Separating fact from fiction in a world of AI fairytales
summary
Covers why sensational claims (sentience, imminent AGI) overstate current models and shows concrete limits and fixes: GPT-4 writes a working but imperfect 3D JS game, yet fails many math puzzles and confidently hallucinates factual errors (one mayor wrongly reported jailed). Training details include CUDA GPUs, Common Crawl, GPT sizes (GPT-1 120M → GPT-4 ~1T est.), fine-tuning + reinforcement learning from human feedback, and memorisation of Codeforces problems (pre-training solved, post-training failed). Retrieval-augmented pipelines reduce hallucinations and improve currency.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:08AI hype cycle and generative models
- 3:40Neural network evolution
- 8:36Transformer models impact
- 11:07GPT models architecture
- 17:55Large model limitations
- 21:27AI intelligence or autocomplete
- 28:10Measuring artificial intelligence
- 35:08Chalearn AI intelligence framework
- 41:07Theory of mind and model brittleness
- 43:39Training data memorization
- 46:07AI hysteria dangers
- 47:51Hallucinations and bias solutions
- 51:23Model evaluation tools
- 53:31AI models future
- 55:00Conclusion and Q&A
related talks
46mConversations with ChatGPT: illusion or reality?
59mMirror, mirror: LLMs and the illusion of humanity
37mGenerative AI: Beyond (Just) Hype
58mMirror, mirror: LLMs and the illusion of humanity
17mPsychology of AI
01h 05mCan you trust your (large language) model?
19mLet's Talk GenAI
29mThe Rise of AI: A Reality Check on Generative AI