AI Roundup: MetaGPT, Style Imitation, Plugins, HVAC | Week 33, 2023
AI Roundup: MetaGPT, Style Imitation, Plugins, HVAC | Week 33, 2023
Week 33, 2023 AI Roundup
In this weekly roundup, let’s take a look at:
Using MetaGPT to build your own software company
MetaGPT is a new multi-agent collaboration approach that uses meta-programming and human workflows to solve complex tasks. It encodes SOPs for structured coordination and assigns roles to agents, mirroring an assembly line. MetaGPT outperforms other methods in problem-solving and is cost-effective, costing $0.2 for an example and $2.0 for a full project.
Read more here.
Controlling HVAC with GPT-4
Microsoft Research conducted an experiment using GPT-4, a leading AI model, to control HVAC systems. GPT-4 was tasked with maintaining a building's temperature at 22 degrees Celsius. The AI model was trained with expert demonstrations and used prompts to suggest actions for temperature control. Initial results showed GPT-4's ability to learn, adapt, and optimize performance, even in noisy external conditions. GPT-4's efficiency, adaptability, and expert decision-making offer significant advantages over traditional HVAC control methods, potentially revolutionizing the industry.
Read more here.
Teaching LLMs to imitate your writing style and voice
From researchers at Google and University of Michigan: A multistage and multitask approach to train large language models for personalized text generation. Using five stages and the user's previous writings, the system can generate content in the user's unique style. This method shows significant improvement over previous approaches and has the potential for creating content tailored to specific audiences and needs.
Read more here.
Fine-tuning LLMs with Open-Platypus
The research paper "Platypus: Quick, Cheap, and Powerful Refinement of LLMs" introduces Platypus, a family of fine-tuned and merged Large Language Models (LLMs) that achieve top performance on HuggingFace's Open LLM Leaderboard. Platypus uses the Open-Platypus dataset, a curated collection of public text datasets focused on STEM and logic, to fine-tune the LLaMa 2 model, a 70B parameter LLM released by Meta. The fine-tuning process is efficient, with a 13B Platypus model trainable on a single A100 GPU using 15k questions in just 5 hours.
Read more here.
Gorilla (fine-tuned LLaMa) to make better API calls
Gorilla, a fine-tuned LLaMA-based model, excels at writing API calls, outperforming even advanced LLMs like GPT-4. It mitigates the issue of hallucination in API calls, a common problem with LLMs. Gorilla's success lies in its unique methodology, using self-instruct fine-tuning and retrieval on the APIBench dataset, a large collection of APIs. Evaluation results show significant improvements over other models, especially when using a ground truth retriever. However, there's room for improvement in retriever accuracy. Gorilla's adaptability to changes in API documentation is another standout feature, making it a game-changer in the world of LLMs and API calls.
Read more here.