AI is eating all the RAM, so your crew fights back with AI. Pool your laptops, phones and iPads into one rig, play games as big as GTA VI better than a 30 fps PS5, and earn for every frame you share.

Inspiration

RAM got expensive this year, and AI is the reason. AI data centers are projected to take about 70% of the world's DRAM output in 2026, and console prices followed: a PS5 now costs $649 and an Xbox Series X $799.99. Games keep getting heavier anyway. GTA VI came out on November 19 and runs at 30 fps on consoles.

When our group plays together, each of us lags on our own machine while the spare laptops, phones and iPads in the room sit idle. We wanted to turn AI on the problem AI created: pool those devices, share the game's work across them, and pay whoever lends their hardware.

What we built

LANlord turns a group of friends' devices into one shared rig for a game. Our demo game, Lanternfall, shows a QR code. A friend scans it with a laptop or phone, signs in with an email, and their device starts taking work the game can afford to get back a moment late.

Job What it does Runs on Work units
NPC dialogue A local Llama 3.1 8B model voices each NPC Mac runner 10
Area generation Builds the next 512×384 area before you reach it Mac or phone 20
Crowd pathfinding Flow fields for 2,500 villagers, every 250 ms Mac or phone 2
Town simulation Runs the parts of town nobody is looking at Mac or phone 1

Rendering and input stay on the main machine. Each finished job earns the helper devnet USDC from our Solana program, and nobody sees a seed phrase or pays gas.

Results

Measured on an M-series MacBook:

Local only Linked to one helper
p95 frame time ~134 ms 16.8 ms
Frame rate stuttering 60 FPS
Job work on the main thread ~540 ms every second 0 ms

The live demo runs the game on a Windows laptop with an i9 and an RTX 4070, helped by an M3 and an M2 MacBook.

How we built it

Everything except the Solana program is TypeScript in one monorepo. The job code lives in a shared package, so the same functions run on the game's main thread, in worker threads on a Mac, and in a Web Worker on a phone. Every job is deterministic from a seed, so a result hashes the same wherever it ran.

Crowds

Crowds move on flow fields. For each of 8 destinations, Dijkstra builds one direction grid over a 384×288 map, with a step cost that rises with crowd density $\rho$:

$$c = 1 + \frac{\rho}{32}$$

The grids are recomputed every 250 ms, and each villager just reads the grid for where it's going.

Payments

A Fastify and Socket.IO server relays jobs between the game and helpers. A helper signs a receipt for each job and the game countersigns it. Every 5 seconds the server checks each helper and settles once it has 10 countersigned receipts or its oldest receipt is 60 seconds old. Our Anchor program on Solana devnet escrows each session's USDC budget and pays per work unit:

$$\text{payout} = 0.001 \times \sum_{j \in \text{batch}} w_j \ \text{USDC}$$

Each settlement emits the Merkle root of its receipts in a JobsSettled event, so anyone can check which jobs a payment covered.

Onchain data

A QuickNode Stream filter keeps successful transactions that touch our program and forwards their logs to a webhook. That webhook is the only code that writes settlements, so every earnings counter, the leaderboard, the in-game payout ticker and our status page come from chain data.

Voice and wallets

ElevenLabs runs the voice: Scribe speech-to-text for the player, streaming text-to-speech with its own voice for each NPC, and generated sound effects. Privy embedded Solana wallets are wired into the helper sign-in. Our demo currently uses the fallback, where the server creates one devnet keypair per email, so helpers still never handle a key.

Grok Bot, the game master

Our server gives Grok Bot a plain status page with onchain activity and alerts, a dispatch form, and an endpoint for pause requests.

Routine When What it does
Daily Dispatch Every morning Writes the world event, quests, NPC fallback lines, and the theme and seed for a new area that helpers generate
Network Watch Every 30 minutes Flags failing helpers and stalled payments
Party Recap When a session ends Posts a summary on X

The stack

Layer Tools
Game and web TypeScript, React, Vite, Phaser 3
Server Node.js, Fastify, Socket.IO, SQLite
Helpers Ollama, Node worker_threads, Web Workers
Onchain Solana devnet, Anchor, Circle USDC, QuickNode Streams
Voice and identity ElevenLabs, Privy
World content Grok Bot on X

Challenges we ran into

Splitting a model over Wi-Fi doesn't work. Our first idea was to split one language model across several devices. Other people's tests showed Wi-Fi is far too slow for the data a split model exchanges on every word, so each job goes whole to one device.

Browsers block local connections. An HTTPS page can't call a plain-HTTP address on the local network, so all traffic goes through our relay. That also lets a judge's phone join from anywhere.

A fast laptop hides the problem. Our host laptop ran a small village fine on its own. Judges read the code, so we refused to fake any slowdown. We scaled the world up until the laptop became the real bottleneck, and made the world size a URL parameter so we could calibrate on each machine.

One job, three runtimes. Running the same job on the host, a Mac and a phone meant banning Math.random, seeding everything, and checking output hashes in tests.

An AI agent as a user. Grok Bot uses a browser like a person does, so the status page and dispatch form had to be server-rendered HTML that works without JavaScript.

Payload size. One pathfinding pass is about 1.3 MB of JSON. That's fine on a home network and tight on conference Wi-Fi, so binary encoding is next.

What we learned

A lot of a game's expensive work can arrive a little late. Once we sorted jobs by how much delay they tolerate, much more could leave the main machine than we expected.

Paying onchain for every job would waste fees. Batching receipts every 10 jobs or 60 seconds, with a Merkle root in the event, keeps payments cheap and still checkable.

Treating an AI agent as one of our users changed how we built the web side. A form a person might fill in once had to be dependable enough for Grok Bot to fill in on a schedule.

Built With

Share this project:

Updates

Submission history