EP15/16: Double Feature Newsletter, Designing with AI & GPT4 IS getting Dumber!

This week it's HTTTA^2...that's right, we're exponential!

Tune in for the live chat with Leonardo.AI co-founder Ethan Smith and Binx, AI Artist:

Howdy, prompt engineers and AI enthusiasts!

In this week’s special 2x issue…

Get ready to boost your AI knowledge in our latest HTTTA podcast episodes! 🌟 First, we delve into the riveting world of GPT-4, exploring its architecture, training process, and capabilities based on speculated leaked reports. We tackle prompt engineering, VR and AR's future, and AI integration optimization for businesses, discussing aspects ranging from power usage costs to the use of smaller inference models. In another episode, Wes and Goda shed light on their CoRise course "AI and ChatGPT for Everyone", demonstrating how AI and design tools are revolutionizing various industries. We examine the continuous evolution of AI models, including Meta's game-changing open-sourcing move, and dissect a research paper on OpenAI's code generation and GPT4's accuracy. Goda also introduces her exciting collaboration with Politecnico di Milano, teaching design with emerging technologies. 🎙️ So, buckle up for these enlightening AI journeys filled with captivating facts, thought-provoking ideas, and the endless potential of emerging technologies. Happy Prompting Everybody!Designing with Emerging Technologies Course: Use code "GodaGo" at the checkout for 15% OFF

Discord (Goda Go#3156 & Commordore_Wesmardo#2912)
AI & Prompt Engineering for Everyone on CoRise.com: sign up here and use code 'AI4ALL' for 10% off
FOR 15% OFF PROMPT PERFECT Click here & use code 'httta' at the checkout!

1. Designing with Emerging Technologies: Goda Leads a Dynamic Course hosted by one of the World’s Preeminent Design Schools:Course Creator Live Stream Q&A Thursday 27 July 2023

In the rapidly evolving world of emerging technologies, POLI.design offers a compelling executive course called 'Designing with Emerging Technologies.' authored and chaired by our very own Goda Go! These technologies, including Artificial Intelligence (AI), Extended Reality (XR), the Internet of Things (IoT), and Web3/Distributed Autonomous Organizations (DAOs), are drastically transforming design methods, artifacts, co-creation strategies, business models, and even cultures at an unprecedented speed. The key challenge industries face is to effectively and timely incorporate these technologies into their design process.

While rapid engineering is indeed a crucial factor in this integration process, there's a broader scope for designers to seize. With these new technologies, designers have the opportunity to elevate their abilities and become 'superdesigners.' They can harness the power of these innovations to not only enhance their design skills but also to revolutionize the field.

POLI.design's 'Designing with Emerging Technologies' course series aims to prepare and empower designers from all disciplines to navigate the dynamic landscape of these emerging technologies. With these courses, learners are expected to gain increased awareness and confidence, and a readiness to use these technologies. More importantly, they are encouraged to augment design to build a more inclusive and sustainable world.

The first course, 'Level-Up: Learning to Design with Emerging Technologies,' encourages learners to explore the transformative tech megatrends that are revolutionizing the design landscape. It provides an opportunity to delve into the intricacies of these new technologies and learn how to incorporate them into design practice.

For a more select audience, the invite-only course, 'Game-On: Designing with Emerging Technologies,' takes the training a step further. It prompts learners to unleash the potential of emerging technologies, pushing the boundaries of innovation and creativity. This course is designed to truly unlock the potential of these technologies, preparing designers to meet and exceed the demands of the future.

2. On CoRise: AI & ChatGPT for Everyone from Synthminds.AI (Cue epic promo video)

Wes will be teaching AI & ChatGPT for Everyone, a cutting-edge course that dives deep into the world of language models and chatbots, with a specific focus on leveraging the power of LLMs like ChatGPT. Designed for individuals across various domains, this course equips learners with the knowledge and skills to harness the capabilities of AI-driven conversation engines. Through hands-on exercises and real-world examples, participants will learn how to optimize prompts for different purposes, whether it's writing persuasive copy, building interactive chatbots, or enhancing customer support systems.

The course taught by Wes with Goda participating as well, convenes August 21, 2023. Click here to sign up for the course, and use the code “AI4ALL” for 10% off!

3. GPT-4 Speculated Parameters released, and it IS getting Dumber (We knew it!)

A leaked and now removed twitter post speculatively revealed that GPT-4, the latest iteration of its generative pre-training transformer, boasts a whopping 1.8 trillion parameters, spread across 120 layers. This model, which is more than ten times the size of its predecessor, GPT-3, has been kept economically feasible by leveraging a Mixture of Experts (MoE) model. OpenAI's GPT-4 employs 16 experts, each carrying roughly 111 billion parameters. However, this method does have trade-offs. While not all sections of the model are utilized in each token generation, leading to lowered utilization rates, research has shown that using a greater number of experts can lead to better loss. OpenAI has chosen to use a lower number of experts to ensure stability in a wide array of tasks and to improve model convergence.

The GPT-4 model was trained on around 13 trillion tokens, counting multiple epochs as additional tokens. This includes two epochs for text-based data and four for code-based data. It utilizes a fine-tuning dataset consisting of millions of instruction rows from both ScaleAI and internal sources. The model was initially trained with an 8k context length, later being fine-tuned to a 32k sequence length.

When it comes to the training cost, OpenAI's GPT-4 required an impressive ~2.15e25 floating-point operations, operating on ~25,000 A100s for 90 to 100 days. Despite a significant number of failures necessitating restarts from checkpoints, the training cost for this run alone amounted to approximately $63 million. This enormous training run led to a conservative approach with regards to the number of experts used in the MoE model.

GPT-4's inference cost is about three times that of the 175 billion parameter Davinci due to the larger clusters required and the lower utilization achieved. Continuous batching and variable batch sizes were implemented to optimize inference costs and maintain maximum latency.

OpenAI's GPT-4 has also ventured into the realm of vision with a separate vision encoder similar to Flamingo's architecture. It was fine-tuned with an additional ~2 trillion tokens after the text-only pre-training. OpenAI had initially aimed to train it from scratch, but chose to start with text to reduce risk. This vision capability primarily targets autonomous agents able to read web pages and transcribe images and videos.

Another potentially exciting development in GPT-4 is the use of speculative decoding. The idea behind this is to have a smaller, faster model decode several tokens in advance and then feed them into a larger oracle model as a single batch. This approach could be contributing to perceived reductions in GPT-4's quality if the oracle model is accepting lower probability sequences from the speculative decoding model.

Despite these remarkable advancements, OpenAI has faced challenges in acquiring high-quality data. Training on a staggering 13 trillion tokens, the organization has used a range of sources like CommonCrawl, RefinedWeb, and rumored sources such as Twitter, Reddit, and YouTube. There are speculations that parts of the data have been collected from extensive databases like LibGen and Sci-Hub, and even GitHub. Some speculate that hand-collected datasets from college textbooks were also used, thereby allowing the model to appear "smart" across a broad range of topics. These are, however, speculations and OpenAI has not confirmed these rumors.

OpenAI's GPT-4 is an ambitious stride in the field of artificial intelligence, bringing both considerable advancements and challenges. As it continues to evolve, the field eagerly awaits to see what future iterations will bring.

The introduction section of this research paper discusses the challenges and uncertainties surrounding large language models (LLMs) such as GPT-3.5 and GPT-4. It is currently unclear when and how these models are updated, making it difficult to integrate them into workflows and reproduce consistent results. The researchers aim to address these concerns by evaluating the behavior of the March 2023 and June 2023 versions of GPT-3.5 and GPT-4 on various tasks including math problem-solving, answering sensitive questions, generating code, and visual reasoning. They find that the performance and behavior of both models vary significantly across releases, with some tasks showing a decline in performance over time. The paper also mentions previous work that evaluated LLMs like GPT-3.5 and GPT-4 on traditional language tasks, but these studies did not systematically monitor the longitudinal shifts in performance over time. The research paper provides a figure illustrating the performance differences between the two versions of GPT-3.5 and GPT-4 on the four tasks mentioned. Overall, the paper aims to address the need for understanding the updates and impact of LLMs like GPT-3.5 and GPT-4 to ensure their reliable and effective integration into various applications. Overview: LLM Services, Tasks and Metrics This research paper focuses on studying the behavioral changes of different language model models (LLMs) over time. To quantitatively answer the research question, the paper specifies which LLM services to monitor, which application scenarios to focus on, and how to measure LLM drifts in each scenario. The LLM services monitored in this paper are GPT-4 and GPT-3.5, which are the backbone of ChatGPT. These two services have been widely adopted by individual users and businesses, making it important to monitor their behavior. The paper specifically looks at the drifts between the March 2023 and June 2023 versions of GPT-4 and GPT-3.5. The evaluation tasks in this study are focused on four frequently studied LLM tasks: solving math problems, answering sensitive questions, code generation, and visual reasoning. These tasks are diverse and commonly used to evaluate LLMs in performance and safety benchmarks. For each task, the authors use one dataset, either sampled from existing datasets or created by themselves for monitoring purposes. It is acknowledged that using one benchmark dataset may not provide a comprehensive assessment of the task, but the goal here is to demonstrate substantial performance drift in ChatGPT on simple tasks. The authors plan to include more benchmarks in future evaluations as part of a broader, long-term study of LLM service behavior. In summary, this section of the research paper outlines the LLM services, evaluation tasks, and datasets chosen for monitoring and studying the drifts in behavior of GPT-4 and GPT-3.5 in different application scenarios. Solving Math Problems: Chain-of-Thought Might Fail 

The section "Solving Math Problems: Chain-of-Thought Might Fail" from a research paper discusses the evolution of GPT-4 and GPT-3.5's math-solving skills over time. The study focuses on their ability to determine if a given integer is prime. The dataset consists of 500 questions, and Chain-of-Thought, a reasoning approach, is used to help the models. Surprisingly, substantial changes in accuracy and response length are observed. GPT-4's accuracy drops significantly from 97.6% in March to 2.4% in June, while GPT-3.5's accuracy improves from 7.4% to 86.8%. The generation length of GPT-4 also decreases, while GPT-3.5's response length increases. The chain-of-thought approach has different effects on the models, with GPT-4 generating a more compact response in June and GPT-3.5 generating the wrong initial answer in March. The update in June improves GPT-3.5's performance. The paper also explores the models' response to sensitive questions and jailbreaking attacks. GPT-4 shows a stronger defense against attacks, with a decrease in answer rate, while GPT-3.5 does not show a significant difference.

The researchers also assess the models' code generation ability and find a drop in directly executable generations. Finally, the research investigates the models' performance in visual reasoning tasks, noting marginal improvements but also instances where previously correct answers become incorrect. The findings highlight the need for fine-grained monitoring of drift in language models for critical applications. Conclusions and Future Work 

The study found that the behavior of GPT-3.5 and GPT-4, which are language models (LLMs), has varied significantly in a short period of time. This highlights the importance of continuous evaluation and assessment of LLM behavior in real-world applications. To address this, the researchers plan to conduct an ongoing long-term study by regularly evaluating GPT-3.5, GPT-4, and other LLMs on diverse tasks. They recommend that users or companies relying on LLM services implement similar monitoring analysis to ensure the effectiveness and reliability of their applications. To further encourage research on LLM drifts, the researchers have made their evaluation data and ChatGPT responses publicly available on GitHub. This dataset can be used to analyze and compare the behavior of different LLMs over time. By sharing their data, the researchers hope to promote collaboration and stimulate further research in this area. In conclusion, the study emphasizes the need for continuous evaluation and assessment of LLM behavior. The researchers' ongoing long-term study will provide valuable insights into the variation of different LLMs over time. By sharing their evaluation data, they aim to encourage further research and analysis of LLM drifts.

5. Llama 2 has Entered the Chat

Meta has launched Llama 2, an open-source language model that promises to revolutionize the AI landscape. This transformative technology, boasting a training volume of 70 billion tokens, outpaces its predecessor, Chinchilla, by a 29:1 ratio. With a commitment to increasing public engagement in AI development, Meta has licensed Llama 2 for commercial use, aiming to foster innovation in the field.

This strategy democratizes access to AI technology, giving businesses, startups, and researchers the tools to scale and innovate in novel ways. It provides an unprecedented opportunity to leverage robust computing power that may otherwise be out of reach for many entities. Meta's blog post announcing the release optimistically highlighted the potential for economic and social benefits from such widespread access.

Unlike Google and OpenAI, Meta has been forthright about the development of their language learning models (LLMs). They've shared specifics such as the number of parameters and the data used for training their models. Llama 2 is accessible via platforms like Hugging Face, Amazon Web Services, and Microsoft Azure. This transparency empowers developers to improve upon the model's code and data. Llama 2 comes in three pre-trained model sizes: 7 billion, 13 billion, and 70 billion parameters.

While Llama 2 might not be the largest model, it holds the mantle as the leading open-source LLM capable of running on standard consumer-grade hardware. This leadership position is anticipated to hold for approximately the next three months.

However, Meta has included some unique provisions in their terms and conditions. The company has essentially excluded certain large-scale commercial entities (dubbed the 'Apple ban') from freely using Llama 2. Companies with a product or service boasting more than 700 million monthly active users must obtain a specific license from Meta. This model is also explicitly prohibited for military use.

In summary, Meta's release of Llama 2 marks a significant stride in democratizing AI. Offering an open model that can run on typical hardware, and Meta encourages innovation while protecting its interests through strategic licensing provisions.

6. Attention Aviators, Synthminds.AI debuts their full SaaS offering with FAR/BOT

YOUR AI AVIATION REGULATION ASSISTANT

Introducing AI for the FAR/AIM, an innovative solution committed to simplifying the labyrinthine realm of Federal Aviation Regulations (FAR) and Aeronautical Information Manual (AIM) compliance. This cutting-edge AI platform, fully developed by the team at Synthminds.AI, serves as a dependable ally for pilots and aviation professionals, empowering them with accurate resources to efficiently decipher and navigate the often convoluted regulatory landscape. With a focus on transforming the aviation industry, we're set on a mission to streamline regulatory compliance and furnish precision-based information, thereby revolutionizing the way aviation norms are understood and adhered to. FAR/BOT will be taking off soon, sign up today for discounted early access.

7. Prompts, served Hot and Fresh weekly

Did you know that GPT-4 can function as a compression/decompression algorithm for text? Use GPT-4 to compress all of Mid Journeys documentation into a few sentences for example. This makes it easier and cheaper to get ChatGPT to generate good prompts for by minifying all the important background information, data, and examples. Compress an entire webpage into 500 tokens, then decompress and recompile the output into a near-perfect summary of the original input text. A game changer!

Encode the information above into a condensed version that another GPT-4 language model can understand. Try use as few characters as possible. Remove redundancy and streamline the content to further optimize on saving space. The goal is to use this new encoded message as context for another GPT-4. Persist as much detail / examples to remain useful. You may use emojis, encoding, abbreviations, etc to achieve the goal.

8. ELI5 AI/ML Term(s) of the week: “Parameters & Hyper-parameters”

Let's imagine you're baking a cake. The recipe you have is a bit like a machine-learning model. It tells you what you need to do to make your cake.

Parameters in AI are like the ingredients in your cake recipe. When you follow a recipe, you'll need things like flour, sugar, eggs, and butter. In a machine learning model, these ingredients are the data that the model learns from. Just like how you adjust the amount of sugar to make your cake sweeter or less sweet, the model adjusts its parameters to make its predictions better.

Hyperparameters are like the instructions in your recipe. They tell you how long to bake the cake, what temperature to set the oven to, and how many times to stir the batter. In a machine learning model, hyperparameters are settings that tell the model how to learn. For example, a hyperparameter might tell the model how many times to look at the data (just like stirring the batter), or how fast to learn from the data (like the temperature of the oven).

So, in summary, the parameters are the "ingredients" that the AI adjusts to learn better, while the hyperparameters are the "recipe instructions" that we set up at the beginning to tell the AI how it should learn.

In Conclusion - What we’re Noodling with:

We’re going to close out with our top new AI tools or learning resources we are trying and loving over the past week. 1000+ new ones get released each week now (no exaggeration there) so here’s a little amuse-bouche to top off the newsletter this week. This week’s theme remains tools to punch up and enhance your AI-generated content even more. These tools layer in generative AI with some traditional ML techniques and algorithms to turbocharge your generated content. Enjoy and Happy Prompting Everybody!

Prompt Perfect (This week’s Sponsor)

Huge update to PromptPerfect, a cutting-edge prompt optimizer for large language models (LLMs), large models (LMs), and LMOps. It streamlines prompt engineering, automatically optimizing prompts for ChatGPT, GPT-3.5, DALL-E 2, StableDiffusion, and MidJourney. Whether you're a prompt engineer, content creator, or AI developer, PromptPerfect makes prompt optimization easy and accessible. Unlock the full potential of LLMs and LMs with PromptPerfect, delivering top-quality results every time. Say goodbye to subpar AI-generated content and hello to prompt perfection!

FOR 15% OFF, Use this link https://bit.ly/3Chmc16 and the code 'httta' at the checkout!

WordAi uses advanced machine learning models to provide high-quality rewriting that is indistinguishable from human content. WordAi understands the meaning of each sentence, then rewrites it from scratch while optimizing for uniqueness and readability. This allows WordAi to completely restructure your sentences while keeping the same overall meaning.

Black logo

Use the AI Writer to create long-form blogs and articles; the Sonic Editor (an enhanced version of Google Docs) to write sales emails, essays, reports, and even ebooks. Moreover, you can polish your content to perfection using our AI-driven editing tools: Paraphraser, Expander, and Shortener.

Improve your store’s Google ranking & boost sales through our eCommerce tools. Generate engaging, SEO-optimized titles, descriptions, and benefit bullet points for your products. When your content is synchronized and semantically accurate, Google will place your products high in the search results.

Writing ad copy that converts is not easy. Boost your conversions, increase your CTRs, and reduce your cost-per-click with our high-quality ad copy tools for Facebook, Google, and other media. Your brand will experience immediate results with these ad copy tools.

Create high-converting landing page copy with minimal input. Furthermore, generate new ideas for your new startup and use our growth ideas generator to get innovative ideas to grow your business.

Reply

or to participate.