This Day in AI Podcast EP51: OpenAI's Sora, Gemini Pro 1.5 10M Context, ChatGPT Memory, GraphRAG, ChatRTX, Microsoft UFO...
8 snips
Feb 16, 2024 OpenAI's Sora, a system for generating 1 minute videos that track objects, is discussed, along with Google's Gemini 1.5. The potential of larger context windows in multi-modal inputs to replace RAG is explored. Microsoft's GraphRAG, which aims to improve RAG using knowledge graphs, is introduced. Nvidia's ChatRTX is tested on high-end graphics cards, and Microsoft UFO, an open-source project using vision AI, is mentioned. OpenAI's memory feature in ChatGPT is highlighted, as well as recent developments at OpenAI, including a researcher leaving and fundraising for chip development.
AI Snips
Chapters
Transcript
Episode notes
Sora Demonstrates Realistic Video Consistency
- OpenAI's Sora shows realistic, consistent object tracking and minute-long video generation that feels production-ready in many examples.
- The models still miss prompt details and exhibit artifacts, indicating cherry-picked demos and remaining limitations.
Memory Made Chatbot 'Haywire' In Practice
- Michael and Chris compare ChatGPT memory to SimTheory's long-running memory system and note similar features like selective storing and deletion.
- They describe how careless memory growth can make chatbots go 'haywire' and break character over time.
Consolidate And Decay Memories
- Consolidate and summarize memories over time so models prioritize important facts and avoid overwhelming the context window.
- Use decay and compression to reduce token usage while keeping reconstructable details available when needed.