OpenAI Consulting

• Global AI Keynotes: CEO of Cazton delivered keynotes for Global AI Developer Days in Fall 2022 in Europe, United States and Latin America. The keynote compared the AI offerings of AWS, Azure and GCP. The demo included multiple programming languages, including Node.js, Python and the latest .NET (C#) framework.
• OpenAI offerings include GPT-3, an AI language model that can generate human-like text, and DALL-E, an AI image generation tool.
• ChatGPT, an OpenAI product released in November, 2022, reached one million users in a week. Facebook and Instagram took 10 and 2.5 months respectively to hit the same milestone.
• At Cazton, we help Fortune 500, large, mid-size and startup companies with infrastructure as code and AI development, deployment, consulting, recruiting services and hands-on training services.

Introduction

Imagine being able to build almost any digital asset just by providing prompts in natural language. Language models (LMs) like T5, LaMDA, GPT-3, and PaLM have demonstrated impressive performance on such tasks. Recent studies suggest that scaling up the size of the model is crucial for solving complex natural language problems. This has led to the development of Large Language Models (LLMs). These models are trained on a very large dataset of text.

Why LLMs? It has traditionally been hard for AI models to generate human-like text, often lacking fluency, coherence, and context. LLMs have achieved impressive results (though they have several weaknesses that we will discuss below) in a variety of natural language processing tasks, such as language translation, summarization, and answering questions. Notable LLMs include GPT (Generative Pre-training Transformer), BERT (Bidirectional Encoder Representations from Transformers), and RoBERTa (Robustly Optimized BERT Pre-training Approach). LLMs are pre-trained to learn to predict the next word in a sequence, given the context provided by the previous words. This pre-training process helps the model learn the structure of language.

Denoising Diffusion Probabilistic Models (or DDPMs, diffusion models, score-based generative models or simply autoencoders) have demonstrated remarkable results for (un)conditional image, audio and video generation. Popular examples (as of Dec, 2022) include GLIDE and DALL-E 2 by OpenAI, Latent Diffusion by the University of Heidelberg, ImageGen by Google Brain and Stability AI by Stable Diffusion.

Generative Adversarial Networks (GANs): The technology of the year for 2022

Before GANs or generative AI, most models were used for processing and analyzing pre-existing data, but they were not able to create something entirely new. Generative AI is a type of artificial intelligence that generates new content, such as text, images, or music, based on a set of input data. This is a significant breakthrough in the field of AI because it allows machines to mimic human creativity and generate novel, human-like output.

ChatGPT reached one million users in a week, compared to Facebook's ten months and Instagram's 2.5 months to hit the same milestone. According to OpenAI, more than 1.5 million users were actively creating over 2 million images a day with DALL· Stable Diffusion by October, less than 2 months from its release date on August 22, 2022. Midjourney reported in September that it had over two million Discord chat members. Clearly, generative AI is a very new and disruptive field.

Features

Generative AI has some great use cases:

• Text and code completion: OpenAI can generate text and code completion based on any input text prompt. This can be used for a wide array of use cases including content or code generation, summarization, expansion, conversation, creative writing, style transfer, and more.
  
  

  Good result: Here's an example of a prompt we tried that had a good result, "Write a tagline for a coffee shop":

  The response from OpenAI is shared below:

  Here are a few ideas for taglines for a coffee shop:

  • "Wake up and smell the coffee!"
  • "Brewed to perfection"
  • "Life is too short for bad coffee"
  • "Savor the moment, one cup at a time"
  • "Where every cup is a new adventure"
  • "Elevate your daily grind"
  • "Coffee: because adulting is hard"
  • "Find your happy place, one sip at a time"
  • "Coffee: the fuel that keeps you going"
  • "Life's too short to drink bad coffee"

  Not so good result (see screenshot below): Here’s an example of a prompt we tried that missed quite a few things, “Machine learning frameworks that are built on Java/JVM and their use cases”. Some notable ones are:

  1. It misses very important libraries that are extremely popular. For example, Deeplearning4 and Spark MLlib, FlinkML, Apache SAMOA and MALLET. Please note that some of these are Scala libraries. However, since both Java and Scala run on JVM, they can be used natively. Please refer to this article for more details.
  2. At best, it includes only partial features supported by these libraries and fails to list the major use cases.

  Note: The above problems can be fixed easily. Generic models are good at answering problems but are not specific enough. That’s why we help our clients build customized models that use the generic ones as starter models.

  
• Image-generation: Simple based on a text prompt, the three major use cases in image generation are:
  • New image generation.
  • Editing an existing image.
  • And creating multiple different variations of an existing image.

  Prompt: "Futuristic flying cars with lambhorgini and ferrari kind of looks flying above a blue lake"

  Conclusion: While these pictures, which were generated in seconds, are extremely creative and speak volumes about how good a generative AI can be, upon closer inspection, they are not pictures that can be used in production. However, generative AI is in its infancy, and it is clear that this will only get better. Therefore, the future is promising.

  

  Prompt: "Design a dashboard with multiple different graphs for a drone manufacturing company"

  Conclusion: While this is great for inspiration, the result is almost unusable (zoom in to see the details). So, clearly it’s a long way to go before we say goodbye to our designers.

  Note: Despite the designs not being of high quality, it can speed up idea sharing within teams. We have helped clients integrate these tools into their existing tools and processes and have seen exponential improvement in productivity. Customers who are not designers (including BAs, developers, executives, etc.) can now easily communicate their generic ideas to our designers and developers, who can help create a high-quality design and app.

  

Proven success strategies for enterprise

While OpenAI is good and will get better with time, it’s not for everyone. We have listed some client concerns below and the solutions:

• Offline access: Some clients prefer not to make a call to an external API (like OpenAI). Can we help you with your own model that could be used offline? Absolutely. We can help create a solution based on open source pre-trained models that can be used offline in a multitude of devices including all major operating systems, Docker, IoT devices etc.
• Cost reduction: OpenAI solution is based on a pay-as-you-go model? However, some clients who would want to use the Generative AI solutions extensively may want to save that cost. There would be no ongoing costs like per-image cost. Superb! That means the cost of running the model is effectively zero. (1)
• Alleviating lack of accuracy: OpenAI models and other AI models are not 100% accurate. There are cases where users only like a few of the generate pictures. A common question customers ask is why do they have to pay for discarded pictures. That said, we have a a great solution which can help alleviate such costs (https://cazton.com/contact-us).
  
  We can help with customized models that can alleviate lack of accuracy and this custom model can be trained on clients’ business domain. Two popular solutions are:
  
  
  • OpenAI model extension: Creating a customized model on top of OpenAI model.
  • Open-source model extension: Creating a customized model on top of a pre-trained open source model.

Limitations, pitfalls and threats

While this is surely a breakthrough, there is a long way to do. We have share a partial list of limitations below:

• SEO challenges: Search engines do not like AI-generated content, and there are talks about the content being watermarked so it can be downgraded. If that happens, most of the content created may not be great for websites that require search engine optimization (SEO). However, there are some ways that AI content may be used without jeopardizing SEO.
• Tokens and cost: The cost is based on the usage of tokens because the OpenAI model understands and processes text by breaking it down into tokens. The number of tokens (refer to the screenshot below) processed in a given API request depends on the length of both the inputs and outputs.
  
  Limitation: The model's maximum context length is, for most models, 2048 tokens, or about 1500 words. The text prompt and generated completion combined must not be more than the maximum context length.
  
  Note: Our team has automated workarounds that can help alleviate this limitation.
• Lack of accuracy: One of the major strengths of GANs is also their biggest weakness. They are creative like humans, but they are not very accurate. What adds pain to misery is that, despite being inaccurate, they appear to be extremely confident. In some cases, accuracy is as low as 25% for code prompts, but you will still be paying for 100% of the use. Our team has devised many ways to mitigate this and help our clients cut unnecessary costs. This is particularly helpful for large customers with high usage.
  
  We asked ChatGPT when (in which month) Stack Overflow conducts its survey every year. The answer was January-February. When we let it know that it was wrong and it was conducted in June 2021, it accepted the mistake and confidently stated that, generally speaking, it is conducted in Jan/Feb, but in 2021 it was June. The truth is it was conducted in May/June 2021, so we tried to correct it again. It accepted the mistake. However, when you ask the same question from a different user name, it still gave the incorrect answer.
  
  Conclusion: It is a great tool for someone who knows how to use it and what to use it for. However, a beginner may end up very confused. There are a ton of meme tweets joking about how bad ChatGPT is at math.
• Lack of specific business domain: While generic models are great, they lack domain expertise. This can be alleviated through the use of customized models that our team can help build.
• Model is offline: Not only that the model is offline but it also trained on old data. Have a look at the screenshot. We asked it to list top phones on Dec 30, 2022 and it listed iPhone 12 instead of iPhone 14.
  
  

How Cazton can help you with OpenAI?

Our team of experts is extremely fortunate to work with top companies all over the world. We have the added advantage of creating best practices after witnessing what works and what doesn't work in the industry. We can help you with the full development life cycle of your products, from initial consulting to development, testing, automation, deployment, and scale in an on-premises, multi-cloud, or hybrid environment.

• Technology stack: We can help create the best machine learning stack using top technologies, frameworks, and libraries that suit the talent pool of your organization. This includes OpenAI, Stable Diffusion, PyTorch, TensorFlow, Keras, Apache Spark, Scikit-learn, and/or others.
• Develop models, optimize them for production, deploy and scale them.
• Best practices: Introduce best practices into the DNA of your team by delivering top quality machine learning (ML) and deep learning (DL) models and then training your team.
• Incorporating ML/DL models in your existing enterprise solutions.
• Customized AI Solutions - The Future of Business Efficiency: Develop enterprise apps or augment existing apps with real time ML/DL models. This includes Web apps, iOS, Android, Windows, Electron.js app.

Note: The results displayed above are based on the queries made on or before December 30, 2022.

1. Requires Master Services Agreement and Statement of Work.