AI Distilled 35: Building LLMs, Rust Inference, SpeechGPT-Gen, and 3D

👋 Hello ,

“[AI healthcare systems] include understanding symptoms and conditions better, including simplifying explanations in local vernaculars…and acting as a valuable second opinion. This, in turn, potentially provides a pathway for medical AI towards superhuman diagnostic performance.” - Vivek Natrajan, Google AI Researcher 

AI is making rapid strides in healthcare with a major disruption underway. The recent study examined how current language models perform in diagnostic settings and it’s clear they’re making tremendous progress with each passing day. 

Step into AI_Distilled's latest installment, where we delve into the most recent developments in AI/ML, LLMs, NLP, GPT, and Gen AI. Join us as we kick off this edition, showcasing news and insights from various sectors: 

AI Industry Updates:  

Apple Focuses on AI for Major iOS Update 

Tencent Chief Raises Concerns Over Gaming Business 

China Greenlights Over Four Dozen AI Systems 

New iPhone App Streamlines Web Searches with AI Assistance 

German Startup Aims to Supercharge AI Chips with Novel "Memcapacitor" Design 

AI Startup Anthropic Suffers Data Breach Amid Regulatory Scrutiny 

New AI Coding Tool Surpasses Previous Models 

New AI Model Launches:  

Meta Unleashes Mighty Code Llama 70B 

Google Unveils New AI Model for Advanced Video Generation 

OpenAI Announces Major Model and API Updates 

AI in Healthcare: 

AI Helps Design Proteins for Improved Gene Therapy Delivery 

AI's Diagnostic Prowess Advancing Rapidly 

AI in Supply Chain Management: 

AI Transforming Global Supply Chain Management 

Edge AI's Potential in Logistics Faces Memory Limitations 

We’ve also got you your fresh dose of LLM, GPT, and Gen AI secret knowledge and tutorials: 

Assessing AI Accuracy: New Leaderboard Evaluates Language Models 

New Techniques to Boost AI Performance 

Neural Networks Learn Non-Linear Functions Through Rectified Activation 

Enhancing Conversations with Contextual Understanding 

We know how much you love hands-on tips and strategies from the community, so here they are: 

Fine-Tuning BERT for Long-Form Text 

Training Giant Language Models the Cost-Effective Way 

Optimizing AI Models for Speed and Scale 

Estimating Depth from a Single Image: Researchers Evaluate AI Models 

Don’t forget to review these GitHub repositories that have been doing rounds: 

rasbt/LLMs-from-scratch 

LaurentMazare/mamba.rs 

0nutation/speechgpt 

naver/roma 

📥 Feedback on the Weekly Edition

Take our weekly survey and get a free PDF copy of our best-selling book, "Interactive Data Visualization with Python - Second Edition." 

📣 And here's the twist – we're tuning into YOUR frequency! Inspired by a reader's request, we're launching a column just for you. Got a burning question or a topic you're itching to dive into? Drop your suggestions in our content box – because your journey of discovery is our blueprint.

We appreciate your input and hope you enjoy the book! 

Writer’s Credit: Special shout-out to Vidhu Jain for their valuable contribution to this week’s newsletter content!  

Cheers,  

Merlyn Shelley  

Editor-in-Chief, Packt 

🗝️ Unlock the Packt library for FREE

Dive into a world of endless knowledge with our 7-day FREE trial! Discover over 7,500 tech books and videos with your Packt subscription and stay ahead in your field.

Plus, check out our ✨NEW feature: the AI Assistant (beta) ✨, available across eBook, print, and subscription formats.  

Don't miss your chance to explore and innovate – start your free trial today and unlock your tech potential! 

SignUp | Advertise | Archives

⚡ TechWave: AI/GPT News & Analysis

💎 Apple Focuses on AI for Major iOS Update: Apple is planning a major iOS 18 update, aiming to enhance AI features, particularly Siri and messaging apps. It will include generative AI in various apps, with Apple in talks with publishers for content. New iPads and MacBooks with M3 chips may be released soon. 

💎 Tencent Chief Raises Concerns Over Gaming Business: Tencent's CEO, Pony Ma, is concerned about competition in the gaming division as recent titles underperformed while rivals thrive. Despite 30% of revenue from games, Tencent feels left behind. They've caught up in AI and plan to integrate it across businesses, with a focus on evolving WeChat platform. 

💎 China Greenlights Over Four Dozen AI Systems: Chinese regulators have recently granted approval to over 40 AI models for public use, with companies including Baidu, Alibaba, and ByteDance receiving the green light. This move reflects China's push to narrow the AI gap with the United States, driven by the success of ChatGPT. 

💎 New iPhone App Streamlines Web Searches with AI Assistance: Arc Search is a mobile app improving search with "Browse for Me" creating concise web pages from multiple sources, AI summaries, tab switching, and reading mode. It's an early-stage tool designed to save users time in the digital era. 

💎 German Startup Aims to Supercharge AI Chips with Novel "Memcapacitor" Design: Semron, a German startup, aims to revolutionize the computer chip industry by replacing transistors with "memcapacitors." Their 3D-stacked memcapacitor design promises improved AI performance and reduced energy consumption. With $10 million in funding, Semron aims to provide efficient chips for mobile and edge AI applications. 

💎 AI Startup Anthropic Suffers Data Breach Amid Regulatory Scrutiny: Anthropic, the creator of LLM, disclosed an accidental data leak by a contractor to a third party. Concurrently, the FTC initiated an inquiry into Anthropic's affiliations with Amazon and Google, fueling concerns about data privacy in the expanding LLM landscape. 

💎 New AI Coding Tool Surpasses Previous Models: AlphaCodium, an open-source AI model by CodiumAI, surpasses Google's AlphaCode and AlphaCode 2 in code generation efficiency. It employs a "flow engineering" approach, emphasizing code integrity through iterative code generation and adversarial testing, aiming to advance upon DeepMind's work and assist global developers.  

New AI Model Launches: 

💎 Google Unveils New AI Model for Advanced Video Generation: Google has unveiled Lumiere, an advanced AI model utilizing space-time diffusion to create lifelike videos from text or images. It addresses motion and consistency problems in AI video generation, yielding 5-second clips with smooth motion. While not yet available for testing, experts anticipate Lumiere could redefine AI video creation capabilities. 

💎 Meta Unleashes Mighty Code Llama 70B: Meta has launched Code Llama 70B, a powerful AI model for automating software development. With 500 billion training examples, it excels in translating natural language instructions into functional code, surpassing previous models. Meta aims to boost global coding and expand its use in translation, documentation, and debugging. 

💎 OpenAI Announces Major Model and API Updates: OpenAI launched new AI models and tools, including enhanced embeddings, GPT-4, and moderation models. They lowered GPT-3.5 Turbo pricing, introduced key management features, and per-key analytics to enhance user control and accessibility, bolstering their technology's capabilities. 

AI in Healthcare: 

💎 AI Helps Design Proteins for Improved Gene Therapy Delivery: University of Toronto scientists created ProteinVAE, an AI model to design unique protein variants, aiming to evade immune responses in gene therapy. By re-engineering adenovirus hexon proteins, they generate novel sequences for improved safety and efficacy, with faster, cost-effective design. Successful experiments could enhance gene therapy. 

💎 AI's Diagnostic Prowess Advancing Rapidly: A study assessed GPT-3.5 and GPT-4's ability to mimic doctors' diagnostic reasoning. AI gave mostly accurate responses to various prompts, suggesting potential for assisting physicians, provided clinicians grasp the AI's response generation process.  

AI in Supply Chain Management: 

💎 AI Transforming Global Supply Chain Management: A recent study revealed that 98% of executives consider AI crucial for enhancing supply chain operations. AI aids in cost reduction through inventory optimization, transportation, and trade expense management. It particularly benefits inventory management, a vital cost-cutting aspect, as businesses increasingly adopt AI and data-driven solutions to overcome ongoing supply chain challenges and achieve strategic objectives like cost reduction and revenue growth. 

💎 Edge AI's Potential in Logistics Faces Memory Limitations: AI and edge computing offer potential for efficient supply chain management through real-time decision-making. However, data processing at the edge strains memory. MRAM tech may alleviate limitations. As data increases, novel storage solutions are essential for maximizing edge AI's logistics benefits, hinging on memory improvements. 

🔮 Expert Insights from Packt Community 

Complete Python Course with 10 Real-World Projects [Video] - By Ardit Sulce 

This Python course benefits both beginners and experienced AI developers by thoroughly covering Python's versatility in supporting different programming paradigms. It begins with a concise introduction and explores fundamental to advanced Python techniques. 

For beginners, the initial 12 sections provide a strong foundation in Python basics. Experienced developers can sharpen their skills by exploring intermediate and advanced concepts such as OOPS, classes, lists, modules, functions, and JSON. Additionally, the course extends into practical application by teaching the use of essential libraries like Matplotlib and NumPy, web development with Flask, and even Android APK file manipulation. 

 Database handling and geographical app development are also included, enriching the skill set of both novices and experts. The course culminates with the creation of ten practical applications, ranging from a volcano web map generator to data analysis dashboards, mobile apps, web scraping tools, and more. 

Ultimately, you will gain the ability to independently create executable Python programs, master coding syntax, and achieve comprehensive proficiency in Python programming, making this course highly recommended for students at all levels of experience. 

Plus, you can access project resources at: https://github.com/PacktPublishing/Complete-Python-Course-with-10-Real-World-Projects.

Discover the "Complete Python Course with 10 Real-World Projects [Video]" by Ardit Sulce, published in February 2023. Get a comprehensive overview of the course content through a complete chapter preview video. Alternatively, access the entire Packt digital library, including this course and more, with a 7-day free trial. To access more insights and additional resources, click the button below.

🌟 Secret Knowledge: AI/LLM Resources

💎 Assessing AI Accuracy: New Leaderboard Evaluates Language Models: The Hallucinations Leaderboard assesses Language Models (LLMs) for generating incorrect information. It includes diverse tasks to test accuracy. Initial findings reveal variations in model performance, aiding in the pursuit of more reliable and less hallucination-prone LLMs. 

💎 New Techniques to Boost AI Performance: Google researchers have created innovative software techniques for optimizing mixed-input matrix multiplication, a critical AI computation. They utilize register shuffling and efficient data type conversions to map mixed-input tasks onto hardware-accelerated Tensor Cores with minimal overhead. On NVIDIA GPUs, their approach matches the performance of hardware-native operations, offering potential solutions to computational challenges in advancing AI technology. 

💎 Neural Networks Learn Non-Linear Functions Through Rectified Activation: The article explains how neural networks approximate complex functions, especially with Rectified Linear Unit (ReLU) activation. ReLU enables multiple neurons to represent linear and curved functions, emphasizing proper architecture over more layers for accuracy without overfitting, revealing neural networks' expressive capabilities. 

💎 Enhancing Conversations with Contextual Understanding: The article explores enhancing conversational agents' contextual understanding in multi-turn conversations. It utilizes models like FLAN and Llama to assess question relationships, incorporates context from prior questions, and reevaluates responses to provide more comprehensive and successful conversation outcomes. These methods showed promising results in testing.

🔛 Masterclass: AI/LLM Tutorials

💎 Fine-Tuning BERT for Long-Form Text: This article discusses how to use powerful NLP models like BERT to analyze long passages. BERT works by splitting lengthy reviews into chunks that fit its 512-token limit. It first pre-trains on general text, then fine-tunes on a task. To classify a long movie review, it is chopped into pieces, tagged with IDs, and stacked into the model. Results are pooled to get an overall sentiment.  

💎 Training Giant Language Models the Cost-Effective Way: This guide explains how to enhance AI model training on AWS using Trainium and EKS. Set up a Kubernetes cluster with Trainium chips, preprocess data with Llama2's tokenizer, and optimize the model with compilation jobs. Monitor training across nodes, checking logs, utilization, and using Tensorboard for cost-effective training of large models. 

💎 Optimizing AI Models for Speed and Scale: This blog advises on efficiently setting up and optimizing large language models for speed and scalability. It emphasizes hardware and configuration choices tailored to each model's requirements, balancing factors like response time and concurrent user capacity. Benchmarks demonstrate how GPU usage, model distribution, and request volume affect performance. Proper tuning enables AI assistants to engage in concurrent conversations with many users without excessive delays. 

💎 Estimating Depth from a Single Image: Researchers Evaluate AI Models: Researchers evaluated neural networks' monocular depth estimation capabilities using single photos and RGB-depth map datasets. DPT and Marigold models were trained and compared, with DPT outperforming in accuracy based on RMSE and SSIM metrics. While promising, further enhancements are needed in this area. 

🚀 HackHub: Trending AI Tools

💎 rasbt/LLMs-from-scratch: Incrementally build a ChatGPT-like LLM from scratch for educational purposes.

💎 LaurentMazare/mamba.rs: Pure Rust version of the Mamba inference algorithm with minimal dependencies for efficient sequence modeling. 

💎 0nutation/speechgpt: Allows multi-modal conversations and SpeechGPT-Gen for scaling chain-of-information speech generation.

💎 naver/roma: PyTorch library offering useful tools for working with 3D rotations, including conversions between rotation representations, metrics, interpolation, and differentiation capabilities.