Discover more from Towards AI Newsletter
This AI newsletter is all you need #53
Our in-depth practical Generative AI course has already reached thousands of sign-ups!
We launched our practical Generative AI course last week and we are excited with the positive reception. Thousands of people have already signed up for the course and some people are already putting their new skills into practice, deploying their models on behalf of customers!
This course is the first of many by Towards AI and builds on our four-year history of providing educational and tutorial AI content. This first course was released in collaboration with Activeloop and the Intel Disruptor Initiative. The course focuses on LangChain and Deep Lake, a vector database for all AI data in production, and stands out with its depth and its focus on practical skills. With over 50 lessons including 10 practical projects across 8 modules and 5 full days of study time, the course offers a deep dive into AI technology, guiding you through creating innovative and production-grade AI tools and solutions.
Enroll today for free at learn.activeloop.ai.
What happened this week in AI by Louie
This week in AI developments, we highlight the acquisition of MosaicML by Databricks which we think demonstrates the rapid potential speed of startup maturity in AI as well as the potential for open-source models for LLMs.
Databricks, a data and AI company, announced its definitive agreement to join forces with MosaicML, a leading generative AI platform. This collaboration aims to democratize generative AI and empower organizations to build, own and secure generative AI models with their own data. MosaicML is behind leading open-source LLMs such as MPT-7B and MPT-30B, and is well known for its cost-effective model training.
The transaction was valued at approximately $1.3 billion and stands as a testament to the rapid pace of AI startups as the company has achieved an exit status just 2.5 years after it was founded and after raising just ~$64 million in capital. This acquisition follows recent significant AI startup funding rounds, such as Adept AI raising $350 million and Claude raising $450 million, and further emphasizes the substantial investments in this space and the amount of capital being invested to remain competitive. This acquisition also underscores how a small number of individuals can create wealth in the AI paradigm, as demonstrated by MosaicML with its 62 employees.
We believe Databricks' acquisition of MosaicML serves as validation for the open-source model of LLMs and where companies may choose to train their own models using customer-specific data to maintain a full understanding of the model training set and control of deployment.
- Louie Peters — Towards AI Co-founder and CEO
Thanks for reading Towards AI Newsletter! Subscribe for free to receive new posts and support my work.
Meta AI has created Voicebox, a new model that employs a Flow Matching approach to train on vast and varied datasets. This enables it to produce top-notch synthesized speech without the need for specific training. The model can match various audio styles, read text passages in multiple languages, and even edit speech segments within audio recordings. While the research paper and audio samples are accessible, the model and code are kept private to prevent misuse.
ElevenLabs, a voice technology research company and global leader in audio AI software, has secured $19 million in Series A funding for their AI-powered synthetic voice technology. This technology empowers content creators to effortlessly manage AI-generated audio content. The company has further plans to launch an AI dubbing tool and an AI Speech Classifier to ensure transparency and safety in generative media.
DeepMind's AI agent, RoboCat, can quickly adapt and enhance its skills with various robotic arms by generating new training data. In just 100 demonstrations, it can learn to operate new robotic arms within hours. The latest version of RoboCat has significantly improved, with its success rate on novel tasks more than doubling compared to its initial version, thanks to its growing experience.
DeepSpeed is an algorithm and system employed in training massive open models. To overcome existing limitations, Microsoft has introduced ZeRO++ with enhancements in memory, throughput, and usability.
Amazon aims to establish a unified gateway for businesses to access both open-source and closed-source generative AI models. In an interview with Matt Wood, AWS VP of Product, he delves into Amazon's perspective on the AI market, their strategy to outperform other tech giants in the AI competition, the future of the internet, and more.
Five 5-minute reads/videos to keep you learning
Last week, we announced our collaboration with Activeloop and the Intel Disruptor Initiative to create LangChain's free AI course, "LangChain & Vector Databases in Production." The course is designed to make AI practical and accessible, tailored to both experienced developers and enthusiasts. It offers 50+ lessons and 10+ projects covering API integration, prompt engineering, and production use.
Generative AI firms are enhancing their services to compete with Google, Microsoft, and Baidu. The article discusses the leading AI search companies, such as You.com and Perplexity.ai, in the consumer search space, and Vectara, Dashworks, Nuclia, Metaphor, and Hebbia in the enterprise space.
The article introduces a reference architecture for the LLM apps stack using in-context learning. It consists of three parts: data preprocessing and embeddings, prompt construction and retrieval, and prompt execution. In-context learning simplifies AI and is particularly useful for smaller datasets, enabling real-time data incorporation.
The article explains the concept of Key, Value, and Query in Attention using a potion metaphor to illustrate their role in predicting the next word in a text within the functioning of GPT. It aims to help readers develop a more intuitive understanding of GPT's inner workings from end to end.
Researchers are shifting focus from larger models to making AI models more efficient, considering the high making and deploying costs involved. This could lead to a new era of AI development, where models can achieve superhuman performance without requiring massive amounts of resources.
Papers & Repositories
This work suggests a roadmap to unify Language Models (LLMs) and Knowledge Graphs (KGs) in AI. It introduces KG-enhanced LLMs, LLM-augmented KGs, and Synergized LLMs + KGs frameworks to enhance knowledge representation and reasoning for improved downstream tasks.
This paper presents Phi-1, a new coding language model that leverages "textbook quality" data. Despite its smaller size, Phi-1 demonstrates remarkable accuracy on HumanEval and MBPP. It excels in Python coding and outperforms larger models by utilizing high-quality examples.
This paper introduces Wanda, a novel pruning approach for LLMs. Wanda achieves sparsity without requiring retraining or weight updates. It efficiently identifies efficient sparse networks from pre-trained models, surpassing pruning approaches and matching or exceeding the performance of other recent methods while minimizing computational costs.
This repository presents a fast method that matches the performance of the recently proposed Segment Anything (SAM) model. SAM is utilized in computer vision for tasks like image segmentation and captioning. However, this new technique employs a standard approach called instance segmentation, delivering comparable results at a speed 50 times faster.
This study reveals that while LLMs excel in various tasks, they face challenges with probabilistic reasoning, leading to performance limitations. To address this, the paper introduces TART, a solution that enhances an LLM's reasoning capabilities through a synthetically trained Transformer-based reasoning module.
Enjoy these papers and news summaries? Get a daily recap in your inbox!
The Learn AI Together Community section!
Weekly AI Podcas
In this week's episode of the "What's AI" podcast, Louis Bouchard interviews Jay Alammar, widely known in the AI and NLP field for his exceptional blog on transformers and attention. They delve into the world of Transformers, AI evolution, training steps of LLMs, and more. Jay shares insights on building LLM Apps and the challenges that accompany them. Tune in to the episode on YouTube, Spotify, or Apple Podcasts.
Upcoming Community Events
The Learn AI Together Discord community hosts weekly AI seminars to help the community learn from industry experts, ask questions, and get a deeper insight into the latest research in AI. Join us for free, interactive video sessions hosted live on Discord weekly by attending our upcoming events.
In critical domains like medical diagnoses and safety-critical systems, quantifying prediction uncertainty in machine learning is crucial. Conformal prediction offers a robust framework for this purpose. It allows the quantification of uncertainty for any machine learning model as a post-processing layer, without requiring model refitting. Join us for an upcoming talk where we delve into the applications of conformal prediction. Attendees are encouraged to familiarize themselves with the insights shared on the MLBoost YouTube channel before the event.
Join the event here and discover how conformal prediction enhances reliable decision-making by providing a measure of uncertainty beyond traditional point predictions.
Date & Time: 7th July 2023, 10:00 am EST
Meme of the week!
Meme shared by bigbuxchungus
Featured Community post from the Discord
Marcklingen has contributed to the development of langfuse, an open-source experimentation platform for LLM-based applications. Langfuse allows users to stay updated on output quality and feature trends, track token usage, segment executions conveniently, compare low-quality executions side-by-side, and more. Check it out on GitHub and support a fellow community member. Share your questions and feedback in the thread here.
AI poll of the week
TAI Curated section
Article of the week
This article tries to answer questions such as: What makes Vision Transformers (ViT) so special in comparison to their CNN counterparts, why these unique architectural features of ViTs are “breaking” almost all NPU and AI hardware accelerators targeting the high-performance edge market, and how Quadric’s Chimera GPNPU architecture can run ViTs today with real-time throughput at ~1.2W.
Our must-read articles
If you are interested in publishing with Towards AI, check our guidelines and sign up. We will publish your work to our network if it meets our editorial policies and standards.
Interested in sharing a job opportunity here? Contact email@example.com.
If you are preparing your next machine learning interview, don’t hesitate to check out our leading interview preparation website, confetti!
Thanks for reading Towards AI Newsletter! Subscribe for free to receive new posts and support my work.