Week of October 25, 2026
Claude Code replaces a smart TV's entire home screen in a few prompts, and Laurie Voss explains why developers using coding agents write 741% more code but ship only 30% faster.
This Week's Top Videos
Cooking with Codex: Ep. 1
By OpenAI
OpenAI's DevDay launches got a live demo in a kitchen: Dots act as always-on background agents, Codex Cloud keeps coding after you close your laptop, and a full-duplex voice model coached the cooking. A community-built Napoleon cipher decode hit nearly 5 million views, a signal of what one-off AI builds can earn in attention.
- Dots are persistent, always-on agents. In the video one ran in the background checking in guests, sending emails and building a recipe from photos of ingredients.
- Spaces give a Dot a shared workspace where it writes out its reasoning and plans. You can follow its steps there and collaborate with it directly.
- Codex Cloud runs coding agents in a cloud environment your team sets up and customizes. Work continues after you close your laptop.
- The refreshed Codex CLI adds a full-screen interface, agent monitoring, forking a conversation into a separate worktree, and voice input to the terminal. Worktree forking lets you run parallel experiments without overwriting each other's changes.
- The voice sous chef runs on a full-duplex voice model, called GPT Live 1 in the video. It waits through pauses and filler words like 'um' instead of interrupting, which matters for anyone building voice agents for hands-busy settings.
- Proof of attention economics: one community member used AI to decode a long-unsolved Napoleon letter and drew almost 5 million views. A playful game where your drawing of a horse sets its race speed earned the show's first 'double cooked' rating.
This Is How You Should Actually Make Launch Videos
By Y Combinator
A founder talking into a phone now outperforms a $20,000 produced launch video placed with the New York Times. Farza Majeed (Buildspace, Hey Clicky) treats launches as constant weekly demos on X, LinkedIn and YouTube, writing the hook and caption before building the feature, because distribution is now an ongoing channel rather than a one-shot event.
- Stop treating launches as one-shot events. On algorithmic platforms like X, LinkedIn and YouTube you are constantly launching, so ship frequent mini demos instead of one big coordinated launch.
- A simple selfie-style video where you say hi and show the product will outperform a $20,000 professionally produced video placed with the New York Times on engagement and traffic.
- Work backwards from the launch. Before building a feature, ask what the Instagram caption is and what the video hook would be, because how you communicate the product matters as much as what you build.
- Run the company around one weekly question: 'What are we launching this week?' The more demos a company ships, the better its distribution gets.
- Use a repeatable demo formula: this is who I am, this is what's going on, this is what I'm working on, let me show you. That structure turns each demo into a mini story.
- Show the process, not the polished ending. Founder videos with a DSLR, a fancy office and a funding announcement are everywhere, and the goal of a launch is users and a repeatable channel, not clout.
Meta Muse Will Create 1,000 Millionaires
By Greg Isenberg
Greg Isenberg says Meta Muse will mint 1,000+ millionaires by automating the matchmaking agencies do manually. His night nurse agency admitted it just scraped Facebook groups. Pick a niche plus city, let Muse mine public Groups, Marketplace, and Instagram, and own the cleaned buyer and seller data as a compounding, sellable asset.
- Service agencies charge fees for work Muse can now automate. Isenberg's night nurse agency admitted it found candidates by browsing Facebook groups like 'South Florida night nurses.' That is the arbitrage: replace the human sourcing layer with an agent.
- Use a four-question framework before launching: What is the niche? What is the city? What are the rules? How do you approve people? Narrow niche-city pairs like night nurses in Tampa, furniture in Austin, or plumbers in Phoenix each count as a standalone business.
- Muse handles the full agency workflow: structuring, matching, drafting DMs, calls, texts, and booking. Margins run high because you no longer hire people to comb through groups manually.
- The real moat is data, not matchmaking. Have Muse clean noisy Marketplace and Groups posts into Airtable or Google Sheets, capturing names, needs, prices, and trust signals. That repository compounds over time and can be sold later.
- Muse searches public Facebook groups, fully public Marketplace, and Instagram when prompted. It cannot see private groups, so join the relevant private groups yourself to widen its reach. Prompts can be as specific as 'night nurses at $30 an hour in Montreal.'
- Isenberg compares this moment to early eBay. The top sellers made millions by spotting unmet supply and demand on a new marketplace. Personal agents open the same buyer-seller connection play across billions of Meta users.
Claude Duet Is Here! (Better than Grok Bot?)
By Peter Yang
Claude Duet's proactivity comes from a simple default: a scheduled check-in that scans your email, calendar, and connected apps every 3 hours from 9am to 10pm. Peter Yang shows it auto-suggesting tasks, generating a styled morning brief, and building a writing style guide from 33 emails, 89 Slack messages, 47 DMs, and 11 docs.
- Duet's proactivity comes from a default scheduled check-in task that runs every 3 hours from 9am to 10pm and scans your email, calendar, and connected apps for work it can do. Five check-ins a day is too many, but you can ask Duet to change the schedule.
- Duet reuses your existing Claude account connections, so onboarding requires no reconnecting of email, calendar, or other apps. In the first conversation it inferred how the user works and suggested concrete tasks like sponsor call prep and podcast interview guides.
- Typing slash in Duet reveals hidden skills, including a morning skill that renders a styled HTML morning brief summarizing meetings and priority items like pending sponsor contracts.
- The 'learn how I write' skill analyzed 33 sent emails, 89 Slack messages, 47 DMs, and 11 docs to produce a reusable writing style guide, such as opening with the point and turning asks into numbered lists. Run it before letting an agent draft emails or documents for you.
- Duet organizes work as tasks: sub-threads within a main chat, sorted into active, upcoming, scheduled, and resolved folders. It works much like Claude Projects, and you can create custom channels and groups that auto-suggest tasks.
- Anthropic is running its familiar playbook: launch Duet as a separate research preview app, prove it works, then likely merge it back into Claude, as it did with Cowork. Duet is work-focused, with less avatar customization than Muse or Grok Bot.
I Built My Own TV Operating System With Claude Code
By Lukas Margerie
Two or three Claude Code prompts can replace your smart TV's entire home screen. Connect your TV over Wi-Fi with USB debugging enabled, show Claude a photo of the About screen, and it builds a custom Android TV launcher app. One X user claims 70% less RAM usage and a 3x faster interface after stripping bloatware.
- Photograph your TV's Settings > About screen, not the TV itself, before prompting Claude Code. The agent uses that photo to identify the exact OS. In the video, it identified a TCL running Android TV OS 14 and found the TV on the shared Wi-Fi network.
- Turn on USB debugging and select 'always allow from this computer' so Claude Code can control the TV remotely. Setup took about 3 to 4 minutes. Claude confirmed the connection by playing a test video, which ran slowly at first and got faster on the second and third tries.
- Any Android TV app that declares itself a home screen can replace the Google TV home. Claude built a working launcher with no design direction. It also rebuilt the UI as a chicken laying eggs, with each egg linking to an app, in about 5 minutes.
- Connect the Mobbin MCP to Claude Code to pull real layouts from HBO Max, Paramount Plus, Netflix and the New York Times. You can then ask Claude to remix elements you like into new home screen variants.
- Import your original TV home UI one-to-one into a MagicPath canvas. That gives you a navigable baseline you can remix and iterate on visually before deploying changes to the TV.
- This TV hack builds on an AI reverse-engineering repo that lets Claude Code or Codex take apart any app down to the binary. Builders have used the same approach to make open-source Adobe clones like Photocraft, Vectorcraft and Filmcraft.
AI Writes the Code. So What's Left for Engineers? - Laurie Voss, Arize AI
By ai.engineer
Developers using autonomous coding agents write 741% more code but ship only 30% faster, because code review is the new bottleneck. Laurie Voss argues engineering jobs will shift to whatever bottleneck agents can't clear, and that will be judgment about what 'good' means. Cognition's 'would you merge it' benchmark drops pass rates from 87% to 54%.
- Developers who turn on autonomous coding agents write 741% more code (roughly 8x) but ship only 30% faster. Writing code is no longer the constraint. Whatever comes after it is.
- A Cisco study found human code review accuracy falls off sharply past about 450 lines per hour. Properly reviewing a 10,000-line agent diff would take about a week, so most teams simply aren't doing it.
- Four AI code reviewers each found real bugs, but none of them flagged the same bug. Run several reviewers in parallel and don't rely on one. GitHub Copilot already handles more than 1 in 5 code reviews on GitHub.
- Skipping code review entirely has failed in practice. Dexter Horthy reversed his six-month 'don't read the code' advice after having to rip out whole systems. Amazon now requires senior engineer sign-off after outages it traced to shipped agent slop.
- Cognition's benchmark asks whether a maintainer would merge the PR, not just whether it passes tests. That question drops pass rates from 87% to 54%, because style, side effects, and unwritten history live in maintainers' heads.
- Voss applies Goldratt's Theory of Constraints from 'The Goal' to engineering. Each bottleneck that gets automated exposes the next one. The last one agents can't clear, a sense of what 'good' means, becomes the human job, and teams need to start training juniors for it now.