This AI newsletter is all you need #78

Author(s): Towards AI Editorial Team

Originally published on Towards AI.

What happened this week in AI by Louie

Towards AI, Activeloop, and the Intel Disruptor Initiative are excited to collaborate to release our “Retrieval Augmented Generation for Production with LlamaIndex and LangChain.” course, consisting of 10 lessons and 8 applied projects. This course is the third part of our Gen AI 360: Foundational Model Certification Course and follows the success of our “LangChain & Vector Databases In Production” and “Training and Fine-tuning LLMs for Production” courses.

The Retrieval Augmented Generation (RAG) for Production with LlamaIndex and LangChain course provides the theoretical knowledge and practical skills necessary to build advanced RAG products. You will start by learning the essential RAG tool stack, such as loading, indexing, storing, and querying in LangChain and LlamaIndex. We’ll also demystify the two libraries to help you select the right one when working with RAG or other LLM applications. You will then move towards more advanced RAG techniques aimed at surfacing and using more relevant information from the dataset. We cover techniques such as Query expansion, Transformation reranking, recursive retrieval, optimization, and production tips and techniques with LlamaIndex. We then progress to the exciting stuff: learning how to build RAG agents in LangChain and LlamaIndex, an introduction to OpenAI assistants, and some other tools & models that can be used in RAG products. We conclude with a summary of RAG evaluation techniques in LlamaIndex together with an introduction to LangSmith in LangChain.

Why This Course?

We believe that after recent AI progress, many human tasks across various industries can begin to be assisted with AI by combining LLMs, prompting, RAG, and fine-tuning workflows.

We are huge fans of RAG because it helps with:

1. Reducing hallucinations by limiting the LLM to answer based on existing documentation.
2. Helping with explainability, error checking, and copyright issues by clearly referencing its sources for each comment
3. Giving private/specific or more up-to-date data to the LLM,
4. Not relying on black box LLM training/fine tuning for what the models know and have memorized.

However, RAG is still a young and emerging field, and RAG-based products still need a lot of experimentation with various methods to retrieve and use the most useful data for your use case. We are pleased to introduce some of the latest advanced RAG techniques with this course.

Who is this for?

Whether you are planning to build a chat with data application for your organization or learning how to leverage Generative AI in various industries and build a robust AI portfolio, this course is for you. And we made it free!

Find more information about the course in this video introduction.

Join the course here: https://learn.activeloop.ai/courses/rag

– Louie Peters — Towards AI Co-founder and CEO

Hottest News

1.DeepMind AI Outdoes Human Mathematicians on Unsolved Problems

DeepMind’s FunSearch AI has surpassed human mathematicians by solving a previously unsolved math problem related to combinatorics. FunSearch utilizes large language models to generate effective solutions and constantly improves through testing and feedback.

2. Phi-2: The Surprising Power of Small Language Models

Microsoft has released Phi-2, a language model with 2.7 billion parameters that outperforms models up to 25 times its size. Phi-2 achieves remarkable reasoning and language understanding abilities using high-quality data and synthetic datasets. Phi-2 is available in the Azure AI Studio model catalog.

3. Google Releases the Gemini Pro API

Google has launched the Gemini Pro API, allowing developers and enterprises in the AI field to experiment and build upon their specific use cases. The current version has a 32K context window for text and is free to use.

4. BioCLIP: A Vision Foundation Model for the Tree of Life

BioCLIP is the first large-scale multimodal model that utilizes images and structured biological knowledge to answer biology questions. By training a vision encoder on the Tree of Life taxonomy, BioCLIP enhances hierarchical understanding of the natural world.

5. Introducing gigaGPT: GPT-3 Sized Models in 565 Lines of Code

Cerebras has released gigaGPT, a model implementation similar to nanoGPT but with over 100 billion parameters. By leveraging Cerebras hardware and different optimizers, gigaGPT overcomes the limitations of GPU memory and the need for complex scaling frameworks, offering a simplified approach for training large models.

Do you believe DeepMind’s AI can replace human mathematicians or act as a force multiplier for them? Share your thoughts in the comments below!

Five 5-minute reads/videos to keep you learning

1.Mixture of Experts Explained

Mixture of Experts (MoEs) models is a technique for large model pretraining that uses sparse MoE layers and gate networks. This blog post looks into the building blocks of MoEs, their training, and the tradeoffs to consider when serving them for inference.

2. Multi-Modal RAG on Slide Decks

Retrieval Augmented Generation (RAG) now supports images, enhancing performance in multi-modal AI applications. This article shares a public benchmark for evaluating RAG on slide decks and provides a template for quickly creating multi-modal RAG apps for slide decks.

3. Benchmarking RAG on Tables

This blog discusses strategies for managing semi-structured RAG, including a mix of unstructured text and structured tables. It proposes three strategies: using long-context LLMs but risking quality degradation, employing targeted approaches to detect and extract tables accurately, and splitting documents to preserve table elements within text chunks.

4. The LLM OS: A Glimpse Into the Future of Tech With Andrej Karpathy

AI visionary Andrej Karpathy has floated a novel concept: the LLM Operating System. This article allows a peek into a future where our devices understand us as humans do — perhaps even better.

5. Top 10 Influential AI Research Papers in 2023 From Google, Meta, Microsoft, and More

The article highlights the top 10 AI research papers of 2023. The papers cover many topics, including advancements in language models, multimodal models, image segmentation, text-to-image models, video editing, and world model-based algorithms.

Repositories & Tools

1. PyApp is a wrapper for Python applications that bootstrap themselves at runtime.
2. Amphion is a toolkit for Audio, Music, and Speech Generation.
3. Lobe Chat is an open-source chatbot framework that supports speech synthesis, multimodality, and extensible Function Call plugin systems.
4. Azure AI Studio is a platform for developing generative AI solutions and custom copilots.
5. DryRun Security is an AI-powered security buddy that adds security context to coding.

Top Papers of The Week

1.Are Emergent Abilities of Large Language Models a Mirage?

Emergent abilities in AI models may not be as significant as they appear, as they can be influenced by how they are measured. Researchers have shown that these abilities may vanish or be less pronounced when using different metrics or improved statistics. This suggests that emergent abilities may not be an inherent characteristic of scaling AI models.

2. Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models

Researchers have developed a simplified self-training method called ReST^EM, which uses expectation maximization to fine-tune LLMs. This approach, which incorporates binary feedback, outperforms strategies that solely rely on human data for fine-tuning. The study demonstrates that ReST^EM scales effectively with the model’s size and suggests that self-training with feedback can potentially reduce reliance on human-generated data in AI applications.

3. LLM360: Towards Fully Transparent Open-Source LLMs

LLM360 is an initiative to fully open-source LLMs, which advocates for all training code and data, model checkpoints, and intermediate results to be made available to the community. The goal of LLM360 is to support open and collaborative AI research by making LLM training transparent and reproducible to everyone.

4. ReST meets ReAct: Self-Improvement for Multi-Step Reasoning LLM Agent

To address the current deficiencies of LLMs in answering complex natural language questions, researchers defined a ReAct-style LLM agent with the ability to reason and act upon external knowledge. The agent is refined through a ReST-like method that iteratively trains on previous trajectories.

5. Distributed Inference and Fine-tuning of Large Language Models Over The Internet

Researchers have demonstrated that large language models can efficiently run on geo-distributed devices over a consumer-grade network, opening up possibilities for pooling compute resources from multiple research groups.

Quick Links

1. Mistral released Mixtral 8x7B, a high-quality sparse mixture of expert models (SMoE) with open weights. It excels in code generation and supports multiple languages, surpassing Llama 2 70B in benchmarks.

2. Deci AI Unveils DeciLM-7B: A Leap Forward in Language Model Performance and Inference Cost Efficiency. It is a 7 billion parameter model, outperforming open-source models such as Llama2 7B and Mistral 7B in accuracy and efficiency.

3. Lightning AI, creator of PyTorch Lightning framework, debuts the platform for building and deploying AI apps to simplify how enterprises can build and deploy AI-infused applications.

Who’s Hiring in AI

Research Scientist, Programming Systems — New College Grad 2024 @NVIDIA (Remote)

Machine Learning Developer @Willow (Sydney, Australia)

Senior Data Analyst @BlueFlag LLP (Remote)

Sr. UX Designer, AI/ML @Amazon (Seattle, WA, USA)

AI Computing Performance Architect @NVIDIA (Remote)

Product Manager, Machine Learning @Augmedix (San Francisco, CA, USA)

Site Reliability Engineer — US @Weights & Biases (Remote)

Interested in sharing a job opportunity here? Contact [email protected].

If you are preparing your next machine learning interview, don’t hesitate to check out our leading interview preparation website, confetti!

Think a friend would enjoy this too? Share the newsletter and let them join the conversation.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI