Tips For Detecting The Use Of AI-Generated Text
summary
Demonstrates that shallow neural networks trained with a single pass and no gradient descent can produce workable language models that train on CPUs for far lower cost than large Transformer runs (GPT-3 training is cited as using 10,000 GPUs and costing $10–20M). Provides prototype software—IMAN and an HR-BPE tokenizer—that trains locally, shows pretraining + transfer is feasible, and reports bot-detection wins: text-only generative features match Botometer-level performance as of 2021 and empirical measures find circulating follower accounts (many cycling ~20 times) tied to spike/sawtooth follower signatures.
AI-written summary · about
chapters (6)
- 0:00Introduction and motivation
- 10:23Bot detection with generative features
- 21:53Tracking social botnets on Twitter
- 30:11AI hallucinations and adversarial triggers
- 38:58Interpretable AI and efficient models
- 45:05Q&A
related talks
47mThe Rise of AI: A Reality Check on Generative AI
40mAI vs AI: Creating Novel Spam and Catching it with Text Generating AI
19mLet's Talk GenAI
25mGenerative Text Modelling: Scratching the surface
39mBuilding Your Own Bot with GPT: Harnessing the Power of AI...
52mBeyond the AI Hype: What's Real, What's Next
28mGPT or how I learned to stop worrying and love the generative AI
02h 50mUnderstanding and Using AI Language Models