AI/ML Digest | Issue 30

Welcome to the newest installment of Roosh Circle Digest, your go-to source for staying abreast of the latest developments in artificial intelligence and machine learning.

1/10 ?? A new prompting guide for Code Llama 70B Instruct has landed, offering a treasure trove of techniques and examples for developers to exploit the power of this open-source code model. Dive into the guide and elevate your coding prowess to new heights.

Check guide: [here]

2/10 ?? Media2Face is pushing the boundaries of 3D facial animation with multi-modality guidance. This innovation tackles the challenges of synthesizing facial expressions from speech despite the scarcity of high-quality 4D data. A boon for animators and AI enthusiasts alike seeking to breathe life into digital characters.

Read more: [here]

Project page: [here]

3/10 ?? Carnegie Mellon University's Open-World Mobile Manipulation System is a giant leap for robotics, venturing into the open world to master tasks like door-opening. This system is set to redefine realistic robotics applications, paving the way for robots to navigate our complex world.

Read more: [here]

4/10 ?? Over in San Francisco, Berkeley's AI lab has showcased a general-purpose humanoid robot adeptly navigating the urban jungle. This display of zero-shot reinforcement learning in action heralds a new era for robotics where adaptability and real-world application are front and center.

Video: [here]

5/10 ?? Introducing MultiPLY, a multisensory embodied LLM changing the game in 3D environments. By interacting with dynamic data across visual, auditory, and even thermal spectrums, this technology redefines our interaction with AI, offering a more immersive experience.

Watch: [here]

6/10 ?? Meta's new guide to prompt engineering is here to demystify the process for AI engineers and the curious-minded. This step-by-step manual highlights the art of crafting precise prompts to fine-tune AI models' performance and accuracy. It is a must-read for those looking to delve into the nuances of AI interactions.

Check: [here]

7/10 ?? Google's latest Chrome browser update is a triple treat of AI features. Manage your tabs easily, personalize your browser with AI themes, and get writing assistance on the fly. These enhancements promise a more streamlined and customized web browsing experience.

More: [here]

8/10 ?? Google's AI team has unveiled Lumiere, a "space-time" neural network that's set to revolutionize AI video generation. With its ability to craft clips in a single pass, we're on the brink of seeing text-to-movie functionalities by 2024. The future of video production is looking brighter than ever.

AI Video: [here]

9/10 ?? TikTok's Depth Anything is harnessing the untapped potential of unlabeled data for Monocular Depth estimation. Trained on an impressive dataset, this technology marks a significant stride in AI's ability to perceive depth, setting a new standard for image and video applications.

Check: [here]

10/10 ?? The concept of knowledge fusion in AI has leaped forward with FuseLLM. This technique aims to merge the prowess of multiple LLMs to create a supercharged AI with a profound understanding of language. It's a promising step for enhancing AI's natural language processing capabilities.

More: [here]

1/11 ??? DSPy is making waves in AI and application development with its revolutionary approach to chaining LLM calls, a technique that's stirring up excitement among tech aficionados and developers. Imagine constructing robust AI systems where one LLM feeds into the next – endless possibilities!

Read more: [here]

2/11 ?? The AI community is buzzing about proxy-tuning, a method that adapts LLMs without the fuss of changing model weights. This innovation simplifies the customization of language models, offering a new level of flexibility for AI developers.

Check: [here]

3/11 ?? Google Deepmind is breaking new ground by enabling robots to display expressive behaviors through large language models. With nods and gestures, robots become more communicative and coordinated, opening doors to more natural human-robot interactions.

Read more: [here]

Paper page: [here]

4/11 ??? X Corp. (formerly Twitter) is set to construct a $700 million AI data center in Atlanta following a $10 million tax break approval. This move signifies a major leap forward for AI infrastructure and innovation in the region.

Read: [here]

5/11 ?? Tesla's Dojo supercomputer is poised to supercharge Full Self-Driving software by processing massive amounts of video data.

Watch: [here]

6/11 ?? Google's Bard has received significant enhancements, with Gemini Pro now outperforming GPT-4 in rankings. s the AI community anticipates the release of Gemini Ultra, Google's advancements are setting new benchmarks in language models.

Check: [here]

7/11 ??? OpenAI's new 'GPT Mentions' feature for ChatGPT is here, allowing users to tag GPTs in conversations with an '@' symbol. Although still in limited rollout, this feature is expected to enhance user interaction with AI.

More: [here]

8/11 ??♂? WalkingRAG defies conventional AI performance improvement methods by not relying on reranking models and achieving superior outcomes. Their unique strategy has piqued the interest of many eager to understand the secret to their success.

Read: [here]

9/11 ?? Structured JSON output is now simpler than ever with the introduction of the JSONformer library! This innovative tool aims to streamline the process of generating structured JSON output from LLM, allowing developers to focus on content tokens and filling in fixed tokens. A real game-changer for AI engineers and non-tech individuals alike!

More: [here]

10/11 ?? LLM research has been off to a strong start in 2024, with several noteworthy papers published. These papers cover a range of topics and introduce innovative concepts that are shaping the future of AI. Stay tuned for more updates and exciting developments in the field of LLM research.

Check: [here]

Proxy-tuning - [here]

Self-rewarding LLMs - [here]

AlphaCodium - [here]

AlphaGeometry - [here]

MoE + Mamba - [here]

RAG vs. Finetuning in Agriculture - [here]

LLM-based evaluation survey - [here]

11/11 ?? AutoAct, an innovative method for training AI agents with limited data, is making waves in the AI industry. This groundbreaking approach, which surpasses existing techniques that rely on synthetic data, promises to revolutionize the way we train AI agents.

Details: [here]

1/11 ?? The realm of multimodal Language and Vision models (MM-LLMs) is buzzing with activity as an influx of research papers emerges. A pivotal survey summarizing 26 MM-LLMs stands out, providing a goldmine of insights and practical training recipes. This is a boon for AI engineers and curious minds eager to grasp the burgeoning capabilities of MM-LLMs.

Check: [here]

2/11 ?? Morpheus-1 is on the horizon, promising to be a linchpin in lucid dreaming technology. This multi-modal generative ultrasonic transformer is engineered to induce and stabilize lucid dreams, potentially transforming dream research. With a release slated for Spring 2024, beta users will soon navigate the uncharted waters of their subconscious.

Video: [here]

3/11 ?? The retrieval aspect of RAG systems is under the microscope, with a study pinpointing the necessity of placing pertinent information close to the query. This insight is crucial for AI developers aiming to refine their RAG models, ensuring that the system can effectively attend to and process the information required.

Read: [here]

4/11 ?? CircleCI's CTO, Rob Zuber, has unveiled a new course focusing on Automated Testing for LLMOps. This educational offering is a bridge between continuous integration principles and the development of LLM-based applications, providing a pathway for developers to bolster the reliability of their AI-driven apps.

More: [here]

5/11 ?? Pinterest has stepped into the spotlight with the introduction of RL Diffusion, a technique that refines pretrained Stable Diffusion models. This advancement is setting new standards in large-scale reinforcement learning, with a human preference rate of 80.3% for generated outputs.

Watch: [here]

Paper page: [here]

6/11 ?? An AI project has shattered the ceiling of classification tasks with long documents, achieving state-of-the-art results. With a DSPy program optimized with around 50 examples, this development is a testament to the evolving prowess of AI models in handling extensive classification challenges.

Check: [here]

7/11 ?? The AI landscape welcomes the Stable LM 2 1.6B Language Model, a compact yet powerful tool trained on multilingual data. Its release caters to a wide audience, overcoming hardware constraints with its small size and swift performance.

Details: [here]

8/11 ?? Wanderer, an AI-powered career exploration app, has tackled growth-induced cost challenges head-on. By implementing cost-saving measures that slash AI expenses by 99% without compromising quality or latency, the app continues to guide users on their professional journeys efficiently.

More: [here]

9/11 ?? A key insight has emerged in the AI/Machine Learning/Deep Learning industry, promising to be easily understood by anyone, even without a technical background.

Source: [here]

10/11 ?? Exciting updates have been introduced to RAG, enhancing its scalability for handling thousands of documents. Thanks to nerdai, users can now perform incremental updates across large data volumes seamlessly.

Read more: [here]

11/11 ?? Llama Index, in collaboration with nerdai, has significantly enhanced the speed of the RAG pipeline by parallelizing the ingestion process. This improvement results in impressive speed improvements, making data handling a breeze.

More: [here]

Data loading: [here]

Data transformations: [here]

1/12 ?? Bard's Advantage: Access to the Web Unveiled The AI landscape just got more intriguing with Bard's latest feature: web access. This AI service stands out from the crowd, offering users the unique capability to scour the web for information. Whether you're an AI engineer or just AI-curious, Bard's new function could be the tool you need to stay informed and ahead of the curve

More: [here]

2/12 ?? Arc Launches 'Arc Search', an AI-Powered Default Browser for iPhones iPhone users, get ready to elevate your browsing experience with 'Arc Search'. Arc's new AI-powered browser is designed to intuitively cater to your internet needs, streamlining your searches to match your preferences. Tailored results and a user-friendly interface could make 'Arc Search' your new go-to for web browsing on the go.

Check: [here]

3/12 ?? In-depth Analysis of AI Examples in LangSmith Dive into the world of AI with a detailed 20-minute YouTube analysis of AI examples in LangSmith. This video walks viewers through practical AI applications and the debugging process, providing a treasure trove of knowledge for both AI professionals and enthusiasts eager to deepen their understanding.

YouTube: [here]

4/12 ML Papers of the Week (6K??) For those hungry for the latest in LLM research, the "ML Papers of the Week" repository is a goldmine ?. It's a curated collection of the most impactful and trending machine learning papers, serving as a go-to resource for the AI community. This repository is perfect for anyone looking to keep up with the field's rapid advancements.

Read more: [here]

5/12 ?? Self-Rewarding LLMs: A Breakthrough in AI Judge Prompting The AI Judge Prompting arena is witnessing a revolution with the advent of Self-Rewarding LLMs. These LLMs are enhancing themselves with self-generated rewards, leading to improved output quality. With impressive scores on the AlpacaEval 2.0 leaderboard, this breakthrough is a beacon of potential for AI engineers and the non-tech curious alike.

Details: [here]

6/12 ?? Exciting Developments in LLM Research LLM research in 2024 is off to a robust start with several groundbreaking papers. From proxy-tuning and self-rewarding LLMs to AlphaCodium and AlphaGeometry, these papers are defining the future of AI. They tackle a range of topics, including geometric reasoning and robustness against adversarial attacks, showing the dynamic nature of LLM research.

More: [here]

7/12 ??? Open Source Tools for Large-Scale Data Processing and Model Training The open-source community has a treat in store with the release of datatrove and nanotron. These tools are designed for large-scale data processing and 3D parallel model training, promising to be a boon for AI engineers and those passionate about AI. Their lightweight and feature-rich nature could significantly enhance data handling and training processes.

Datatrove Github – all things webscale data processing: deduplication, filtering, tokenization: [here]

Nanotron Github - all things 3D parallelism: lightweight and fast LLM training: [here]

8/12 ?? AI Revolutionizing ETL Pipelines and ORMs AI is transforming ETL pipelines and ORMs, automating the generation of typespecs and DDLs with remarkable precision. This leap forward is a game-changer for businesses, offering a more streamlined approach to data integration and freeing up resources for strategic endeavors. It's a clear indicator of AI's growing impact across various sectors.

More: [here]

9/12 ?? New Short Course on LLMOps! LLMOps enthusiasts, there's a new short course on the horizon. Taught by Erwin from Google Cloud, this course brings MLOps principles to the world of large language models. It's an opportunity to stay at the forefront of LLMOps, enhancing your skillset in this emerging field.

Check: [here]

10/12 ???? The AI community is buzzing about a new concept in advanced RAG (Retrieval-Augmented Generation) - ensembling and fusion. This powerful technique combines multiple retrievers to enhance overall performance. To help AI engineers and non-tech individuals implement this technique, a comprehensive resource called LlamaPack has been introduced. So, if you're interested in improving the accuracy and effectiveness of your retrieval models, LlamaPack is worth exploring.

LlamaPack: [here]

Notebook Guide: [here]

11/12 ?? A new inference technique, IP-Adapter-FaceID Plus, is changing the game in the creation of personalized stock photos. Using advanced facial recognition technology, it allows anyone to create their own stock photos in seconds. This opens up exciting possibilities for both AI engineers and non-tech individuals.

More: [here]

12/12 ?? The integration of ChatGPT in spreadsheets is set to revolutionize data processing. With this advancement, users can generate content based on other cells and summarize information, making it easier to extract key insights from large datasets. This application of AI is predicted to gain widespread popularity by 2024.

Check: [here]

??Thanks to the joint efforts of our community, we have created a new Discord channel where we have collected 70+ AI tools for you.

We hope you find them not only useful but also inspiring for your future endeavors in the AI and ML space.