Today’s TechPulse: Cheap Macs, Repairable ThinkPads, and AI Talent Turmoil

The loudest signal in today’s news isn’t a moonshot lab demo or a courtroom drama—it’s a price tag. Apple has decided that $599 is now a Mac price, and it’s doing it with a product that reads like a strategic memo as much as a laptop: the new 13-inch MacBook Neo, announced March 4, up for preorder now, and arriving March 11. Apple is pitching it squarely at “students, families, and new Mac users,” with an A18 Pro chip, a 13-inch Liquid Retina display, and macOS Tahoe featuring Apple Intelligence and iPhone integration. If you’re keeping score, that’s Apple pushing its modern macOS experience—and the company’s on-device AI story—down into the same budget territory where Chromebooks and entry Windows machines typically spar.

Apple’s claim set is predictably bold: up to 50% faster everyday tasks and up to 3x faster on-device AI workloads than a “top-selling Intel Core Ultra 5 PC,” plus up to 16 hours of battery life. The Neo’s hardware list sounds designed to preempt the usual “cheap laptop” compromises: 1080p FaceTime camera, dual mics, spatial audio speakers, Magic Keyboard, a “large” trackpad, and two USB‑C ports, with Touch ID called out as part of the overall experience. Even the aesthetics are doing work here—four colors (Silver, Blush, Citrus, Indigo) and a recycled aluminum enclosure (Apple says 60% recycled content by weight), framing affordability without the stigma of “budget.”

But Apple is also testing how much the market will tolerate when affordability collides with spec-sheet expectations. The debate, as it’s already forming, centers on the Neo’s modest base configuration: 8GB unified RAM. Apple’s bet is that vertical integration, macOS optimization, and the A18 Pro’s characteristics can keep the experience feeling “premium enough” even at that floor. In education—where the Neo hits $499—that’s potentially a disruptive wedge, especially if Apple can credibly attach “on-device AI” to the value proposition without requiring cloud subscriptions or constant connectivity. The broader implication is less about one laptop and more about a new low-end reference point for Macs: if macOS plus Apple Intelligence can live at $599, competitors in the Chromebook and low-end Windows ecosystem may feel pressure to respond on performance-per-dollar, battery life, and AI messaging all at once.

That push toward a more accessible mainstream is mirrored—almost like a counterpoint—by what Lenovo is doing at the other end of the laptop conversation: not the sticker price, but the lifecycle. Lenovo’s ThinkPad T14 Gen 7 and T16 Gen 5 earned a provisional 10/10 repairability score from iFixit, the first time a T-series model has hit the top rating. That matters because these are not niche machines aimed at hobbyists with spudgers; they’re corporate workhorses. When a fleet laptop becomes genuinely serviceable, it changes the economics of IT departments and the practical sustainability of enterprise hardware in a way that glossy “green” marketing rarely does.

What’s striking in iFixit’s account is how unglamorous the path to 10/10 sounds—and that’s the point. Lenovo didn’t just “add repairability” as a late feature; it pulled repairability conversations earlier into development and coordinated across design, engineering, service, quality, and sustainability teams to make “small but consequential” decisions. The results include easier keyboard replacement, a return to modular LPCAMM2-style memory, and standard M.2 SSDs. This is the kind of design shift that, once normalized, tends to spread: if mainstream enterprise buyers start to factor repairability into procurement, service-friendly construction becomes a competitive advantage instead of a concession.

It’s also hard not to read this as a quiet rebuke to the industry’s long drift toward sealed, frictionless minimalism. iFixit emphasizes that these changes aim at serviceability “without sacrificing performance or reliability,” and Lenovo’s collaboration suggests a more intentional balance between thinness, thermals, and maintainability. For sustainability, the biggest win is often not recycled materials (though those help) but keeping devices in service longer, reducing e-waste and downtime. If the MacBook Neo is about expanding the funnel into a platform, Lenovo’s repairable ThinkPads are about keeping a platform stable and cost-effective at scale.

Stability is the theme that feels least assured in AI today, and the most human story in the briefing comes from Alibaba’s Qwen team. Reports compiled by Simon Willison describe a sudden wave of resignations among key Qwen researchers, including lead researcher Junyang Lin, who announced his departure March 4. Multiple other senior contributors—leads spanning code, post-training, and VL/Coder work—have reportedly resigned amid an internal reorganization. The reporting frames a particularly combustible detail: a recent hire from Google’s Gemini team was placed above Qwen, and that appears to have been a breaking point. Alibaba’s CEO reportedly held an emergency all-hands, underscoring the disruption.

The timing is what makes this feel like more than routine churn. Qwen 3.5 has been released to strong reception, including a massive 397B model and a range of smaller variants (27B, 35B, 9B, 4B, 2B) that have been praised for coding and multimodal capabilities. In the open-weight ecosystem, momentum isn’t just about model weights—it’s about the team that knows how to train, tune, evaluate, and ship them. When the people who hold that tacit knowledge walk out, continuity becomes a question mark even if the code and checkpoints remain. VentureBeat’s framing—“Did Alibaba just kneecap its powerful Qwen AI team?”—captures the stakes: organizational decisions can reshape the open-model competitive landscape faster than any benchmark chart.

If the Qwen situation is a reminder that AI progress depends on fragile social systems, Google’s latest Play Store shift is a reminder that platform economics can change abruptly too—especially when antitrust pressure and developer frustration converge. Engadget reports that Google is ending its standard 30% Play Store fee, replacing it with lower, region- and program-specific rates: 20% generally, 15% for certain new installs, and 10% for subscriptions. Developers in the UK, US, and EEA using Google’s billing pay 5%, while other regions get market-specific rates. Just as consequential, Google is formally enabling third-party billing and third-party app stores on Android, including the ability for apps to offer alternative billing or link users to external purchase flows.

This isn’t an overnight flip; it’s a phased rollout, region by region, through September 2027. But the direction is clear: Android’s app economy is being re-architected toward a world where Google Play is a major distribution channel, not the toll gate. Google is also launching a voluntary Registered App Stores program with a streamlined installation interface for approved stores, while sideloading remains possible. If you want a tangible indicator of how meaningful this could be, Engadget notes that Epic Games’ Tim Sweeney confirmed Fortnite will return worldwide to Google Play. The subtext is obvious: when pricing and billing constraints loosen, big developers who previously treated Play as hostile territory may reconsider—reshaping not just where apps are found, but how they’re priced, bundled, and updated.

While app store economics are being renegotiated in public, developer workflows are being renegotiated in private—inside teams that are trying to make AI actually useful without turning codebases into haunted houses. Raycast’s newly announced Glaze is an attempt to push “conversational AI” into a very specific niche: building local-first macOS utilities in minutes, with the kinds of OS-level hooks web-first builders can’t easily provide. Glaze promises native utilities with file-system access, menu-bar integration, background processes, and hardware hooks, plus internal team stores, a public marketplace, and optional code editing. Pricing is positioned as accessible—Raycast says a free tier with daily credits is planned, with paid plans starting at $20/month, and enterprise/team features. Windows and Linux support are planned after the Mac launch.

At the same time, the less flashy but arguably more foundational tooling story is about merges—because nothing says “the future of software” like arguing with a conflict marker. Weave proposes replacing Git’s line-based diffing with an entity-aware merge algorithm built on tree-sitter parsing. Instead of treating code as text lines, Weave tries to match semantic entities—functions, classes, keys—across base/ours/theirs, auto-resolving independent changes even when they happen in the same file, and surfacing conflicts with context when changes genuinely collide (like modify vs delete). The project reports benchmarks across major open-source repositories (including git, Flask, CPython, Go, and TypeScript projects) showing fewer false conflicts with zero regressions. In a world where “multi-agent development” and high-concurrency edits are becoming normal, semantic merging feels like one of those unsexy innovations that might quietly save thousands of human hours.

Those human hours also depend on trust—particularly that the network isn’t leaking more about you than it needs to. The IETF’s publication of RFC 9849, standardizing TLS Encrypted Client Hello (ECH), is a concrete step toward better handshake privacy. ECH encrypts the ClientHello contents, including identifiers like SNI and ALPN, under a server public key. The RFC frames ECH as a way to reduce on-path leakage about which domain a client is trying to reach, building an anonymity set among co-located servers with similar external behavior. It’s not a magic cloak: the RFC notes limitations like plaintext DNS and IP address visibility, and stresses that real-world protection depends on deployment details—server configs, record formatting, and DNS privacy among them. Still, standards-track clarity matters; implementers now have a defined path to making “what site are you visiting?” harder to infer from the handshake metadata alone.

And then there’s the story that reads like a dare: a CPU that runs entirely on GPU, built out of neural networks. The nCPU project implements registers, memory, flags, and the program counter as PyTorch tensors, and executes ALU operations via trained .pt models—23 models totaling 135 MB. The details are delightfully specific: addition uses a learned Kogge-Stone carry-lookahead, multiplication is a byte-pair lookup-table model, bitwise ops are neural truth tables, and shifts use attention-based routing. It loads models in about 60 ms, stays on-device with no host round-trips, and achieves verified 100% integer correctness across tests. Performance is quirky on Apple Silicon (MPS): multiply is vastly faster than add, producing an overall reported rate around 4.9k IPS. Nobody is pretending this replaces silicon arithmetic tomorrow; what it does is force a fresh look at the boundary between “model” and “machine,” and at how correctness, latency, and architecture interact when you express computation as learned components.

The through-line across all of this is that 2026’s tech landscape is being rebuilt at the seams: a cheaper Mac that tries to redefine the entry tier, enterprise laptops that treat repairability as a first-class feature, AI teams whose momentum can be derailed by org charts, mobile platforms that are loosening their grip on billing, developer tools that make AI workflows and collaboration less painful, and network standards that reduce passive surveillance opportunities. The next few months will test which of these shifts stick—not as announcements, but as habits. If Apple’s Neo lands in classrooms, if repairable ThinkPads become procurement defaults, if Qwen’s open-weight cadence slows (or doesn’t), and if Google Play’s fee and billing changes actually alter developer behavior, we’ll look back on this period as less about singular breakthroughs and more about who controls the pipes, who pays the tolls, and who gets to keep building.