can yall tell me if you found this in search or cause you saw someone else lov/fav the project? ✦ ━━━━━━━━━━━━━━━━━━━━━━ ✦ ⚡ MythicGPT 4 Pro ⚡ small AI language model — runs in Scratch ★ real AI, like a tiny ChatGPT ★ scroll down for more information ✦ ━━━━━━━━━━━━━━━━━━━━━━ ✦ ★ RUNS FULLY IN SCRATCH — NO API, NO SERVERS ★ Runs in vanilla Scratch, no Turbowarp required. (use turbowarp tho if its running rly slow in scratch) Check out the other ai models I made if this one is running very slowly on your computer :D https://scratch.mit.edu/studios/51676816 ★ CHANGES FROM 4 MINI ★ - Architecture: 5-layer Transformer with a full-softmax output head (4 Mini used a 4-layer factorized-softmax head) — this is a structural fix, not just a data tweak, for the "word-salad" garbling - New: remembers made-up names and words it's never seen before: tell it "call me Zorblax" and it remembers "Zorblax" later, even though that exact word isn't in its vocabulary - New: adjustable memory — choose how many messages back it remembers (off / 2 / 4 / 6 / 8), not just one - New: handles typos better — sloppy kid-typing ("helo", "gud mornin") doesn't confuse it into thinking you asked about something it's never heard of - Vocabulary: 3,337 words (was 2,209) - Context window: 192 tokens (was 128) - Denser weight packing — 1.33 characters per weight instead of 2 - Much better at saying it doesn't know something and now names the specific unknown word instead of a generic non-answer - Still does math via <calc> tool calls, more accurate than 4 Mini's ★ BENCHMARK ★ MythicGPT Benchmark — 2,421 held-out questions, scored 0-100 (deliberately hard, so scores are low) - Overall score — 4 Pro: 38.6 / 4 mini: 29.1 - Perceived-smartness — 4 Pro: 51.5 / 4 mini: 42.5 - Coherence (never-garbles) — 4 Pro: 99.8 / 4 mini: 97.4 ★ THE STATS ★ - 10,001 blocks *o* - Parameters: ~1,278,000 - Architecture: 5-layer decoder-only Transformer, full-softmax output head, H=128 - Vocabulary: 3,337 words - Context window: 192 tokens - Tool calls: math - Quantization: int8 (dense packing, ~1.33 chars per weight) - Weights file: ~5.4 MB - .sb3 file: ~1.6 MB
