AppTheory

Most Chord Apps Show You Shapes No Hand Can Play

2026-08-28T00:34:14.800Z

Search the App Store for a guitar chord dictionary and you will find apps advertising thousands of voicings per chord. It sounds generous. It is closer to the opposite.

A six-string guitar with fifteen frets offers an enormous number of ways to sound the four notes of a Cmaj7. Enumerate them naively and you get a very large number very quickly. The trouble is that the overwhelming majority are unplayable — they need a six-fret stretch, or a fifth finger, or they ask the ring finger to sit behind the index on an adjacent string.

Advertising the raw count is advertising the size of the haystack. What a guitarist wants is the needle.

The more useful approach is to generate everything and then throw most of it away. Reject any shape spanning more than four frets. Reject anything needing a fifth finger. Reject voice crossings between adjacent strings. Reject semitone clashes, and reject intervals crowded too tightly in the bass, where they turn to mud.

Each of those rules is an opinion, and reasonable guitarists would argue about the thresholds. That is fine. An opinionated filter that returns forty usable shapes is worth more than a neutral one returning four thousand, because the second one has quietly handed the filtering work back to you.

The number worth putting on the box is not how many you generated. It is how few you kept.

Building an iOS App with Claude in Xcode: An Honest Account

2026-08-28T00:34:14.399Z

A note on what building an iOS app alongside an agentic coding assistant is actually like, minus the marketing.

Where it genuinely helps: scaffolding, boilerplate, and the long tail of API surface you would otherwise be looking up. SwiftUI and SwiftData have wide APIs with sharp edges, and having something that has read all of it is a real advantage. Pure functions with clear inputs and outputs — schedulers, parsers, solvers — come out well, and they are testable, so you can verify rather than trust.

Where it does not help: anything requiring taste. Layout that feels right, the wording on a button, whether a feature should exist at all. It will produce something plausible and you will not notice it is mediocre until later.

The failure mode worth naming is confident wrong code that compiles. It looks finished, it runs, and the bug surfaces a week later in a case nobody tested. The defence is the same as it has always been: tests around the logic you actually depend on, and reading the diff rather than accepting it.

The honest summary is that it moves the bottleneck. Typing was never the slow part of building software. Deciding what to build, and noticing when what you built is wrong, still is — and neither has been automated.

Voice Leading Is a Shortest-Path Problem

2026-08-28T00:34:14.375Z

Ask a guitarist to play Cmaj7 to Am7 to Dm7 to G7 and most will jump between memorised shapes, often travelling half the neck between chords. It works, but it sounds like four separate events rather than one moving line.

Good voice leading means each note moves as little as possible — ideally not at all. Framed that way it stops being a matter of taste and becomes a graph problem.

Each chord in the progression is a layer of nodes, one per playable voicing. Each edge between consecutive layers carries a cost: how far the fingers travel, how many notes are held in common, whether the shape jumps position. Then it is a shortest-path search through the layers.

The interesting part is the cost function, because that is where the musical judgment lives. Total fret distance alone produces sequences that are efficient and lifeless. Weighting held notes more heavily produces something that sounds intentional — the ear rewards a voice that stays put more than it rewards one that moves a little.

The scoring is where the opinions are, and it is worth arguing about. The search itself is the easy half.

Anthropic Opens a Preview of the Model Hardware Standard

2026-08-27T23:19:12.617Z

Anthropic has opened a research preview of the Model Hardware Standard, described as a shared specification for AI agents to safely operate physical devices. It is going out initially to scientific research labs and manufacturers.

The framing matters more than the spec sheet. Most agent safety work so far has assumed the agent acts on software — files, APIs, browsers — where a mistake is recoverable. Hardware is not like that. An agent driving a robot arm or a lab instrument produces consequences that cannot be rolled back, and the failure modes are physical rather than logical.

A shared standard is an attempt to get the safety layer agreed before the deployment wave rather than after it, and to make it something manufacturers implement once rather than each vendor inventing privately.

Whether it becomes a genuine standard depends entirely on adoption outside the company that wrote it. That is the part worth watching over the next year.

Source: Anthropic Newsroom, August 27, 2026.

Claude Gets Text Watermarking

2026-08-27T23:19:12.575Z

Anthropic has detailed its watermarking approach for Claude's text output, covering how the mechanism works and what it does — and does not do — to the text itself.

Text watermarking is a harder problem than image watermarking. An image has enormous redundancy to hide a signal in; prose does not, and any scheme has to survive editing, paraphrasing, and truncation while not degrading the writing. Those goals pull against each other, and most published approaches trade some of one for the other.

The practical question for anyone building on these models is what survives a round of human editing. A watermark that disappears the moment someone rewrites a paragraph answers a different question than one that persists.

Source: Anthropic, August 14, 2026.

Mac mini Picks Up the M6 and M5 Pro

2026-08-27T23:19:12.560Z

Apple has refreshed the Mac mini with two new chips. The M6 model starts at $899, and the M5 Pro at $1,699. Pre-orders opened August 25, with units shipping September 22.

The M6 brings a 12-core CPU and 12-core GPU with neural accelerators, a dual 16-core Neural Engine, 16GB of unified memory as standard, and up to 170GB/s of memory bandwidth. The M5 Pro scales to an 18-core CPU, a 20-core GPU, 64GB of memory, and 307GB/s of bandwidth, and moves up to Thunderbolt 5. Both gain Wi-Fi 7 and Bluetooth 6.

Apple's headline claims are up to 4x faster AI performance on the M6 versus the M4, and up to 8.5x faster local LLM processing on the M5 Pro versus the M2 Pro. Vendor benchmarks always deserve a raised eyebrow, but the direction is clear enough: the neural hardware is where the generational effort is going.

For anyone doing iOS development, this is the interesting one. The mini remains the cheapest way into a Mac capable of running Xcode comfortably, and at $899 the M6 undercuts the practical alternatives while running local models far better than the machine it replaces.

Source: Apple Newsroom, August 2026.

Apple Sets September 9 Event: Surprise and Shine

2026-08-27T23:19:12.541Z

Apple has confirmed its autumn event for Wednesday, September 9 at 10:00 a.m. Pacific, held at Apple Park in Cupertino. The invitation carries the tagline "Surprise and Shine" — which, going by the rumour cycle, is doing a fair amount of work.

The expected line-up is the iPhone 18 Pro and 18 Pro Max, alongside a foldable that may ship as the iPhone Ultra. Reporting puts the foldable at roughly 5.5 inches closed and 7.6 inches open, using Touch ID rather than Face ID, and dropping the telephoto camera. All three phones are expected to run the A20 Pro, built on a 2nm process.

The Apple Watch Series 12 and Watch Ultra 4 are also anticipated, with a chance of home hardware — a home hub, a refreshed Apple TV 4K, an updated HomePod mini. Release dates for iOS 27, iPadOS 27, and the next macOS should be confirmed at the event too.

Worth bracing for: price increases are widely expected this cycle, attributed to memory shortages rather than anything Apple is adding.

Source: Apple Newsroom and MacRumors, August 26, 2026.

Anthropic Funds Research Into AI and Wellbeing

2026-08-27T23:19:12.521Z

Anthropic has launched a grant programme aimed at improving how the effect of AI systems on human wellbeing is measured.

Evaluation is the weak point here. Benchmarks for capability are plentiful and well understood — you can measure whether a model solves a problem. Measuring whether sustained use of a conversational system leaves someone better or worse off is a far messier question, closer to public health research than to machine learning, and the field has comparatively little to work with.

Funding the measurement side rather than the product side is the more useful move, if the resulting instruments end up in independent hands.

Source: Anthropic, August 25, 2026.