Flipboard Friday this week shares my ChatGPT journey. I started using Chat-GPT last December and wrote about How ChatGPT Will Change Your Website, and to say it has made my Content Marketing Director job easier would be to have a gift for vast understatement. I've used OpenAI's gpt 3.5 chatbot (or ChatGPT 3, not sure) daily and am so addicted to their generative artificial intelligence and language model that I fear withdrawal. "Where the awesome interacts with every day" was something I heard yesterday as an apt description for describing the experience of using ChatGPT.
I was curious enough to get on OpenAI's waitlist. Still, I quickly discovered the liability of depending on an ai chatbot more than a hundred million people were using - it went down frequently. Like most marketing teams, ours works on deadline, so using chatgpt became so mission-critical that I paid $20 monthly. I didn't even wait for accounting to approve the money; I threw down my card and am happy I did.
Let's call my Chat-GPT subscription the chatgpt plus experience. Since subscribing, I've only encountered OpenAI's "too busy" signal twice, and once was on ChatGPT-4 this week when you, me, and everyone we know were using ChatGPT-4.
I am not using ChatGPT's large language model to pass the bar exam but to accomplish real-world marketing tasks faster and better. I have yet to test Google's Bard or any of the Microsoft-enabled tools such as Bing, but the Big M is doubling down on generative ai, and suddenly they feel relevant again, but ChatGPT has a moat around me. Something new must be beyond amazing for me to even think about switching.
My ChatGPT-4 review explains a little about generative pre-trained transformers with the caveat that I'm more of a right-brained marketing geek than a left-brained technical. However, marketing these days requires even those far right-brained marketing geeks to know enough tech to be dangerous, and brain cross-training might be ai tools' most significant contribution as I described in AI Gets Practical. So when our tech team or my way to wise for me boss Eric discusses anything from shards to NoSQL, I'm only a question with a natural language response away from not groping in the dark. Use the links below to discover why I'm addicted to ChatGPT-4; it is probably unhealthy.
AI systems start as ai models, and as any newborn chatbots depend on human "labelers" to provide "supervised learning," teaching the model "more this, less that" lessons. Step two is where Pavlov's puppy becomes a dog because this step is where the AI reward model begins.
An AI reward model represents the artificial intelligence goals and objectives. In reinforcement learning, a subfield of machine learning, where AI learns to make decisions by interacting with an environment and receiving feedback in the form of rewards or penalties, so chatbot operant conditioning.
I'm not clear on if the chatbot gets a ding dong for good answers and a shock when the AI model is off the mark, but you want your AI dog to be accurate, do things the "right" way, avoid smartest AI in the room exploitations, encourage exploration to learn about "the environment" and discover better solutions, and learn fast and efficiently, so we don't have to mortgage the house to pay for the CPUs needed to produce quality answers.
The third phase, the PPO or Proximal Policy Optimization, is where I get lost. This "on-policy, model free-algorithm" created by OpenAI to coach "deep neural network policies" sounds like we're speaking at the UN. Instead, this phase is about adding speed and fury to millions' human-like conversations with ChatGPT as I write this.
The idea is to "train" an ai model on content, reinforce good and wrong answers, and turn the amp to 11 so millions of simultaneous users can get their questions answered without needing the Pentagon's computing power. However, I suspect OpenAI's server farms rival anything in Washington. Read Eric's AI Guide to What's Coming for more on where the AI tsunami lands next.
I depend on ChatGPT to do three digital marketing tasks, so here they are in order of my use case priority:
Tech Geek Translator.
Content Marketing Assistant (discover all the AI tools I used to write online content).
Trained at P&G and M&M/Mars, I'm a right-brained marketing geek, but I've been fascinated with technology since purchasing a PC with five thousand bucks I didn't have in 1984 (an Apple II I wish I still had). After developing a simple territory management system with Lotus 1-2-3, the first spreadsheet, and DB1, I was promoted to M&M's national office to help create the Sales, Management, Analysis, Reporting, and Telecommunications (SMART) system before many reading this post were born.
I work for a software and website development company, so almost everyone here has forgotten more about IT and tech than I'll ever know. Taking notes and running to ChatGPT to translate geek speak is a lifesaver. The experience of having a human-like conversation with a chatbot is much more efficient than wading through Google search. I've noticed that I'm five times more likely to ask questions and learn from ChatGPT's answers than wade through SERPS. I use Google to check my chatbot conversations but rarely start there anymore.
I'm typing this post on Clearscope after asking it to help me know the keywords related to a "ChatGPT-4 Review." SEMRush and UberSuggest have similar write-to keywords for a grading tool, but I prefer Clearscope's UI, fantastic keyword research, and grading system.
I feed my Clearscope draft into Grammarly, another artificial intelligence tool, to clean grammar, ensure I'm using an active voice instead of a passive voice, and be specific. For example, OpenAI's chatbot uses "it" and "they" and other non-keyword stop words more than I like, so I clean that tendency in Grammarly.
I need to be careful because Grammarly doesn't know keywords, so I feed my cleaned copy back into Clearscope, hoping my A+++ grade is okay with how I cleaned the document. Finally, I use Neil Patel's Ubersuggest for meta descriptions and headlines. Sometimes I write first and add meta and headlines after; other times; I start with Ubersuggest. I've been thinking about my ChatGPT-4 review for days, so I started writing and will backfill headlines and meta.
Read my Guide to Modern Ecommerce for a favorite idea generator example. I told ChatGPT-3 I'd like to create a sustainable mouthwash to disrupt the billion-dollar American mouthwash category. EcoFresh, fresh breath, and a clean planet was the brilliant name and tag created by ChatGPT.
Next, I used OpenAI's Dall-e text-to-image tool to create the blog post's logo and hero image. Nothing from OpenAI's chatbot or image generator goes directly onto our website. I work the copy to put our active voice in, use Grammarly to clean the grammar and check my consistently poor spelling, and photoshop to clean Dall-e's images and create versions we need for desktop, mobile, and social media.
The latest version ChatGPT-4, needs 30% less Grammarly, and answers are about 50% more robust, meaning there's more there "there" than ChatGPT-3. I'm learning too. Yesterday I asked ChatGPT-4 to write a script for data engineering and may use that prompt for the other four whitepapers I'm working on because the dataset was complete. Asking OpenAI's AI tool to write a script provides the thirty thousand foot view and maps every road I must take in that journey.
ChatGPT-4 has been sluggish as everyone with a paid subscription tries new features, but I've only received a "too busy, try again later" message once, and ChatGPT-3 flies. ChatGPT-3, I go back and forth by changing the URL, and answers like it is on speed now, but OpenAI's latest generative pre-trained transformer answers are more informative and need less clean-up. If that sounds like I'm already addicted to 4, you're on it.
I haven't used multimodal or any of the fringe new features yet because I work in a production B2B marketing shop where today's deadlines will give way to tomorrow before I know it, so I'll bank having a better geek speak translator, content marketing assistant, and idea generator before worrying about multimodal.
Check out Sign of the Times Flipboard magazine for great posts on OpenAi's LLM's newest features.
It's all about the prompts; we've seen new tools to help find or create great prompts. Take away my life marketing soap, candy, and tech; those EcoFresh prompts wouldn't produce much. Stealing great prompts is highly recommended and easier, but there's a rub.
AI isn't static. Ask ChatGPT or Dall-E the same thing twice, and you'll receive different answers, so I've learned to test fire and combine ChatGPT's solutions with Google's search engine flood to refine responses while checking accuracy. I stay awake over OpenAI's biases and veracity because nothing goes from chat.openai.com to our website, but those outputs make my life easier, faster, and more fun.
If you visit our Sign of the Times Flipboard magazine, where we share the ML and AI conversations, you'll ask what ChatGPT can't do because people are testing everything from diagnosing disease to taking the bar exam. We'd ask ChatGPT how to shower soon, but let's ask how Amazon and B2B software as a service (SaaS) and web designers like WTE will use OpenAi's applications program interface (API).
Internal Search & Interactive FAQ.
ChatGPT Content and Social Media Marketing Assistance.
ChatGPT Code & Multimodal Web Design Assistant.
We've been working on "Ask Eric," where we use ChatGPT's generative ai to answer questions about WTE's content. Eric told me ChatGPT-3 would've needed a second mortgage to pay for using OpenAI's generative pre-trained transformer to create a more interactive internal search and FAQ experience, but 4 shows real promise.
Internal search on any website is a cocaine bear - easy to do poorly. But interactive and personal is where every website is going, and AI will help us get there. And better internal search and frequently asked questions are the logical starting point for our more interactive future.
We've come a long way since I bought PPC ads by hand. Everything moves too fast today, so algorithms control everything. Soon we'll use AI tools to design our content marketing campaigns, algorithms to buy ads, and machine learning to adjust everything from headlines to images based on results automatically.
Don't read my first paragraph; marketers like me will become extinct because the opposite feels authentic. However, if you are a fellow marketer knowing how to use ChatGPT will be crucial to getting hired, doing your job, and creating successful campaigns.
Our coders use GitHub's Pilot, a ChatGPT-enabled code assistant, because when a misplaced semicolon stops everything efficiently, finding such small mistakes and writing better code faster is crucial to thriving in the highly competitive custom software and web design business.
Like marketing, don't read that paragraph as the end of coding and coders; the opposite is true again. Coders who know how to code faster with AI will be priceless. And I haven't touched on AI security, another area we see gaining vast efficiencies and protections thanks to AI tools.
Can ChatGPT become an effective CEO? Outstanding CEOs have two jobs - vision thing and motivation. ChatGPT is a great help in creating "the vision thing" or the direction a company and brands should take, but motivating and leading are different. Can we use AI to inspire and motivate others? I'm not saying a stiff "no," but I see motivating and inspiring people as a reach.
What about you? How are you using ChatGPT? Subscribe to the WTE tribe to comment or email me at martin (at) wte.net. Have a great weekend.
Chris Duke is a friend I met at one of Phil Buckley's genius lunches or one of the SEO meetups Phil used to host. Phil is one of the best SEOs in the world, a good friend to me WTE founder Eric, and a great connector. He also contributes to the WTE BLog; read his SEO Secret Weapon Advanced Search Operators post to see how a pro uses search to help SEO. Chris is a fantastic cook and entrepreneur who knows I will forget more about technology than ever. He wrote me a great email about his highly technical use of ChatGPT, a perfect left-brained compliment to my natural right-brain marketing take. Here is how ChatGPT helps Chris. I'm lucky to have Phil and Chris as friends!
After I posted Chris's note to LinkedIn Phil shared a ChatGPT story.
I switched to chatGTP4 last week to do some translation work from a family history that was in typewriter format from the 1930s that I had scanned. I can copy and paste that Canadian French PDF typewritten page into chatGTP4 and it's is amazing. Not only does it do pretty flawless Canadian-French to American-English translations, it also can figure out the weird characters that come from a non-perfect scan. For example, the copy of 1680 becomes l O 8 0, but because of the other context around it helps it understand that it should be 1680 - amazing.
I read your post and you asked for feedback - so I thought I'd share.
First of all - yes, I am late to the party. I applied for beta access a while ago and yes - OpenAI gave it to us. So, we can login and look at the latest release.