Glyph GPT |

𓉡𓀬𓉢

Defragmenting pottery shards...

𓇋 𓆑 [F51D] 𓈖 𓄂 𓏏 [F51D] 𓈖 𓄂 𓏏 [F51D] 𓋴 𓂋 𓅾 𓎺 𓍿 𓂋 𓊪 𓅾 𓎺 𓊃 𓏏 𓅾 𓎺 𓋴 𓅾 𓎺 𓏠 𓏌 𓏏 𓅾 𓎺 𓏤 𓏏 𓏐 𓊃 𓇋 𓆑 𓏖 𓏤 𓈙 𓂝 𓏏 𓏐 𓏤 𓏤 𓈖 𓊪 𓅮 𓏏 𓎺 𓏤 𓏤 𓅓 𓊃 𓏏 𓈓 𓎺 𓏤 𓏤 𓂦 𓏏 𓏊 𓎺 𓏤 𓏤 𓂦 𓂋 𓏏 𓏊 𓏤 𓏤 𓂦 𓂋 𓏏 𓏊 𓏤 𓎛 𓈎 𓏏 𓏊 𓐍 𓈖 𓋴 𓀍 𓏤 𓏤 𓎛 𓈎 𓏏 𓏊 𓏤 𓏤 𓋴 𓐍 𓊪 𓏏 𓎺 𓏤 𓏤 𓊪 𓐍 𓄿 𓎺 𓏤 𓏤 𓈋 𓃛 𓏊 𓏤 𓏤 𓂧 𓄿 𓃀 𓄿 𓈒 𓈒 𓈒 𓈒 𓎺 𓏤 𓏤 𓇋 𓂋 𓊪 𓇭 𓎺 𓏤 𓏤 𓇋 𓂋 𓊪 𓇭 𓎺 𓏤 𓏤 𓇋 𓂋 𓊪 𓇭 𓎺 𓏤 𓏤 𓎛 𓃀 𓇑 𓇑 𓏏 𓈓 𓎺 𓏤 𓏤 𓐍 𓈖 𓆑 𓅱 𓎺 𓏤 𓏤 𓇋 𓈙 𓂧 𓈓 𓎺 𓏤 𓏤 𓋴 𓄡 𓏏 𓇅 𓎺 𓏤 𓏤 𓋴 𓄡 𓏏 𓇅 𓇅 𓎺 𓏤 𓏤 𓂝 [S105] 𓎼 𓏏 𓂝 𓎺 𓏤 𓏤 𓇠 𓂝 𓎼 𓏏 𓎺 𓏤 𓏤 𓃀 𓅡 𓏏 𓇠
1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time, 1 time 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 portion, 1 por 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 portions, 2 por The grain
Text copied

Gardiner codes: M17 I9 F51D N35 F4 X1 F51D N35 F4 X1 F51D S29 D21 G54 W10 V13 D21 Q3 G54 W10 O34 X1 G54 W10 S29 G54 W10 Y5 W24 X1 G54 W10 Z1 X1 X2 O34 M17 I9 X6 Z1 N37 D36 X1 X2 Z1 Z1 N35 Q3 G40 X1 W10 Z1 Z1 G17 O34 X1 N33A W10 Z1 Z1 D45 X1 W22 W10 Z1 Z1 D45 D21 X1 W22 Z1 Z1 D45 D21 X1 W22 Z1 V28 N29 X1 W22 Aa1 N35 S29 A11 Z1 Z1 V28 N29 X1 W22 Z1 Z1 S29 Aa1 Q3 X1 W10 Z1 Z1 Q3 Aa1 G1 W10 Z1 Z1 N26 E9 W22 Z1 Z1 D46 G1 D58 G1 N33 N33 N33 N33 W10 Z1 Z1 M17 D21 Q3 M43 W10 Z1 Z1 M17 D21 Q3 M43 W10 Z1 Z1 M17 D21 Q3 M4

How does it work?

GlyphGPT begins with nanoGPT, Andrej Karpathy's GPT architecture implementation, trained from scratch exclusively on hieroglyphic texts painstakingly prepared by Mattia De Cao. Optimized with Claude Code autoresearch. This generates streams of hieroglyphic sequences, which we then translate into English using Mattia's hiero-transformer - which iteself is based on Facebook's M2M-100 translation model, fine-tuned on the same hieroglyphics dataset. Essentially a mini AI learns to babble Middle and Old Egyptian, and big AI does its best to translate that into English. Both models have some room for improvement, but what could go wrong? Actually one thing for sure going wrong, Gardner notation has a larger vocabulary than unicode; so [Z3A] indicates the A variant of [Z3] but it is displayed as [𓏪A].

Contact

EgyptGPT GitHub

Thanks to Andrej Karpathy's for NanoGPT, and Mattia De Cao for releasing the translation model and dataset.