OpenAI has recently announced the development of SearchGPT, a groundbreaking prototype that revolutionizes how users search for information online. This new AI-driven search feature combines the strengths of OpenAI’s conversational models with real-time web data, promising to deliver fast, accurate, and contextually relevant answers. SearchGPT is currently in a testing phase and is available to…
Designing computational workflows for AI applications, such as chatbots and coding assistants, is complex due to the need to manage numerous heterogeneous parameters, such as prompts and ML hyper-parameters. Post-deployment errors require manual updates, adding to the challenge. The study explores optimization problems aimed at automating the design and updating of these workflows. Given their…
Large Language Models (LLMs) are a subset of artificial intelligence focusing on understanding and generating human language. These models leverage complex architectures to comprehend and produce human-like text, facilitating applications in customer service, content creation, and beyond. A major challenge with LLMs is their efficiency when processing long texts. The Transformer architecture they use has…
In the rapidly evolving field of natural language processing (NLP), integrating external knowledge bases through Retrieval-Augmented Generation (RAG) systems represents a significant leap forward. These systems leverage dense retrievers to pull relevant information, which large language models (LLMs) then utilize to generate responses. However, while RAG systems have improved the performance of LLMs across various…
The number of academic papers released daily is increasing, making it difficult for researchers to track all the latest innovations. Automating the data extraction process, especially from tables and figures, can allow researchers to focus on data analysis and interpretation rather than manual data extraction. With quicker access to relevant data, researchers can accelerate the…
Developing AI agents that can autonomously perform a wide variety of tasks with the same flexibility and capability as human software developers presents a significant challenge. These tasks include writing and executing code, interacting with command lines, and browsing the web. Current AI agents often lack the necessary adaptability and generalization for such diverse and…
The rapid advancements in Generative AI have underscored the importance of text embeddings. These embeddings transform textual data into dense vector representations, enabling models to efficiently process text, images, audio, and other data types. Various embedding libraries have emerged as front-runners in this domain, each with unique strengths and limitations. Let’s compare 15 popular embedding…
Reinforcement Learning (RL) finetuning is an important step in training language models (LMs) to behave in specific ways and follow human etiquette. In today’s applications, RL finetuning involves multiple goals due to various human preferences and uses. The multi-objective finetuning (MOFT) is needed to train a multi-objective LM to overcome the limitations of single-objective finetuning…
Parameter-efficient fine-tuning (PEFT) methods have become essential in machine learning. They allow large models to adapt to new tasks without extensive computational resources. By fine-tuning only a small subset of parameters while keeping most of the model frozen, PEFT methods aim to make the adaptation process more efficient and accessible. This approach is crucial for…
Recent advancements in LLM capabilities have increased their usability by enabling them to do a broader range of general activities autonomously. The existing methods for expressing and running LM programs could be more efficient, although they are widely used. There are two main obstacles to effective LM program utilization: The non-deterministic character of LLMs makes…