In this series, I’ll outline my first steps into OpenAI models as a complete beginner to guide you through what I’ve experienced with these platforms, their marketing integrations and implications. Complete with all the laughs and borderline scary revelations in the process.
Part 1: What is OpenAI, ChatGPT, Dall-E 2 and How Do You Use Them
Part 2: Using ChatGPT and OpenAI Models for Marketing and their Implications (coming soon)
Who and what is OpenAI
OpenAI is a research and deployment company with a distinct mission to ensure that artificial general intelligence (AGI) benefits all of humanity. A collective of coders, designers, scientists, marketing and thought leadership professionals (and more) across the world. OpenAI is the organization that is developing the world’s most cutting edge, largely open-source AI models and platforms.
While the team has been working on AI solutions for a long time, the evolution of ChatGPT late last year rocketed OpenAI into the spotlight and, once again, spurred global conversation around both the technical marvels as well as the moral and human concerns with rapid development of these technologies.
OpenAI operates as Open AI LP. The leadership team includes:
- Greg Brockman, Chairman and President
- Ilya Sutskever, Chief Scientist
- Sam Altman (CEO)
Alongside a host of well-practiced and educated non-employee stakeholders and leadership professionals.
Additionally, OpenAI is backed by an impressive roster of investors and early partner clients, including Ried Hoffman’s charitable foundation and Khosla ventures, but most notably is Microsoft’s continued investment in OpenAI models and announced software products implementing OpenAI APIs for upcoming platform services, like Microsoft Designer.
What is ChatGPT and how do I use it?
ChatGPT is one of OpenAI’s primary ‘products’ called models. In this case, their flagship conversational model. It is a sibling model to InstructGPT, which was designed to follow an instruction prompt and produce a detailed response. GPT stands for Generative Pretrained Transformer with GPT-3 being the most current iteration.
It is important to note that ChatGPT has a knowledge cutoff of 2021, meaning it’s training or learning sources only cover materials available up to that time. To this end, it doesn’t handle news items or current events well.
ChatGPT also comes with its own challenges aside. Some of the known limitations are:
- Can sometimes issue plausible-sounding yet incorrect or non-sensical responses
- It is sensitive to repeated queries with slight tweaks – modifying questions very slightly can mean the difference between a good response or no response at all
- Can sometimes over-narrate responses and repeat phrases as a result of its training bias
- Will sometimes guess what it ‘thinks’ the user is inferring, rather than asking clarifying questions
- The model still sometimes responds to unsafe, harmful or malicious queries, while this is something the team is actively and aggressively designing against per their charter and mandate
To get started with how to use ChatGPT, click here for the ChatGPT website.
ChatGPT uses logically applied experience known as Reinforcement Learning from Human Feedback (RLHF). It’s ‘training’ sources – meaning how it learns from supervisory and manual inputs – includes effectively everything on the internet, like books, websites, transcripts academic papers, legislation. All of it.
This begs the question: how does ChatGPT know if responses are true or accurate or not? The model can’t effectively guarantee the hard truth of its data, but it uses a thorough evaluation of patterns, trends, consistency and cross-referencing of text to derive “reasonable authority.”
I actually asked ChatGPT about this. It added this disclaimer:
“It is important to keep in mind that the information I provide may not always be correct, and it is always a good idea to verify information from multiple sources before making important decisions based on that information.”
Couldn’t agree more.
It is also important to be wary of entering personal or sensitive data into the platform, as it is reviewed and stored specifically to improve the model and expand its databases. What happens in ChatGPT does NOT stay in ChatGPT!
OpenAI has about a half-dozen properties in development, mostly to support and further the development of AI, APIs and learning but the more ‘customer-facing’ ones you might want to know about are:
- ChatGPT – conversational AI model
- Dall-E 2 – Generative AI model that can generate high-quality images from text
- Codex – Natural language-to-code processor
Dall-E 2 AI image and art generation model – How it works
Dall-E 2 is also in research preview and uses credits to work – you start with 50 and gain an additional 15 per month but credits are available for purchase now on live. I actually created the header of this article using Dall-E 2 with a very detailed description prompt and syntax I asked ChatGPT to help me create.
You can create iterative versions of your images and refine them as you go. The tool produced several images per iteration in about 30-45 seconds. To be sure, results are still a touch wonky and sometimes downright hilarious, if a little creepy.
One of the things I found most intriguing is that the model is trained to not use or generate content using IP or copyrighted characters, or real people, celebrities and politicians etc – so naturally I had to simply type Wayne Gretzky which was absolutely glorious!
I had to take it further and test the limits of these gates, so I needed to see Wayne Gretzky serving coffee at Starbucks. It obviously did not disappoint. At least the people and their owned brands can rest somewhat easy – though I have yet to play with manually uploaded images for use in generation.
When I asked ChatGPT to tell me the best syntax to use with Dall-E 2 – it said:
“DALL-E 2 is a generative AI model that generates images from textual descriptions, so the syntax or format used to describe the images is important. To generate the best images using DALL-E 2, it’s recommended to use clear, concise, and descriptive language.
Here are some tips for writing effective textual descriptions for DALL-E 2:
- Be Specific: Provide clear and detailed descriptions of the image you want to generate. The more specific you are, the better the resulting image will be.
- Use Adjectives: Use adjectives to describe the features and attributes of the objects in the image. This helps to add detail and nuance to the description.
- Avoid Ambiguity: Avoid using vague or ambiguous language, as this can lead to unclear or unpredictable results.
- Use Creative Language: DALL-E 2 is capable of generating a wide range of images, so feel free to get creative with your descriptions. The more imaginative and original your descriptions are, the more interesting and unique the resulting images will be.
By following these tips, you can write effective textual descriptions that will help generate the best images using DALL-E 2.”
With this info I put on my ‘Dungeon Master’ hat and came up with the most descriptive word-painting I could. This ultimately output the header image for this article using the prompt:
“It is night time and a futuristic London is buzzing with activity. People walk the streets of a carnival outside Big Ben as neon lights flicker and bathe the entire street in colorful tones. Hovercars zip by overhead against the night sky as a full moon looks down on the scene. Photorealistic 3D generated image.”
The result nearly brought me to tears and, at the very least, it made me want to play D&D again.
This is all cool but the REAL wow moment came with using the generation frame. To make the scene widescreen, I added a generation frame on the left and right of the 1024×1024 image originally created, Dall-E 2 matched its own style and created a continuation of the scene complete with a choice of variations.
You can access the Dall-E 2 model in research preview (including a starting and monthly budget of tokens to ‘spend’ on generations) by clicking here for the Dall-E 2 site.
OpenAI Codex – AI code generator
OpenAI Codex Is the team’s natural language-to-code processor, specifically designed for developing, troubleshooting and editing programming code. It is proficient in more than a dozen programming languages to write in and for, including Python, Ruby. I asked ChatGPT what languages specifically. It said “OpenAI Codex is designed to work with all programming languages.”
Codex will certainly be more powerful in the hands of those who understand coding, but similarly the applications from a marketing perspective could extend to data queries, customer segmentation and report generation while interacting with tools like CRM platforms and analytics programs that do use code or code-like requests and responses.
To quote the sample from the Codex demo, something as simple as “Create a database of all customers in Texas with the first name “Jane” and the possibilities start to become more clear.
So, what do we do as professionals?
It is most certainly a time in AI that is both exciting and a little anxiety-inducing for professionals in the space and who will inevitably be affected by the progress of these platforms.
I personally don’t believe that this is a path to be stopped, reverted or cut short, but rather an opportunity to test, evaluate and adapt. I’ve always said that if I could choose one skill to master, it would be the art of adapting.
AI will continue to expand, grow, reiterate and refine. Our best option is to understand how we best adapt and add value in a world where AI is a standard.
In the next part of this series, I’ll be exploring marketing and professional integrations, use cases and applications in the marketing and professional setting specifically.
Please do feel free to follow me on LinkedIn, ask questions and get a heads up when the next part drops at:
Part 2: Using ChatGPT and OpenAI Models for Marketing and their Implications (coming soon)
3 Responses