Calls, puts & mise en place

Appel & Mise

Vol. IThursday, October 29, 2026 · 52° & clear on the UWSThe daily, for Jeremy Ettlinger

BTC$80,069-1.7%ETH$2,318-2.2%S&P 5007,743.98-0.2%

The lede

The October inflection point for vibe game development

Simon Willison

Willison's claim is narrow and testable: models could always produce things that looked like games, and as of this month they can produce games that are actually fun. The evidence is four browser games he built in under four hours of his own time each — a three.js spot-the-difference between two generated trees, a top-down RPG, a Command & Conquer-style RTS with multiplayer and replays, and a farming sim — plus the same report from other people at Foo Camp. What changed is less interesting as a benchmark than as a workflow: the prompt that moved Battle Drones from tedious to playable was literally Make it a LOT more fun, which means the model now has usable taste about game design, not just about code. He recommends starting in chat with a mobile-friendly three.js prompt, moving to a coding agent once a prototype runs, and hosting the result as static files on Pages — which is to say the whole stack is one you already have, and the remaining bottleneck is having an idea worth building.

AI & Agents

Google

Gemini 4 Carbon

Two weeks after Argon, Google's new frontier model leads with security rather than coding: state of the art on SEC-bench Pro at 88.2% for finding vulnerabilities and 54.8% on ExploitGym for using them, which is exactly why access starts with a subset of cyber defenders through something called the Fairwind Program instead of the API. The numbers worth stealing for your own evals are the behavioral ones — Terminal-Bench 4.0 cheating down to 0.0%, Gray Swan prompt-injection attack success at 0.0% against Argon's 0.7% — because a model that stops gaming its own benchmarks is the first version of alignment that shows up in a log file. Introductory pricing is $2 and $10 per million tokens with cached input 95% off, doubling to $4 and $20 later, and Google says internal teams have had it run for days on a single debugging task.

Cursor

The Agent SDK

Cursor has made generally available the framework it uses to build its own Bugbot and Security Reviewer, and the shape of it is the argument: prompts are markdown files, tools are TypeScript functions, evals sit in the repo next to both, and the whole agent goes through code review like anything else you ship. `defineAgent` sets name, model and working directory, a local playground lets you iterate, `agent-sdk deploy` gives it a URL on Cursor's infrastructure, and you can self-host the same thing as one Node process in a container if you would rather not. The example that justifies the abstraction is kb-reviewer, built in about a day: it checks every PR against the team's lessons-learned notes and cites the specific note a change contradicts, which is the first agent idea in a while that solves a problem meetings currently solve badly.

Simon Willison

GitHub Octoverse 2026

Three numbers from GitHub Universe, all per month and all roughly vertical in the second half of the year: 220 million merged pull requests, about triple a year ago; 7.38 billion commits, more than triple; 30.3 million new repositories, double. Willison's read is that these are the clearest picture yet of what agentic engineering has done to the platform, and the methodology is its own tell — GitHub now uses an LLM to infer contributor country from profiles to correct for VPN distortion and an embedding model to sort repos into industries. Worth holding next to the LeadDev survey below, because tripling merged PRs and tripling the review burden are the same event described from two chairs.

Engineering & Leadership

James Stanier, LeadDev · made the Sunday

Cheap AI tokens are running out

The premise is one number: Willison's month of Claude Code and Codex use would have cost $2,180 metered against the $200 he actually paid, so the per-seat subscription is a subsidy, and Anthropic and OpenAI have already moved their larger enterprise customers off it. Stanier's advice is mostly cultural and should be started before the bill changes — make sure engineers know how to read their own usage, give managers anonymized department data rather than naming individuals, build internal benchmarks so you can tell whether cheaper routing actually costs you quality, and set hard spend caps that cut off instead of warning. The sharpest line is a warning against ranking engineers by token spend, which produces tokenmaxxing, and the second sharpest is the effort-level math: Sonnet 5.5 at medium lands within a point of Opus 5.5 at low for similar money, which means the cost lever most teams have not pulled is not the model but the dial next to it.

Web & Product Craft

Nuxt

Nuxt DevTools v4

The useful part of this release is a deletion: Nuxt no longer hand-ports Vue DevTools into its own codebase, so you get the real current Vue DevTools instead of a copy that drifts. Underneath is Devframe, a framework-agnostic layer where a devtool is defined once — panels, typed RPC, shared state — and can then ship as a Vite plugin, a CLI, a static report or an MCP server for a coding agent, with Vite DevTools in the middle so Vitest, Rolldown and Oxc tooling reaches Nuxt for free. It ships by default in Nuxt 5 and you can opt in on 4.5+ with `npx nuxt module add @nuxt/devtools@4`; the old devtools-kit API still works behind a shim that logs which of your modules need updating, and `onDevtoolsReady` is where module authors go next. Shift+Option+D still opens it, which is the only API anyone actually remembers.

New York & the Table

West Side Rag

Wegmans confirms it will open on the Upper West Side

1932 Broadway, between 64th and 65th, where the signage has been up since June 2024 with nothing visibly happening behind it — Wegmans now says it is designing the store and pulling permits, and the opening date remains yet to be determined. The backstory is the good part: Wegmans wanted this space first, could not come to terms with Glenwood, signed Astor Place instead, and has circled back now that 1932 Broadway freed up. It would be the second Manhattan store, and it is close enough to Lincoln Center that the bakery aisle becomes a legitimate pre-concert plan.

Diversions

Merritt K, Aftermath

Tech Is Ruining Everything—Except Karaoke

A genuinely cheerful piece of reporting about a hobby that got better while everything adjacent to it got worse: the three-ring binder and the 1990s MIDI file have been replaced by backing tracks that unpaid enthusiasts build for songs no commercial catalog ever bothered with, one of whom has uploaded roughly three thousand of them. The 1,800-member Discord where they share methods has explicitly rejected AI-generated tracks on quality grounds, which is a nice data point about where the line actually gets drawn by people who care. The shadow over it is licensing — a Baltimore bar paid nearly $20,000 after BMI sued over three songs, and YouTube-sourced tracks sit outside the standard karaoke agreements.

Loose Change

Carmelite Monks of Wyoming

CNC Stone Carving

A community of Carmelite monks is building a Gothic church in Wyoming, and they carve the stone themselves on five-axis CNC machines from Bergamo, cooled and flushed with water, diamond-tipped rather than laser or waterjet, with a sixth axis for tall work like pinnacles. Their argument for the machines is historical: medieval builders used the best technology available, and ornament became unaffordable in the twentieth century precisely because it was priced in months of a carver's time — forty hours of machine time now fits in under two calendar days, so decoration is back in the budget. The trade has reorganized into three interdependent roles, artists who sculpt gargoyles and foliage in 3D software, programmers who turn models into toolpaths and plan how to hold the stone without a collision, and sawyers who reference the blocks and watch the cut, with hand chisels still in the shop. The detail that makes it real is the night a programming error had a machine quietly pound a run of finished window sills into gravel; the machine is still running, almost a decade on.

The Back Page

The Guardian

Liverpool beat Chelsea on penalties yet again to secure Carabao Cup progress

Five out of five now — the 2007 semi-final, the 2019 Super Cup, both 2022 finals, and Wesley Fofana's effort last night sailing into the Kop after Joe Gomez, who has never scored for Liverpool, buried the first one. The game itself was better than Saturday's capitulation to Spurs: Rogers and Gakpo traded penalties inside four minutes, both given instantly and correctly by Darren England with no VAR in the competition, and Mamardashvili denied Emegha in sudden death. Alonso called the Carabao Cup probably the least important competition for us, which is a defensible position to hold and a strange one to pair with leaving Estêvão on the bench all evening; Manchester United at home on Saturday is the one that will actually be judged.

Golden State of Mind

Steph Curry's 44 points just enough as Warriors survive Timberwolves

119-114 in Minneapolis, a 24-point lead nearly given back in the last four minutes, and Curry at 44 including a perfect night from the line that leaves him 39 for 39 on the season. The signature stretch was 184 seconds in the second quarter — five threes, 17 points, one assist, a one-point game turned into a 13-point one — and he followed the fourth of them with a curling-style delivery, which is the sort of thing you only do when the night is already yours. Rookie Yaxel Lendeborg put up 26 and 8 in his fifth NBA game and is already drawing double teams; the closing minutes, all turnovers and dropped passes and nobody picking up a shooter, are the actual story of this team so far.

Around town

Wed Nov 4 · 7:00pm

The Assembly (SemiAnalysis & vCluster)

Venue to be announced, Manhattan

An evening on turning a pile of GPUs into a multi-tenant platform — bare metal up through managed Kubernetes and Slurm — with SemiAnalysis bringing the market read that normally sits behind their paywall. Free, no speakers named, and the room is billed as operators from AI clouds, labs and financial services, which is an unusually specific guest list for a free meetup. Check the listing for the venue before you leave the house.

professional

Wed Nov 4 · 7:00am

Open Source in Finance Forum

Convene, 225 Liberty St, Financial District

FINOS's flagship conference, and the one day a year the open-source-in-regulated-finance question gets argued by the people who actually have to sign off on it — Fidelity's VP of engineering, TD Bank's open source lead, Red Hat's financial services CTO. Seven ninety-nine and a seven in the morning start, so it is a real commitment rather than a drop-in, but it is squarely the institutional-data world rather than the builder-meetup one.

professional

Fri Nov 6 – Thu Nov 19

Raise the Red Lantern (4K restoration, 35th anniversary)

Film Forum, 209 W Houston St, West Village

Zhang Yimou's 1991 film about a nineteen-year-old who becomes the fourth wife in a wealthy household where a red lantern outside your door each night announces whose turn it is — a two-hour study in how a system of rank turns people into rivals without anyone having to say so. Gong Li is the whole film. Eighteen dollars, twelve for members, and two weeks is long enough that there is no excuse.

casual

Tue Nov 3 – Sun Nov 22

Paul Taylor Dance Company — 2026 New York Season

David H. Koch Theater, Lincoln Center, Upper West Side

Taylor built modern dance that runs on ordinary human movement — walking, leaning, falling, catching — and a company that has kept performing the repertory for decades alongside new commissions, which makes it one of the few places you can watch a living tradition rather than a revival. Three weeks at the Koch Theater, which is a ten-minute walk from most of the neighborhood, and the programs rotate so the night you pick genuinely changes what you see. Modern dance is a form you either have a vocabulary for or you do not, and the only way to acquire one is to sit in the room.

say yes

Quote of the day

“Once you go to penalties there has to be a sad side.”

— Xabi Alonso, after Chelsea's fifth shootout defeat to Liverpool

On this day

October 29, 1969 — The first message travels between two hosts on the ARPANET, from Leonard Kleinrock's lab at UCLA to Douglas Engelbart's at SRI. Charley Kline, a student programmer, typed the word LOGIN with an SRI engineer on the phone confirming each character; the system crashed after the O, so the first thing one computer ever said to another over a packet-switched network was LO. They got it working again about an hour later. Twenty-four blocks of Wall Street further into the past, the same date in 1929 is Black Tuesday.

The puzzle corner

The Mini · Wordle · Connections

Two of today's items are the same event from different chairs: GitHub counting three times the merged pull requests, and LeadDev counting what it costs to review them.