Initial expectations of a GPT Store full of low-quality AI were disproven by the presence of highly sophisticated, custom GPTs and recent improvements like GPT-4o.
Metacognition is susceptible to stochasticity
Sam Schillace, Deputy CTO of Microsoft, emphasizes the importance of using code for planning when building with AI. He advises that while LLMs excel in thinking, their planning capabilities are weak, making them unpredictable. Thus, integrating structured code enhances predictability and effectiveness.
Retrieval Unleashed with OpenAI’s new Assistants API
OpenAI’s updated Assistants API introduces advanced features like the file_search tool which optimizes query processing through parallel searches and relevant reranking, handling up to 10,000 files per assistant.
ChatGPT and Memory
OpenAI incorporates “memory” into ChatGPT, allowing user details to persist between sessions.
The Point of LangChain with Harrison Chase of LangChain
Harrison Chase introduced LangChain on The AI Engineer Podcast, emphasizing its utility for building context-aware reasoning applications using chains and language model agents. Excited about elements like plan-and-execute agents and MultiOn, Harrison addressed comments about the predominance of prompt tuning and data serialization in the workflow. The system works best with OpenAI models, with supportive tools like jsonformer and LCEL available.