“[AI] will touch every sector, every industry, every business function, and significantly change the way we live and work..this isn’t just the future. We are already starting to experience the benefits right now. As a company, we’ve been preparing for this moment for some time.”
We’ve also handpicked some great GitHub repos for you to use on your next AI project!
What do you think of this issue and our newsletter? Please consider taking the short survey below to share your thoughts and you will get a free PDF of the “The Applied Artificial Intelligence Workshop” eBook upon completion.
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
⚡ TechWave: AI/GPT News & Analysis
OpenAI Introduces ChatGPT Enterprise: AI Solution for Businesses: OpenAI has unveiled ChatGPT Enterprise with advanced features. The enterprise-grade version offers enhanced security, privacy, and access to the more powerful GPT-4 model. It includes unlimited usage of GPT-4, higher-speed performance, longer context windows for processing lengthier inputs, advanced data analysis capabilities, customization options, and more, targeting improved productivity, customized workflows, and secure data management.
Meta Introduces Code Llama: A Breakthrough in AI-Powered Coding Assistance: Code Llama is a cutting-edge LLM designed to generate code based on text prompts and is tailored for code tasks and offers the potential to enhance developer productivity and facilitate coding education. Built on Llama 2, Code Llama comes in different models, including the foundational code model, Python-specialized version, and an instruct variant fine-tuned for understanding natural language instructions. The models outperformed existing LLMs on code tasks and hold promise for revolutionizing coding workflows while adhering to safety and responsible use guidelines.
Nvidia Reports Over 100% Increase in Sales Amid High Demand for AI Chips: Nvidia has achieved record-breaking sales, more than doubling its revenue to over $13.5 billion for the quarter ending in June. The company anticipates further growth in the current quarter and plans to initiate a stock buyback of $25 billion. Its stock value soared by more than 6.5% in after-hours trading, bolstering its substantial gains this year. Nvidia's data center business, which includes AI chips, fueled its strong performance, with revenue surpassing $10.3 billion, driven by cloud computing providers and consumer internet firms adopting its advanced processors. With a surge in its market value, Nvidia joined the ranks of trillion-dollar companies alongside Apple, Microsoft, Alphabet, and Amazon.
Businesses Facing AI Trust Gap with Customers, Reveals Salesforce's State of the Connected Customer Report: Salesforce's sixth edition of the State of the Connected Customer report highlights a growing concern among businesses about an AI trust gap with their customers. The survey, conducted across 25 countries with over 14,000 consumers and business buyers, indicates that as companies increasingly adopt AI to enhance efficiency and meet customer expectations, nearly three-quarters of their customers are worried about unethical AI use. Consumer receptivity to AI has also decreased over the past year, urging businesses to address this gap by implementing ethical guidelines and providing transparency into AI applications.
Microsoft Introduces "Algorithm of Thoughts" to Enhance AI Reasoning: Microsoft has unveiled a novel AI training method called the "Algorithm of Thoughts" (AoT), aimed at enhancing the reasoning abilities of large language models like ChatGPT by combining human-like cognition with algorithmic logic. This new approach leverages "in-context learning" to guide language models through efficient problem-solving paths, resulting in faster and less resource-intensive solutions. The technique outperforms previous methods and can even surpass the algorithm it is based on.
Google's Duet AI Expands Across Google Cloud with Enhanced Features: Google's Duet AI, a suite of generative AI capabilities for tasks like text summarization and data organization, is expanding its reach to various products and services within the Google Cloud ecosystem. The expansion includes assisting with code refactoring, offering guidance on infrastructure configuration and deployment in the Google Cloud Console, writing code in Google's dev environment Cloud Workstations, generating flows in Application Integration, and more. ̌It also integrates generative AI advancements into the security product line.
OpenAI Collaborates with Scale to Enhance Enterprise Model Fine-Tuning Support: OpenAI has entered into a partnership with Scale to provide expanded support for enterprises seeking to fine-tune advanced models. Recognizing the demand for high performance and customization in AI deployment, OpenAI introduced fine-tuning for GPT-3.5 Turbo and plans to extend it to GPT-4. This feature empowers companies to customize advanced models with proprietary data, enhancing their utility. OpenAI assures that customer data remains confidential and is not utilized to train other models.
Google DeepMind Introduces SynthID: A Tool to Identify AI-Generated Images: In response to the growing prevalence of AI-generated images that can be indistinguishable from real ones, Google Cloud has partnered with Imagen to unveil SynthID. This newly launched beta version aims to watermark and identify AI-created images. The technology seamlessly embeds a digital watermark into the pixels of an image, allowing for imperceptible yet detectable identification. This tool is a step towards responsible use of generative AI and enhances the capacity to identify manipulated or fabricated images.
✨ Unleashing the Power of Causal Reasoning with LLMs:
Join Aleksander Molak on October 11th and be a part of Packt's most awaited event of 2023 on Generative AI!
In AI's evolution, a big change is coming. It's all about Causally Aware Prompt Engineering, and you should pay attention because it's important. LLMs are good at recognizing patterns, but what if they could do more? That's where causal reasoning comes in. It's about understanding not just what's connected but why.
Let's distill the essence:
- LLMs can outperform causal discovery algorithms on some tasks
- GPT-4 achieves a near-human performance on some counterfactual benchmarks
- This might be the case because the models simply memorize the data, but it's also possible that they build a **meta-SCM** (meta structural causal models) based on the correlations of causal facts learned from the data
- LLMs can reason causally if we allow them to intervene on the test time
- LLMs do not reason very well, when we provide them with verbal description of conditional independence structures in the data (but nor do (most of) humans).
Now, catalyze your journey with three simple techniques:
Causal Effect Estimation: Causal effect estimate aims at capturing the strength of (expected) change in the outcome variable when we modify the value of the treatment by one unit. In practice, almost any machine learning algorithm can be used for this purpose, yet in most cases we need to use these algorithms in a way that differs from the classical machine learning flow.
Confronting Confounding: The main challenge (yet not the only one) in estimating causal effects from observational data comes from confounding. Confounder is a variable in the system of interest that produces a spurious relationship between the treatment and the outcome. Spurious relationships are a kind of illusion. Interestingly, you can observe spurious relationships not only in the recorded data, but also in the real world.
Unveiling De-confounding: To obtain an unbiased estimate of the causal effect, we need to get rid of confounding. At the same time, we need to be careful not to introduce confounding ourselves! This usually boils down to controlling for the right subset of variables in your analysis. Not too small, not too large.
If you're intrigued by this, I invite you to join me for an in-depth exploration of this fascinating topic at Packt's upcoming Generative AI conference on October 11th. During my power-talk, we'll delve into the question: Can LLMs learn Causally?
Regularization serves as a valuable approach to enhance the success rate of ML models in production. Effective regularization techniques can prevent AI recruitment models from exhibiting gender biases, either by eliminating certain features or incorporating synthetic data. Additionally, proper regularization enables chatbots to maintain an appropriate level of sensitivity toward new tweets. It also equips models to handle edge cases and previously unseen data proficiently, even when trained on synthetic data.
Key concepts of regularization
Let us now delve into a more precise definition and explore key concepts that enable us to better comprehend regularization.
Bias and variance
Bias and variance are two key concepts when talking about regularization. We can define two main kinds of errors a model can have:
Bias is how bad a model is at capturing the general behavior of the data
Variance is how bad a model is at being robust to small input data fluctuations
Let’s describe those four cases:
High bias and low variance: The model is hitting away from the center of the target, but in a very consistent manner
Low bias and high variance: The model is, on average, hitting the center of the target, but is quite noisy and inconsistent in doing so
High bias and high variance: The model is hitting away from the center in a noisy way
Low bias and low variance: The best of both worlds – the model is hitting the center of the target consistently
By Vincent Vandenbussche and published in July 2023. To get a glimpse of the book's contents, make sure to read the free chapter provided here, or if you want to unlock the full Packt digital library free for 7 days, try signing up now! To learn more, click on the button below.
Google’s RO-ViT: Region-Aware Pre-Training for Open-Vocabulary Object Detection with Vision Transformers: Google's research scientists have unveiled a new method called "RO-ViT" that enhances open-vocabulary object detection using vision transformers. Learn how the technique addresses limitations in existing pre-training approaches for vision transformers, which struggle to fully leverage the concept of objects or regions during pre-training. RO-ViT introduces a novel approach called "cropped positional embedding" that aligns better with region-level tasks.
Tiered AIOps: Enhancing Cloud Platform Management with AI: Explore the concept of Tiered AIOps to manage complex cloud platforms. The ever-changing nature of cloud applications and infrastructure presents challenges for complete automation, requiring a tiered approach to combine AI and human intervention. The concept involves dividing operations into tiers, each with varying levels of automation and human expertise. Tier 1 incorporates routine operations automated by AI, Tier 2 empowers non-expert operators with AI assistance, and Tier 3 engages expert engineers for complex incidents.
Effective AI-Agent Interaction: SERVICE Principles Unveiled: In this post, you'll learn how to design AI agents that can interact seamlessly and effectively with users, aiming to transition from self-service to "agent-service." The author introduces the concept of autonomous AI agents capable of performing tasks on users' behalf and offers insights into their potential applications. The SERVICE principles, rooted in customer service and hospitality practices, are presented as guidelines for designing agent-user interactions. These principles encompass key aspects like salient responses, explanatory context, reviewable inputs, vaulted information, indicative guidance, customization, and empathy.
How to Mitigate Hallucination in Large Language Models: In this article, researchers delve into the persistent challenge of hallucination in Generative LLMs. The piece explores the reasons behind LLMs generating nonsensical or non-factual responses, and the potential consequences for system reliability. The focus is on practical approaches to mitigate hallucination, including adjusting the temperature parameter, employing thoughtful prompt engineering, and incorporating external knowledge sources. The authors conduct experiments to evaluate different methods, such as Chain of Thoughts, Self-Consistency, and Tagged Context Prompts.
💡 MasterClass: AI/LLM Tutorials
How to Use Code Llama: A Breakdown of Features and Usage: Code Llama has made a significant stride in code-related tasks, offering an open-access suite of models specialized for code-related challenges. This release includes various notable components, such as integration within the Hugging Face ecosystem, transformative integration, text generation inference, and inference endpoints. Learn how these models showcase remarkable performance across programming languages, enabling enhanced code understanding, completion, and infilling.
Make Data Queries with Hugging Face's VulcanSQL: In this post, you'll learn how to utilize VulcanSQL, an open-source data API framework, to streamline data queries. VulcanSQL integrates Hugging Face's powerful inference capabilities, allowing data professionals to swiftly generate and share data APIs without extensive backend knowledge. By incorporating Hugging Face's Inference API, VulcanSQL enhances the efficiency of query processes. The framework's HuggingFace Table Question Answering Filter offers a unique solution by leveraging pre-trained AI models for NLP tasks.
Exploring Metaflow and Ray Integration for Supercharged ML Workflows: Explore the integration of Metaflow, an extensible ML orchestration framework, with Ray, a distributed computing framework. This collaboration leverages AWS Batch and Ray for distributed computing, enhancing Metaflow’s capabilities. Know how this integration empowers Metaflow users to harness Ray’s features within their workflows. The article also delves into the challenges faced, the technical aspects of the integration, and real-world test cases, offering valuable insights into building efficient ML workflows using these frameworks.
Explore Reinforcement Learning Through Solving Leetcode Problems: Explore how reinforcement learning principles can be practically grasped by solving a Leetcode problem. The article centers around the "Shortest Path in a Grid with Obstacles Elimination" problem, where an agent aims to find the shortest path from a starting point to a target in a grid with obstacles, considering the option to eliminate a limited number of obstacles. Explore the foundations of reinforcement learning, breaking down terms like agent, environment, state, and reward system. The author provides code examples and outlines how a Q-function is updated through iterations.
🚀 HackHub: Trending AI Tools
apple/ml-fastvit: Introduces a rapid hybrid ViT empowered by structural reparameterization for efficient vision tasks.
openchatai/opencopilot: A personal AI copilot repository that seamlessly integrates with APIs and autonomously executes API calls using LLMs, streamlining developer tasks and enhancing efficiency.
neuml/txtai: An embeddings database for advanced semantic search, LLM orchestration, and language model workflows featuring vector search, multimodal indexing, and flexible pipelines for text, audio, images, and more.
Databingo/aih: Interact with AI models via terminal (Bard, ChatGPT, Claude2, and Llama2) to explore diverse AI capabilities directly from your command line.
osvai/kernelwarehouse: Optimizes dynamic convolution by redefining kernel concepts, improving parameter dependencies, and increasing convolutional efficiency.
morph-labs/rift: Open-source AI-native infrastructure for IDEs, enabling collaborative AI software engineering.
mr-gpt/deepeval: Python-based solution for offline evaluations of LLM pipelines, simplifying the transition to production.
United States
Great Britain
India
Germany
France
Canada
Russia
Spain
Brazil
Australia
Singapore
Hungary
Philippines
Mexico
Thailand
Ukraine
Luxembourg
Estonia
Lithuania
Norway
Chile
South Korea
Ecuador
Colombia
Taiwan
Switzerland
Indonesia
Cyprus
Denmark
Finland
Poland
Malta
Czechia
New Zealand
Austria
Turkey
Sweden
Italy
Egypt
Belgium
Portugal
Slovenia
Ireland
Romania
Greece
Argentina
Malaysia
South Africa
Netherlands
Bulgaria
Latvia
Japan
Slovakia
