OpenAI and the AI that may overtake humans
summary
OpenAI’s GPT-3 is a transformer trained on web crawl, books and Reddit and claims scale with 175 billion parameters; it demonstrates coherent chat, translation and also systematic failures like hallucinations, bias and arithmetic mistakes. A code-specialized fork Codex trains on 55 million public repositories (snapshot May 2020) and powers GitHub Copilot, which generates, refactors and inserts code via the playground and VS Code. DALL·E generates and edits images from prompts. Live demos produce working JavaScript UI and refactored algorithms but require human review.
AI-written summary · about
chapters (15)
- 0:00Intro
- 2:15International Speaker
- 6:55NLP Models comparison
- 9:47Conversational AI
- 12:44GPT-3 Model Stages
- 17:00GPT-3 Training DataSet
- 18:40Training phase
- 20:45Hyper-dimensional Graph
- 24:25Coding Tools
- 25:51GPT-3 can be used for...
- 28:37Codex Model
- 30:12GitHub Copilot Architecture
- 38:06Generating images
- 39:31Editing images
- 49:13OpenAI Playground
related talks
16mUnveiling the practical prowess of OpenAI's ChatGPT
51mBeyond the AI Hype: What's Real, What's Next
46mConversations with ChatGPT: illusion or reality?
25mThe Story of OpenAI GPT3 - Improve Observability with Tracing
43mBecome a 10x developer by harnessing the power of Generative AI
01h 01mKeynote: Transformers: The Rise of ChatGPT
31mAI-Driven Development in Go
01h 01mUnlocking the Potential of AI