Talking to my past self (without introducing temporal paradoxes!)
summary
Claims a working pipeline that trains a seq2seq RNN in TensorFlow on archived chat logs, specifically an MSN-era bot and a modern Hangouts bot, using epochs of training and sampling. Built a pair of conversational agents that reproduce typos, teenage style and odd political quirks; output is often syntactically plausible but semantically silly, remembers almost nothing long term, fails at consistent personality, and produces amusing exchanges rather than passing the Turing test.
AI-written summary · about
chapters (6)
- 0:00MSN Messenger to chatbot
- 1:48Training bot to sound like me
- 3:54Teaching bot to answer questions
- 6:09High school me meets modern me
- 8:51Witty self-conversations impossible
- 10:15Lessons from experiment
related talks
01h 27mDeprecating the state machine: building conversational AI with the Rasa stack
41mConversational Agents at Scale: Retrieval and Generative approaches
30mConversational AI: Building clever chatbots
01h 38mConversational AI with Rasa Core & NLU
55mEveryday AI: Practical Applications Beyond the Chatbot
46mConversations with ChatGPT: illusion or reality?
46mGenerative Chatbot with tensorflow
22mVoice Agents That Think While They Talk