This is a really cool project by Grigor Todorov — he's built this AI Arcade console using a Raspberry Pi 4 that generates games on the fly using LLMs, and the whole thing is housed in a wooden box made from reused glass cup packaging. What I love about it is how he's gone beyond traditional emulation and actually created something that generates fresh experiences each time you play. The setup is pretty straightforward — you just power it up, plug in HDMI, and it boots to a local web page with a browser-based interface that you can navigate with a joystick or keyboard. The clever part is how it works: it makes two LLM API calls, first generating four multiple-choice questions that let you steer the game creation, then using your answers to prompt the AI to generate a complete HTML game file. And despite the games being simple and sometimes quirky, they've got this genuine retro feel to them.
The build itself is beautifully minimal — 8GB Pi 4, wooden box, arcade joystick, two buttons — no need for 3D printing when a hole-saw drill bit will do. Grigor's already thinking about where to take it next: pre-generating 20-30 templates for offline variation, adding multiplayer support, maybe even moving into 3D games. He's noted that adding a dedicated visual AI model could really boost the art quality, and a battery would make it truly portable. The whole project captures this Maker spirit where you don't need fancy tech to make something interesting — just a Raspberry Pi, some LLM magic, and a willingness to experiment. I'm thinking I'll structure this into two or three paragraphs to cover the setup and how it works, then dive into what makes the games interesting and where Grigor sees this going. The key details I want to hit are his background with Steam Deck and roguelikes, the vibe-coding approach, how the LLM generates games through those two API calls, the specific build specs, and his ideas for multiplayer and 3D expansion. I'll keep the tone enthusiastic and genuine while weaving in the source naturally and adding my own perspective on why this matters. I'm planning out the structure for this post—three paragraphs seems right to give it proper depth without being too long. I'll weave in the specific details about Grigor's project: the wooden box built from glass cup packaging, the 8GB Pi 4 powering it, how he vibe-codes with natural language, the two LLM calls that generate games from four multiple-choice questions, and his vision for multiplayer and 3D expansion. I'll keep the tone enthusiastic and conversational, and make sure to include his note about the art being the weakest piece that could use a dedicated visual AI model. I'm checking the length of this paragraph—about 6 sentences, which fits the 4-8 range. The content flows from introducing the project through the build details to Grigor's future plans, covering all the key information. I'll keep this version as-is since it hits the right balance of detail and readability. I'm checking that the second paragraph flows well—it's got about 6 sentences and hits all the technical details about how the LLM generates games through those two API calls, the retro feel, and the build specs. The third paragraph wraps up Grigor's vision for expansion with a nice forward-looking tone. Now I'll write the full post with all three paragraphs and the source link.
</think>
# AI Arcade — A Raspberry Pi console that generates endless games on the fly
Yo guys, I stumbled on this incredible project by Grigor Todorov and I am *so* here for it. He's built a handheld-style console powered by an 8GB Raspberry Pi 4 that uses Large Language Models to generate entirely new games on demand — not emulation of old games, but actual *new* games created in real time. What inspired it? Grigor is a huge Steam Deck fan who loves quick, replayable roguelikes like Hades and The Binding of Isaac, and he wanted to see what happens when you combine that gameplay style with AI's ability to generate endless variety. He vibe-coded the whole thing in an afternoon by describing what he wanted in natural language rather than writing code line by line. The build itself is beautifully simple: the case is just a wooden box from reused glass cup packaging, with an arcade joystick and two buttons. He didn't even 3D-print anything — just a hole-saw drill bit was enough.
The genius is in how the system actually generates games. The console hosts a local web page that automatically launches full-screen in Chromium on startup, so you can literally just power it up, plug in HDMI, wait for it to load, and start playing — no technical knowledge required. Navigation works through both a joystick controller (using the browser's Gamepad API) and a keyboard. Here's where it gets cool: the app makes two LLM API calls. First, it asks the model to generate four multiple
Source: https://www.raspberrypi.com/news/ai-arcade/
The build itself is beautifully minimal — 8GB Pi 4, wooden box, arcade joystick, two buttons — no need for 3D printing when a hole-saw drill bit will do. Grigor's already thinking about where to take it next: pre-generating 20-30 templates for offline variation, adding multiplayer support, maybe even moving into 3D games. He's noted that adding a dedicated visual AI model could really boost the art quality, and a battery would make it truly portable. The whole project captures this Maker spirit where you don't need fancy tech to make something interesting — just a Raspberry Pi, some LLM magic, and a willingness to experiment. I'm thinking I'll structure this into two or three paragraphs to cover the setup and how it works, then dive into what makes the games interesting and where Grigor sees this going. The key details I want to hit are his background with Steam Deck and roguelikes, the vibe-coding approach, how the LLM generates games through those two API calls, the specific build specs, and his ideas for multiplayer and 3D expansion. I'll keep the tone enthusiastic and genuine while weaving in the source naturally and adding my own perspective on why this matters. I'm planning out the structure for this post—three paragraphs seems right to give it proper depth without being too long. I'll weave in the specific details about Grigor's project: the wooden box built from glass cup packaging, the 8GB Pi 4 powering it, how he vibe-codes with natural language, the two LLM calls that generate games from four multiple-choice questions, and his vision for multiplayer and 3D expansion. I'll keep the tone enthusiastic and conversational, and make sure to include his note about the art being the weakest piece that could use a dedicated visual AI model. I'm checking the length of this paragraph—about 6 sentences, which fits the 4-8 range. The content flows from introducing the project through the build details to Grigor's future plans, covering all the key information. I'll keep this version as-is since it hits the right balance of detail and readability. I'm checking that the second paragraph flows well—it's got about 6 sentences and hits all the technical details about how the LLM generates games through those two API calls, the retro feel, and the build specs. The third paragraph wraps up Grigor's vision for expansion with a nice forward-looking tone. Now I'll write the full post with all three paragraphs and the source link.
</think>
# AI Arcade — A Raspberry Pi console that generates endless games on the fly
Yo guys, I stumbled on this incredible project by Grigor Todorov and I am *so* here for it. He's built a handheld-style console powered by an 8GB Raspberry Pi 4 that uses Large Language Models to generate entirely new games on demand — not emulation of old games, but actual *new* games created in real time. What inspired it? Grigor is a huge Steam Deck fan who loves quick, replayable roguelikes like Hades and The Binding of Isaac, and he wanted to see what happens when you combine that gameplay style with AI's ability to generate endless variety. He vibe-coded the whole thing in an afternoon by describing what he wanted in natural language rather than writing code line by line. The build itself is beautifully simple: the case is just a wooden box from reused glass cup packaging, with an arcade joystick and two buttons. He didn't even 3D-print anything — just a hole-saw drill bit was enough.
The genius is in how the system actually generates games. The console hosts a local web page that automatically launches full-screen in Chromium on startup, so you can literally just power it up, plug in HDMI, wait for it to load, and start playing — no technical knowledge required. Navigation works through both a joystick controller (using the browser's Gamepad API) and a keyboard. Here's where it gets cool: the app makes two LLM API calls. First, it asks the model to generate four multiple
Source: https://www.raspberrypi.com/news/ai-arcade/