Launching today

TapeOutTank

Launching today

Battle a fruit-fly-inspired tank. Inspect its circuit brain.

2 followers

Battle a fruit-fly-inspired tank. Inspect its circuit brain.

2 followers

Visit website

Built with GPT-6 Astra in Codex, TapeOutTank turns trained controllers into playable circuit NFTs. Challenge Atlas, Swift or a 64-node fruit-fly-derived network; watch them duel, inspect decisions and download replays. Fly11 is deployed on BNB Chain as 36 child circuits plus a root NFT. Free to play, no wallet needed. Matches run off-chain; the fly is a trained subnetwork, not a whole-brain simulation.

Free

Launch tags:Open Source•Games•OpenAI Day

Launch Team / Built With

Framer AI AgentsDesign and publish professional sites with AI

Promoted

OpenAI

GPT-6 Astra Challenge

What became possible in your product with Astra that was not practical before?

Maker

Astra in Codex helped us turn a small tank-game experiment into a reproducible, playable circuit-research demo. It supported the cross-disciplinary iteration: diagnosing training failures, implementing fixed-point arithmetic, debugging circuit compilation, evaluating ten compression stages, and integrating a 64-node fruit-fly-derived controller as 36 child circuits plus a root NFT on BNB Chain. Visitors can challenge three frozen controllers, inspect their decisions and download replays in four languages. The practical gain was making this end-to-end research-and-release workflow manageable for a small project; we have not measured a controlled productivity comparison against earlier models. Astra is our development agent, not the runtime player: matches execute frozen controllers off-chain, with no live LLM calls or whole-brain simulation.

Report

23h ago

Maker

📌

Hi Product Hunt! I started with a small question: could a tiny logic circuit become a tank player you can actually challenge? Then the experiment got wonderfully out of hand. We trained Atlas and Swift, explored a fruit-fly-derived network, diagnosed students that could not even fit their training examples, and worked through ten compression stages to make the frozen fly controller deployable. Try playing against the three players, watching them duel, or inspecting their observations, calculations and actions. You can download replays too. GPT-6 Astra in Codex helped with implementation, experiment tooling, debugging, verification and release work. Gameplay itself does not call an LLM. Fly11 preserves a selected 64-node, 1,371-connection anatomical structure with engineered, trained numerical dynamics. Its NFT bundle contains 56,702 NAND and 1,615 LATCH. Compression traded circuit area for sequential work. This is a technical demo: matches run off-chain, the model does not learn during play, and our comparisons have not established an advantage from biological topology. The anatomical point cloud is a visualization, not an entire brain simulation. Free to play, no wallet required, with English, Japanese, Korean and Chinese interfaces. Code and research notes: https://github.com/tizerluo/tape... Can you understand why a tank chose its move? Which failure would you teach the next student to handle?

Report

3d ago