AI_Distilled #24: Google Invests $2 Billion in Anthropic, Perplexity's AI Search Engine, Biden's AI Executive Order, Data Mining with GPT-4, RL and AWS Deepracer

👋 Hello ,

Welcome to another captivating edition of AI_Distilled, featuring recent advancements in training and fine-tuning LLMs, GPT and AI models for enhanced business outcomes.

Let’s begin our news and analysis with an industry expert’s opinion.  

“Artificial intelligence is the science of making machines do things that would require intelligence if done by humans” – John McCarthy, Computer Scientist and AI Visionary. 

AI does indeed make machines intelligent, so much so that industry titans are now waging a proxy AI war with billions in startup funding. Without a doubt, AI is onto something big! 

In this week, we’ll talk about Biden's AI Executive Order, which has been praised for scope but deemed insufficient without legislation, Perplexity's AI Search Engine, OpenAI launching new team and challenge to prepare for catastrophic risks of advanced AI, Google Invests $2 Billion in Anthropic, and updating its Bug Bounty program to address AI security concerns. 

Look out for your fresh dose of AI resources, secret knowledge, and tutorials on how to use custom AI models to enhance complex technical workflows, improving LLM understanding with user feedback, and essential text preprocessing for effective machine learning with Python. 

📥 Feedback on the Weekly Edition

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 

SignUp | Advertise | Archives

⚡ TechWave: AI/GPT News & Analysis

🔹 OpenAI Launches New Team and Challenge to Prepare for Catastrophic Risks of Advanced AI: The ChatGPT creator announced new efforts to prepare for potential catastrophic risks associated with highly advanced AI systems. The company is forming a new internal team called "Preparedness" to assess risks ranging from cybersecurity threats to autonomous biological replication. It is also launching an "AI Preparedness Challenge" with prize money to crowdsource ideas for preventing misuse of advanced AI. OpenAI says it aims to benefit humanity with cutting-edge AI while taking seriously the full spectrum of safety risks.

🔹 Biden's AI Executive Order Praised for Scope but Deemed Insufficient Without Legislation: President Biden recently issued an executive order on AI that experts say covers important ground but lacks teeth without accompanying legislation from Congress. The order establishes guidelines and oversight for AI development and use, including in healthcare. However, many provisions simply codify voluntary industry practices. Stakeholders say Congress must pass more comprehensive AI regulations, but partisan disputes make near-term action unlikely.  

🔹 Google Updates Bug Bounty Program to Address AI Security Concerns: Google has expanded its vulnerability rewards program to include incentives for discovering potential abuses of artificial intelligence systems. The update comes as worries grow over generative AI being exploited maliciously. Under the revised guidelines, security researchers can earn financial rewards for uncovering AI training data extraction that leaks private information. The move aligns with AI companies' recent White House pledge to better identify AI vulnerabilities.  

🔹 Perplexity's AI Search Engine Garners $500M Valuation After New Funding: The AI startup Perplexity recently secured additional funding led by venture capital firm IVP, garnering a $500 million valuation. Perplexity is developing a conversational search engine to challenge Google's dominance using artificial intelligence. The company's iOS app and website traffic have been growing steadily amid rising interest in AI like ChatGPT. With deep ties to Google researchers, Perplexity leverages LLMs and has attracted investments from major industry figures.  

🔹 Tech Giants Wage Proxy AI War with Billions in Startup Funding As Google Invests $2 Billion in Anthropic: Major technology companies like Google, Microsoft, and Amazon are investing billions in AI startups like OpenAI and Anthropic as surrogates in the race to lead the AI space. Unable to quickly build their own capabilities in large language models, the tech giants are funneling massive sums into the AI leaders to gain ownership stakes and technology access. Anthropic's $2 billion funding from Google follows similar multibillion investments from Microsoft and Amazon, fueling an expensive AI innovation war by proxy.  

🔹 Poe Unveils Monetization for Third-Party Conversational AI Developers: The AI chatbot platform Poe has introduced a new revenue sharing model to let creators’ profit from building specialized bots. Poe will split subscription fees and pay per-message charges to offset infrastructure costs. An open API also allows adding custom natural language models beyond Poe's defaults. The moves aim to spur innovation by empowering niche developers. Poe believes reducing barriers will increase diversity, not just competition.  

🔮 Expert Insights from Packt Community 

Generative AI with Python and TensorFlow 2 - By Joseph Babcock , Raghav Bali 

Kubeflow: an end-to-end machine learning lab 

As was described at the beginning of this chapter, there are many components of an end-to-end lab for machine learning research and development (Table 2.1), such as: 

A way to manage and version library dependencies, such as TensorFlow, and package them for a reproducible computing environment 

Interactive research environments where we can visualize data and experiment with different settings 

A systematic way to specify the steps of a pipeline – data processing, model tuning, evaluation, and deployment 

Provisioning of resources to run the modeling process in a distributed manner 

Robust mechanisms for snapshotting historical versions of the research process 

As we described earlier in this chapter, TensorFlow was designed to utilize distributed resources for training. To leverage this capability, we will use the Kubeflow projects. Built on top of Kubernetes, Kubeflow has several components that are useful in the end-to-end process of managing machine learning applications. 

Using Kubeflow Katib to optimize model hyperparameters 

Katib is a framework for running multiple instances of the same job with differing inputs, such as in neural architecture search (for determining the right number and size of layers in a neural network) and hyperparameter search (finding the right learning rate, for example, for an algorithm). Like the other Customize templates we have seen, the TensorFlow job specifies a generic TensorFlow job, with placeholders for the parameters: 

apiVersion: "kubeflow.org/v1alpha3" 
kind: Experiment 
metadata: 
 namespace: kubeflow 
 name: tfjob-example 
spec: 
parallelTrialCount: 3 
 maxTrialCount: 12 
 maxFailedTrialCount: 3 
 objective: 
   type: maximize 
   goal: 0.99 
   objectiveMetricName: accuracy_1 
 algorithm: 
   algorithmName: random 
 metricsCollectorSpec: 
   source: 
     fileSystemPath: 
       path: /train 
       kind: Directory 
   collector: 
     kind: TensorFlowEvent 
 parameters: 
   - name: --learning_rate 
     parameterType: double 
     feasibleSpace: 
       min: "0.01" 
       max: "0.05" 
   - name: --batch_size 
     parameterType: int 
     feasibleSpace: 
       min: "100" 
       max: "200" 
 trialTemplate: 
   goTemplate: 
       rawTemplate: |- 
         apiVersion: "kubeflow.org/v1" 
         kind: TFJob 
         metadata: 
           name: {{.Trial}} 
           namespace: {{.NameSpace}} 
         spec: 
          tfReplicaSpecs: 
           Worker: 
             replicas: 1  
             restartPolicy: OnFailure 
             template: 
               spec: 
                 containers: 
                   - name: tensorflow  
                     image: gcr.io/kubeflow-ci/tf-mnist-with- 
                            summaries:1.0 
                     imagePullPolicy: Always 
                     command: 
                       - "python" 
                       - "/var/tf_mnist/mnist_with_summaries.py" 
                       - "--log_dir=/train/metrics" 
                       {{- with .HyperParameters}} 
                       {{- range .}} 
                       - "{{.Name}}={{.Value}}" 
                       {{- end}} 
                       {{- end}} 
 

which we can run using the familiar kubectl syntax: 

kubectl apply -f

https://raw.githubusercontent.com/kubeflow/katib/master/examples/v1alpha3/tfjob-example.yaml 

This content is from the book “Generative AI with Python and TensorFlow 2” by Joseph Babcock , Raghav Bali (April 2021). Start reading a free chapter or access the entire Packt digital library free for 7 days by signing up now. To learn more, click on the button below. 

🌟 Secret Knowledge: AI/LLM Resources

🔹 How to Use Custom AI Models to Enhance Complex Technical Workflows: In this post, you'll learn how Nvidia’s researchers leveraged customized LLMs to streamline intricate semiconductor chip design. The research demonstrates how to refine foundation models into customized assistants that understand industry-specific patterns. You'll see how careful data cleaning and selection enables high performance even with fewer parameters. The post explores step-by-step instructions on how researchers built a specialized AI that helps with writing code, improving documentation, and optimizing complex technical workflows.  

🔹 How to Build Impactful LLM Applications: In this post, you'll explore lessons learned from creating Microsoft's Copilot products, such as Viva and PowerPoint. It discusses how combining LLMs with app context and other ML models can be a game-changer and demonstrates how parsing user queries and responses enables precise skill activation. By following their approach of utilizing multiple models to summarize insights without losing nuance, you can gain practical tips for your own LLM application development. 

🔹 Understanding Convolutional Neural Networks and Vision Transformers: A Mathematical Perspective: You'll learn about convolutional neural networks and vision transformers in this post. They're great for image classification but differ in math, especially for generative tasks. You'll see how their training budgets work and understand their unique math. We'll also discuss their differences in complexity and memory usage. Plus, you'll learn why convolutional nets handle spatial coherence naturally, while vision transformers might need some help. By the end, you'll know why transformers are better for generating sequential data.  

🔹 Improving Large Language Model Understanding with User Feedback: The post focuses on improving user intent detection for LLMs by utilizing disambiguation, context, and MemPrompt. These techniques enhance LLM responses, enabling better understanding of user intent, offering real-time feedback, and enhancing LLM performance and utility. 

🔹 The Power of High-Quality Data in Language Models: The article emphasizes the significance of high-quality data for Large Language Models (LLMs). It introduces the concept of alignment, discussing how it influences LLM behavior. The article stresses the vital role of data quality and diversity in optimizing LLM performance and capabilities. 

💡 Masterclass: AI/LLM Tutorials

🔹 Enhance Language Model Performance with Step-Back Prompting: This guide explores the use of Step-Back Prompting to enhance LLMs' performance in complex tasks, like knowledge-intensive QA and multi-hop reasoning. It offers a step-by-step tutorial, including package setup and data collection, to implement this approach, potentially improving AI model behavior and responses. 
 

🔹 Boosting AI at Scale with Vectorized Databases: This guide explores how vectorized databases are transforming LLMs like GPT-3 by enhancing their capabilities and scalability. It explains the principles of LLMs and the role of vectorized databases in empowering them. It discusses efficient data retrieval, optimization of vector operations, and scaling for real-time responses. The guide highlights use cases, including content generation and recommendation systems, where vectorized databases excel, and addresses the challenges of adopting them for LLMs. 

🔹 Mastering Data Mining with GPT-4: A Practical Guide Using Seattle Weather Data: This guide explores the use of GPT-4 for data mining using Seattle's weather dataset. It covers AI's potential in data mining, detailing the process from exploratory data analysis to clustering and anomaly detection. GPT-4 assists in data loading, EDA, data cleaning, feature engineering, and suggests clustering methods. The post highlights the collaborative aspect of AI-human interaction and how GPT-4 can improve data mining and data analysis in the field of data science. 

🔹 Introduction to Reinforcement Learning and AWS Deepracer: This post introduces reinforcement learning, a machine learning approach focused on maximizing rewards through agent-environment interactions. It compares it to motivating students based on performance. It explores practical applications via AWS Deepracer for self-driving cars, explaining key components and mentioning the Deepracer Student League as a learning opportunity. 
 

🔹 Essential Text Preprocessing for Effective Machine Learning with Python: This post highlights crucial text preprocessing techniques for machine learning. It emphasizes the need to clean text data to avoid interference and unintended word distinctions. The methods, including removing numbers and handling extra spaces, enhance text data quality for effective machine learning applications. 

🚀 HackHub: Trending AI Tools

🔹 Pythagora-io/gpt-pilot: Boosts app development speed 20x via requirement specification, oversight, and coding assistance through clarifications and reviews. 

🔹 hkuds/rlmrec: PyTorch implementation for the RLMRec model, enhancing recommenders with LLMs for advanced representation learning in recommendation systems. 

🔹 THUDM/AgentTuning: Empowers LLMs by instruction-tuning them with interaction trajectories from various agent tasks, enhancing their generalization and language abilities. 

🔹 cpacker/MemGPT: Enhances LLMs by intelligently managing memory tiers, enabling extended context and perpetual conversations.