Martin Förtsch, Thomas Endres & Jonas Mayer
summary
Claim: an ironic system that automates trash‑posting and replies on social media by fine‑tuning language models and synthesizing many voices. Implementation details: GPT‑2 fine‑tuned on scraped Reddit + Twitter (bucketed daily; Twitter API ~500k tweets/month), tokenization/attention explained, model sizes from 500MB to 6GB considered. Audio stack: multi‑speaker Tacotron / Coqui TTS producing varied voices. Hardware: five glass robot heads running client RPis with a central GPU server; demo runs and generates plausible but drifting and often toxic comments.
AI-written summary · about
chapters (12)
- 0:00Posting AI project intro
- 5:32Online debates to automated discussions
- 9:02Natural language processing fundamentals
- 21:58Generating comments with GPT
- 30:32GPT's transformer architecture
- 33:43Training data and model selection
- 38:04Demo: generated comments
- 39:49Human-like robot voices
- 42:45Building robot heads
- 44:10Connecting AI to robot heads
- 45:16Live demo: Posting AI robots
- 47:04Summary and future
related talks
01h 03mIt's All About AI
45mArtificial Irreverence: Improvised Theatre with Robots
40mHow I Made My Computer Write it's First Short Story
01h 22mWho needs ChatGPT? Rock solid AI pipelines with Hugging Face and Kedro
35mSpeech Synthesis with Tacotron2 and PyTorch
47mThe Digital Hearth: An In-the-Wild Automated Group Speech to Text System
22mBuilding next-gen speech synthesis for Bookmate audiobook service
01h 01mKeynote: Transformers: The Rise of ChatGPT