MythicGPT 4 releasing June 18th [DELAYED] (reasoning + tool calls) MythicGPT 4 mini will be noticeably smarter while being roughly the same speed or faster than Mythic 3 (the project). ✦ ━━━━━━━━━━━━━━━━━━━━━━ ✦ ⚡ M Y T H I C G P T 3 ⚡ FASTEST & SMARTEST AI LLM IN SCRATCH ★REAL AI, LIKE CHATGPT★ scroll down for more information ✦ ━━━━━━━━━━━━━━━━━━━━━━ ✦ ★RUN FULLY IN SCRATCH WITHOUT AN API★ Does NOT need Turbowarp, just run it directly in scratch :D (might run a bit slow for some people) !!! START A NEW CONVERSATION BY CLICKING THE GREEN FLAG IF THE AI STARTS ACTING WEIRD !!! ★★★ Pushing the limits of what is possible with ai fully inside Scratch. ★★★ There is only so much that I can add to this model. Right now, I focus on general logic "What color is the red box", but it only knows 1500 words, and due to the limitations of Scratch, there is only so much that I can do to improve it before it stops working entirely. ★FEATURES★ - Real-time chat, runs entirely inside a Scratch project - No internet, no servers - 2,000+ hand-crafted conversational patterns - Multi-turn context: remembers your last message (try "tell me about dogs" then "what about cats") - Riddles with reasoning, logic puzzles, geography depth, animal facts, science basics - Slang recognition: wbu, ty, idk, lol, nvm, bruh, omg, tbh, fr, and more - Anti-hallucination: graceful "I don't know" responses for off-corpus questions - Pen-rendered chat UI with colored speech bubbles and autoscroll - Live tokens/sec display in the top-left - "generating" placeholder while the AI is thinking - Punctuation tolerance: handles "hi!", "what's up?", "don't you know?" - Self-identifies as MythicGPT 3 ★UPGRADES OVER 2.1 PRO★ - 4x more parameters (~467K vs ~122K) - 6x more training pairs (~2,000 vs ~340) - 2-layer LSTM (vs single layer) - NEW: multi-turn conversation memory - NEW: riddles with reasoning traces ("what gets wetter the more it dries" → "a towel") - NEW: logic / if-then questions ("if today is monday what was yesterday") - NEW: geography depth — every country, capital, continent chained reasoning - NEW: pen-stamped bubble chat UI replacing the plain list monitor - NEW: tokens-per-second stats display + "generating" indicator - FIXED: first-token collapse — 2.1 Pro's Scratch if_else block bug made every response start with "hi" or "no" regardless of input. v3 always emits the correct first word. ★THE STATS★ Parameters: ~467,000 Vocabulary: 1,500 words Training pairs: ~2,000 unique (oversampled to ~52,000 training examples) Quantization: int8 (1 byte per weight) Weights file: 2.0 MB .sb3 file: 605 KB Max response length: 22 tokens Conversation memory: 1 message of prior context ★ARCHITECTURE★ - 2-layer LSTM with tied input/output embeddings - V=1500, H=130, ~467,000 parameters - Int8 per-tensor quantized weights stored as Scratch lists - Sigmoid / tanh approximated by 161-entry lookup tables - Word-level tokenizer with punctuation removal - Greedy argmax decoding with no-repeat-bigram filter - Multi-turn encoder input -------------------------------- im happy that the project finally got indexed by scratch search lol
