Uber's agent orders barely exist; Meta buys Conductor; Apple rations engineer AI budgets

🔥The one thing to know

🧱New ground broken

• Uncracked City — is a mobile detective game where you interrogate AI-played suspects — typed or spoken, in six languages — and some of them deliberately mislead you; you crack the lie, then file a report, with three attempts per case. It shows a consumer product turning the model's ability to be wrong into the gameplay itself. If your team treats model unreliability purely as a defect to engineer away, here is a shipped product where it is the value proposition — useful framing when a feature idea keeps getting killed for not being trustworthy enough. The Halloween case, set in Salem, has four suspects of which one lies, and nobody has solved it yet; no account is needed and data can be erased inside the game. Uncracked City

• Siri Recaps — is one of four Audio Intelligence features Apple just put into the first developer betas of iOS 27.3 and watchOS 27.3: the Watch records and summarizes your day's conversations into notes of the things worth keeping. A hands-on reviewer found it genuinely useful for capturing verbally delivered details, and unsettling when it logged dinner-table chatter. This is the reference implementation for shipping an always-listening feature: audio processed on-device, no speaker identification, no raw audio stored, and government IDs redacted. If you're considering anything similar, the note-taking value was real but content triage was the weak spot. It is a developer beta, not shipped; the companion Live Rewind feature replays only the previous 15 seconds of conversation. TechCrunch

🚀Agent features and products that impact product management

• ChatGPT Images 2.6 — lets users turn a text prompt or an uploaded photo into an animated GIF, then adjust playback speed and edit individual frames, sharing straight from the app. It rolls out to all ChatGPT and Codex users on web, desktop and mobile, and developers get two new API models, Flare for speed and Sunburst for quality. If any part of your product generates or hosts AI imagery, motion is now the default expectation at consumer price points — static AI images stopped being a differentiator today. The Verge

• AT&T Connected Life — is now integrated with Gemini for Home: the assistant describes what security cameras saw (who delivered a package and how many items they left), answers natural-language searches over recorded video such as what dug up the yard, and arms the system based on your proximity to home. This is the platform pattern to watch: Google's Home APIs let incumbents bolt a frontier assistant onto their existing hardware, so expect assistant features to appear in competitors' lineups through partnership, not a build cycle you can match feature-for-feature. The features sit behind AT&T's professional monitoring plan at $29.99 a month, and ADT has hinted at bringing Gemini for Home in too. The Verge

• Conductor — the tool for running multiple coding agents in parallel workspaces, has been acquired by Meta, and its cloud service is winding down fast: new sign-ups stopped November 4, cloud workspaces go read-only November 13, and the cloud shuts down November 20 with user data deletion starting that day. Active subscribers get refunds, and the local app keeps models and critical fixes until early 2027. If any of your teams run agents through Conductor Cloud, you have a two-week migration on your hands. More broadly, it's the sharpest example yet that venture-backed agent tooling can vanish between sprints — keep an exit plan for any tool your delivery process depends on. Founder Charlie Holtz says the team has "something new" coming early next year at Meta; the agents Conductor deployed were never in-house, coming from Anthropic and OpenAI. Conductor

🧭How product teams work now

• One ops team's Claude Code fleet — The ops lead at a Korean B2B distributor has published the full bill for 17 bots doing daily work — complaint filing, monthly statements, error watching, pricing — on a single office MacBook and one $100/month Claude Max plan, with zero API spend. Only three of the bots touch a model at all, and those shell out to the Claude Code CLI rather than paying per token. A real cost benchmark for internal automations: the token line is trivial, but the author's hours are the real budget — ten hours a week early on, now down to one, spent writing rules and reading alerts. Fourteen of the bots are plain Python; the one evening writes were allowed to run in parallel produced around 300 database-lock errors and seven abandoned chats. Bots in Production

  “US$100 a month, one laptop that doesn't sleep, and about an hour a week of my time. The bill was never the hard part.”

• QuickProof — is a tool where you write a C specification as a small C program, much like a test, and AI agents turn it into a formal theorem that a proof checker verifies — no trust placed in the agents themselves. Running it across popular open-source projects surfaced more than 300 places where code violates its own documented promises. This is the concrete version of "specs agents can act on": the acceptance checks you write become machine-checkable claims, and the gap between documentation and behavior becomes a bug list. It also shows why spec quality, not test count, is becoming the bottleneck. GenProof

📰Market and platform news

• Apple is reportedly rationing its engineers' AI budgets. Apple has cut the daily amount some engineers can spend on AI tools from $300 to $200, with further reductions to as little as $100 planned, and is weighing a limit as low as $25 outside engineering — all per Bloomberg's sources, not confirmed by Apple. Staff reportedly read the cuts as a layoff signal. The year's tokenmaxxing push is officially being walked back at the biggest company in tech. Put a per-seat AI spend line in your 2027 budget now and decide what it covers, before finance decides for you. The reduction reportedly happened in October; TechCrunch notes the broader pattern of companies curbing usage as token bills came due. Bloomberg via explainx

• Paid Gemini subscribers discover their chats feed training. A Hacker News post that hit 606 points flags that Google uses saved Gemini chat history for model training even on paid personal subscriptions; the only opt-out is switching off Keep Activity, which drops chats after 72 hours. A subset of chats is read by human reviewers, and reviewed chats can be retained for up to three years. If your team uses consumer AI subscriptions for anything work-adjacent, treat those seats as a training surface — the privacy posture you need comes from enterprise tiers or your own tenant, not the monthly fee. Commenters in the thread say Pixel assistant features break entirely without chat history enabled. Google's own documentation warns users not to enter confidential information, and one top commenter claims every frontier provider monitors interactions for problematic behavior regardless of settings. Hacker News

• Replit Pro teams have until Saturday to opt out of arbitration. A new Replit Pro Agreement quietly added mandatory arbitration and a class-action waiver for existing customers in October, and the 30-day opt-out window closes November 8 — by email to [email protected], with each team member required to send their own notice including name, mailing address, username and account email. Vendor terms are shipping changes without release notes; a calendar entry for opt-out windows on tools your team relies on is cheap insurance. While you're in settings, the Model Improvement flag — training on your content — was switched on by default for eligible Pro accounts in September and must be turned off per person. The agreement states no one may opt out on another's behalf, and Replit says turning Model Improvement off is not retroactive. Read the Diff

• PHP2Shell — is the fallout when an AI-assisted security audit took a two-year-old out-of-memory bug in PHP — reported in 2024 and ignored as harmless — and worked out how to use it to run arbitrary code on servers, putting WordPress, Drupal and sites including NASA's and Wikipedia's at risk. A team of four engineers needed two months to build the patch. The oldest, dullest bugs on your dependency list are now reachable attack surface, and the fix side is dramatically slower than the find side. A triage pass over critical legacy dependencies, this quarter, is a defensible roadmap item. The finding came from a two-week audit by firm Calif, sponsored by OpenAI; Wiz confirmed and expanded the exploits. IEEE Spectrum

🌍The big AI stories today

• Data centers lost the midterm elections. Voters approved 33 local measures restricting data center development, with six more leading and only one failing — and Elon Musk's record-breaking roughly $450M push for Republican candidates fell short, with Tesla and SpaceX shares dropping on the news. Local permitting is now a real variable in how fast compute capacity gets built, which eventually shows up in the prices you pay per token. Tom's Hardware

• Cat Wu — left Anthropic last week and is joining OpenAI to lead product management for its core products, having been the product lead for Claude Code. The people shaping the coding-agent tools your engineers use daily are being traded between labs — worth tracking who owns which roadmap. Business Insider via Techmeme

💡One takeaway