Best AI Creations This Week: AI Agents Clean Up Smart TVs (Oct 23)

The best AI creations this week (October 17 to 23, 2026) had one thing in common: people pointed AI at the hardware they already own and made it behave. The standout is a coding-agent prompt that strips ads and bloat from smart TVs, which spread fast enough for TechCrunch to cover it. Close behind: a 360M-parameter model that pretends to be a whole terminal, and llama.cpp running on a jailbroken PS5. On the design side, an AI studio that outputs print-ready dielines for stickers and gift boxes. I am a product designer and design engineer who builds with AI every week, so I picked these for craft, not for hype. Every entry links the original post.

The best AI creations this week

Six picks, each checked against its original post. I only use numbers the source itself shows.

1. CleanYourTV: a Claude Code prompt that debloats your TV

Gathin T. (@gathint on X) saw a sponsored ad on his parents' TV before Netflix, so he gave Claude Code the TV's model number and let it work over Wi-Fi. He says it removed 25 Xiaomi packages and wrote a new launcher. He then shared the prompt and tools as CleanYourTV on GitHub, where the repo showed about 26.6k stars when I checked. TechCrunch reported that a few thousand people ran the same prompt within two days.

What I like as a designer: the repo is built around a PROMPT.md file that asks before touching anything you might use, and every change is reversible. That is a good consent pattern for an agent touching someone's living room.

2. dumbtv: the Fire TV version, with an undo button

Muzzammil Yasin Sabuwala (@DevMuzzammil on X) built dumbtv, a set of scripts that turns off ads, telemetry and autoplay on Fire TV and swaps in a home screen with no internet permission. He says it was coded with DeepSeek V4.1 Flash, Claude Opus 5.5 and Claude Sonnet 5.5, and the README is written for coding agents to follow. Every step comes with a restore.sh. Two independent projects in one week landing on the same idea is the signal here: people want their devices back, and an agent is now the screwdriver.

3. WorldTUI: a 360M model that pretends to be a terminal

Simon Edwardsson wrote up WorldTUI, a tiny language model that imitates a terminal. There is no real shell behind it. Each keypress goes in and the model predicts the next screen, so you can open vim, play snake or start a Python prompt inside a hallucinated computer. He fine-tuned Qwen2.5-Coder-0.5B on about 336k synthetic sessions in roughly 75 GPU-hours on one RTX 4090, and it runs in the browser through WebGPU. On his ShellBench test it scored 80 out of 106, against 52 and 60 for the prompted baselines. The Hacker News thread passed 560 points.

Why it matters for interface people: this is a world model for UI. Today it is a terminal. The same trick pointed at app screens is a prototype that invents itself as you click.

4. UseIt Studio: an AI design agent that outputs real dielines

Lindia Chen and Daniel Gao launched UseIt Studio on Product Hunt. You describe a physical product in one sentence (a pin, a sticker sheet, a gift box, a pop-up card), it asks clarifying questions, and it hands back an editable file with real dimensions, bleed and dielines, checked against a manufacturing knowledge base. You can circle a part to change only that part. It showed 329 points on the weekly leaderboard when I checked.

This is the one I would test first. Most AI image tools stop at a pretty picture. A file a print shop can actually cut is a much more honest promise.

5. PS5LM: llama.cpp running on a jailbroken PS5

PS5LM by cobanov runs llama.cpp on the PS5's GPU and serves models to the rest of your network through OpenAI- and Anthropic-compatible endpoints, plus a chat page and a TV interface. The README lists about 55 tokens per second for Qwen 3.5 9B and about 20 for a 2-bit Qwen 3.8 27B on a PS5 Slim. It reached the Hacker News front page on October 17. Commenters were split on how useful a 2-bit 27B model is, which is fair. It still says something that a game console in a living room can now host an assistant for the whole house.

6. Beam.win: run agents on one Mac, use their windows on another

Hiten Shah and Sam Asante launched Beam.win, which streams single app windows from a Mac mini or Studio to the Mac in front of you. Each remote window gets its own Dock icon and a place in Cmd-Tab, so an agent's terminal and browser feel local while the work runs elsewhere. It is less a creation and more a tool for creators, but it sat second on Product Hunt's weekly leaderboard with 379 points when I checked, and it solves a real problem: agents need a machine, and you need your laptop back.

This week's AI creations at a glance

Creation	Creator	Built with	Why it matters
CleanYourTV	Gathin T.	Claude Code (also works with Codex, Gemini CLI)	A shared prompt became a reusable product
dumbtv	Muzzammil Yasin Sabuwala	DeepSeek V4.1 Flash, Claude Opus 5.5, Claude Sonnet 5.5	Reversible agent actions as a design rule
WorldTUI	Simon Edwardsson	Fine-tuned Qwen2.5-Coder-0.5B, WebGPU	A small world model that simulates an interface
UseIt Studio	Lindia Chen, Daniel Gao	AI design agent with manufacturing rules	AI design that ends in a file you can produce
PS5LM	cobanov	llama.cpp, open-weight Qwen models	Local AI on hardware people already own
Beam.win	Hiten Shah, Sam Asante	Not stated (a tool for running agents)	Agents get their own machine, you keep yours

From my lab

Every week I add one or two of my own experiments to the list. These are playable builds, not mockups, and both fit this week's theme of AI meeting physical things.

Muse AI Gadget: hardware firmware running in a browser

My Muse AI Gadget experiment runs Meta's open-source Muse firmware inside your browser as a 3D gadget sitting on a desk. You can drag it around, poke it, and ask it for flights and Airbnbs. The stack is the Muse SDK, Claude, WebAssembly, LVGL for the device UI, three.js for the object, and ElevenLabs for sound. Try the live Muse demo. If the TV and PS5 hacks above made you think about device UI, this is the cheap way to prototype it: no hardware, real firmware.

Haribo Pick & Mix: a candy store with real physics

The Haribo Pick & Mix experiment is a candy store where 3D gummy bears pour into the bag and a live scale weighs every bear and updates the price. It was designed in Figma and built with Claude through the Figma MCP, using three.js, Rapier physics and ElevenLabs sound. Play the live pick and mix demo. Next to UseIt Studio, it is the other half of the same idea: physical products deserve interfaces that feel physical.

All my builds live on the AI experiments page.

What this week means for designers

The trend is clear: AI is leaving the chat box and going into objects. TVs, consoles, a second Mac, a sticker sheet. The interesting work is no longer "can a model do it", it is "how do we let it touch something real without breaking trust".

The hype check: the TV projects are early and uneven. TechCrunch's report includes a user who bricked an Android TV and another who got stuck halfway through a Fire TV. WorldTUI forgets files. A 2-bit model on a PS5 is a demo, not a daily driver. None of that makes them less interesting, but it should make us careful about calling any of it finished.

The design lesson is the same one I learned shipping AI tools at ViewStats: designing AI features is designing trust. Ask before acting, show what will change, and always ship the undo. Both TV projects got that right, and I think that is why people were willing to run them on the screen in their living room. I wrote more about where that leaves our role in what AI actually replaced in my product design work, and the trust story behind those AI launches is in the ViewStats creator tools case study.

If you want this kind of thinking inside your product, with someone who designs it and ships the code, work with me as your design engineer.

Frequently asked questions

What were the best AI creations this week?

For October 17 to 23, 2026: CleanYourTV and dumbtv (coding agents that remove ads and tracking from smart TVs), WorldTUI (a 360M model that simulates a terminal), UseIt Studio (an AI design agent for print-ready physical products), PS5LM (llama.cpp on a jailbroken PS5) and Beam.win (remote windows for agents on another Mac).

What AI tools are people using to build apps and hardware hacks right now?

This week the named tools were Claude Code, Claude Opus 5.5 and Sonnet 5.5, DeepSeek V4.1 Flash, Codex and Gemini CLI for agent work, plus open-weight Qwen models and llama.cpp for running AI locally. In my own experiments I mostly use Figma, Claude and the Figma MCP, with three.js, SwiftUI and ElevenLabs.

Is it safe to let an AI agent change my smart TV?

It is a risk you should take knowingly. Both projects avoid root and include restore scripts, but users still reported a bricked TV and a stuck install. Read the README, back up first, and turn network debugging off when you are done.

What is new in AI this week for designers?

Design agents are starting to output production files instead of images, as UseIt Studio shows with dielines and bleed. Small world models like WorldTUI hint at prototypes that generate their own screens. Both reward designers who understand constraints, not just visuals.

How often is this AI roundup published?

Once a week, every week, with six to ten creations from the last seven days and one or two of my own experiments.