A silly conversational AI that runs entirely in Scratch 3 blocks. Model weights: 4.4M weights (linear + neural network) are quantized to 36 levels, one base-36 character each. Small weights are set to zero (within ±3 steps for linear, ±2 for the MLP), and runs of zeros are run-length encoded. That's 2.5 MB instead of about 48 MB as float JSON. Using Blorp Save the project as a remix to keep Blorp's memory. The brain-jar, past topics and chat counter are stored in the project. Next time, Blorp greets you by name, remembers what you talked about, wishes you happy birthday on the right day, and checks in if you were feeling down. Things to try: "my name is …" "i have a dog named …" "my favorite food is …" "remember that …" "what do you remember about me" "what did we talk about?" "forget my name" Knowledge "what is a narwhal?" "tell me about saturn" "tell me more" (491 topics, 1,473 facts) Opinions "do you like broccoli?" "what do you think about minecraft?" "cats or dogs?" "do you like them?" Games "knock knock" "tell me a knock knock joke" "riddle" "rock paper scissors" "guess the number" "would you rather" Other jokes, fun facts, stories, songs, tongue twisters, "what is 12 times 7", "what time is it", "ask me a question"
Blorp was trained on about 13,900 example messages, including about 2,100 context-dependent replies. On held-out data (averaged over 3 splits) it gets 83% accuracy, 91% precision when it's confident, and 85–86% on context-dependent replies. A typical message takes about 50 ms in the official Scratch engine, and never more than 0.5 s. Messages about self-harm always get a caring reply that points to a trusted adult or the 988 helpline. This works through about 40 keyword phrases as well as the classifier, so it doesn't depend on the model. A 101-word blocklist stops Blorp from echoing, saving or giving opinions on those words; it changes the subject instead. Mild swearing gets the "squishy ears" reply. Memory wipes only happen on explicit requests. Blorp's lexicon was grown by about 2,500 words using GloVe 6B word vectors (Stanford NLP, from Wikipedia and news text). A frequent word joins a concept only when at least 5 of its 7 nearest already-labeled neighbors agree. This mainly widens the opinion engine and relevance ranking: "do you like kayaking?" gets a sports opinion.
