domain knowledge and review loops become the core asset. And a piece on engineering leadership pushes back on the AI-productivity narrative, arguing that executive statements about needing fewer people destroy psychological safety and erode the conditions that actually make teams productive.