Today’s Newsletter Contains:

• Voyager: Revolutionizing Minecraft Gameplay with GPT-4 and Lifelong Learning

• New King of Open-source LLMs: Falcon 40B Beats LLaMa 65B

• Microsoft Boosts Windows 11 with Copilot AI Assistant and Dev Home for Streamlined Workflows

• QLoRA: Efficient Fine-tuning of Quantized LLMs on Consumer GPUs

• Breaking Language Barriers: Meta MMS Enables Speech Recognition across 1,100 Languages

• Awesome AI Tools

• 6 New Learning Resources

Voyager: Revolutionizing Minecraft Gameplay with GPT-4 and Lifelong Learning

Researchers from Nvidia, Caltech, UT Austin, Stanford, and ASU have introduced Voyager, a groundbreaking lifelong learning agent in Minecraft utilizing GPT-4. The agent continuously explores the world, acquires a diverse array of skills, and makes novel discoveries without human intervention.

In experiments, Voyager outperformed counterpart agents, progressing through the Minecraft tech tree significantly faster and exhibiting exceptional exploration capabilities. Voyager serves as a foundation for developing powerful generalist agents without fine-tuning model parameters.

New King of Open-source LLMs: Falcon 40B Beats LLaMa 65B

The Technology Innovation Institute (TII) in Abu Dhabi has developed FalconLM. It has 40 Billion parameters, but it beats LLaMa 65 Billion parameters model. It holds the top two positions on the Hugging Face OpenLLM Leaderboard currently.

FalconLM uses a high-quality dataset combined with an optimized architecture. As a result, it requires only 75% of the computational effort compared to GPT-3 during training and offers inference costs one-fifth of GPT-3.

Microsoft Boosts Windows 11 with Copilot AI Assistant and Dev Home for Streamlined Workflows

Microsoft announced Windows Copilot, an AI assistant built into Windows 11 designed to turn every user into a "power user." Accessible through a taskbar and available across applications, Copilot offers AI assistance, summaries, paraphrases, and content explanations.

They also introduced Dev Home, a new developer experience to streamline workflows and improve productivity. Additional features include unattended dev machine setup using WinGet configuration and Dev Drive, a storage volume tailored for developers offering improved performance and security.

Dev Home incorporates a customizable dashboard for tracking project information and enables easier collaboration with GitHub widgets for tracking coding tasks, pull requests, and performance metrics.

QLoRA: Efficient Finetuning of Quantized LLMs on Consumer GPUs

QLoRA is a groundbreaking finetuning approach that enables finetuning 65B parameter models on a single 48GB GPU, while preserving 16-bit task performance.

QLoRA introduces innovations like 4-bit NormalFloat, double quantization, and paged optimizers to save memory without sacrificing performance.

QLoRA-trained Guanaco models outperform previous open-source models on the Vicuna benchmark, attaining 99.3% of ChatGPT's performance with just 24 hours of finetuning on a single GPU.

Breaking Language Barriers: Meta MMS Enables Speech Recognition across 1,100 Languages

Meta AI has released Massively Multilingual Speech (MMS) project, which uses wav2vec 2.0 and a new dataset to provide speech-to-text, text-to-speech, and language identification for over 1,100 languages, including many at-risk dialects.

It achieves half the word error rate of OpenAI's Whisper, and preserves linguistic diversity for at-risk dialects.

By supporting thousands of languages and outperforming existing models, MMS aims to preserve linguistic diversity and make information more accessible to a wider audience. MMS is a crucial step toward a future where a single model can handle various speech tasks for all languages.

Awesome AI Tools

• Codium (LINK) - A Free AI-Powered Toolkit for Developers. In-house models and infrastructure, not another API wrapper. Extensions in all your IDEs. Autocomplete and Search, with more coming.

• PlaygroundAI (LINK) - A free-to-use online AI Image generateor and editor. Supports Stable Diffusion, Dalle, and custom stable diffusion models like RevAnimated, Protogen, RPG 4 etc.

• Poe (LINK) - Poe lets you ask questions, get instant answers, and have back-and-forth conversations with AI. Gives free access to ChatGPT, gpt-3.5-turbo, Claude Instant and a lot of custom Chatbots.

• Fine-Tuner.ai (LINK) - Build sophisticated, tailored AI agents without technical skills or coding - just bring your data and ideas.

• PrivateGPT (LINK) - Your offline personal AI Assistant. Chat with your data securely without privacy issue.

Learning Resources

• Practical Deep Learning (LINK) - A free course designed for people with some coding experience, who want to learn how to apply deep learning and machine learning to practical problems.

• Prompt Engineering Guide (LINK) - An awesome free course to learn advanced prompt engineering. Covers basic as well as advanced techniques.

• ChatGPT Prompt Engineering for Developers (LINK) - Free and beginner-friendly course by Andrew Ng in partnership with OpenAI.

• LangChain for LLM Application Development (LINK) - Learn to make AI powered apps with Langchain.

• How Diffusion Models Work (LINK) - This technical course teaches the concepts behind diffusion-based image generation.

• Building Systems with the ChatGPT API (LINK) - In this course, taught by OpenAI’s Isa Fulford together with Andrew Ng, you’ll learn to build complex applications using large language models (LLMs).

That's all for now!

If you have any query or interesting ideas, please reach out to me by responding to this email or by sending me a DM on Twitter: @ShivamKumar212

As always, thanks for reading, and see you next time. 🙂

If you like this, share it with your friends.