For more information on ChatGPT and GPTs read our previous article.
It’s been a long time since I’ve been this geeked out about a technology. I haven’t been this excited since the Pentium.
I’m going to try my hardest to write this article hype-free, but it’s damned hard. I haven’t geeked out about performance specs since bragging about mHz and gigabyte numbers.
Ladies and gentlemen we are at the threshold of a new human-computer interface.
I say that without hype.
Moore’s Law for AI
Moore’s Law was a handy metric for predicting computer processor performance increases. The former CEO of Intel said CPUs will double the number of transistors on a chip every year.
GPT-4 is rumored, rumored, to be train on more than 100 trillion parameters. That’s trillion with a ‘T’. That’s 10 times more than the 175 billion parameters GPT-3 is trained on. GPT-3 is trained on 10 times the parameters Microsoft’s most powerful natural language generation model.
GPT-4 was trained using a custom built Microsoft Azure supercomputer. They invested real money into building this new model!
GPT-4 is a multimodal AI. Which means it can incorporate computer vision as well as language, and perform multiple tasks. GPT-4’s latest feat is being shown a refrigerator of food and being asked meals it can make from available ingredients. TheVerge put this capability to test and it was fascinating.
Intern Level Intelligence
Yesterday’s demonstration had GPT-4 completing many tasks that are available in apps like Apple’s Magnifier and Midjourney, but OpenAI’s main selling point is GPT-4 understands the context of your prompts. OpenAI says GPT-4 has, “the same level of understanding and context as a human volunteer.” An intern.
This is very powerful. Conceptually GPT-4 will be able to explain its environment and provide insight on what it sees. At its rumored 100 trillion parameters, GPT-4 can handle more text and produce novel results. TheVerge staff asked it to write summaries using only a single letter, and when told it made mistakes the AI happily acknowledged its mistakes and made corrections.
Still No Fear
GPT-4 suffers from the same frailties as its predecessor. GPT-4 is still prone to hallucinating. Tests still show its prone to rambling and making things up if it doesn’t know what it’s talking about. The AI’s underlying information is still suspect. Even though it’s using a more recent corpus, it’s only as knowledgeable as what it’s been told. Finally, GPT-4 is still biased. To be fair, bias is a hard thing to remove from these systems. OpenAI publicly admits GPT-4’s bias as well as noting their efforts to root out much of it.
Don’t worry about losing your job to GPT-4.
Real World Uses
My friends and I have been putting these GPTs through their paces. (I’ve spent more time talking to AIs than humans these past few months)
“I have a too long CV problem from my first career where long CVs were the norm.”
One of my friends is making a career transition from being a private school director and wanted help rewriting her resume. She turned to ChatGPT to streamline her resume and condense her 20 year career to two pages of text. The results have been overwhelmingly positive. Her resume is meatier and highlights her experience in more direct and concise tones.
Another friend is a web developer and is using ChatGPT to rewrite customer provided copy for their websites. She said ChatGPT’s output is punchier and more persuasive than the original copy. Her customers granted her consent to have ChatGPT eyeball their content.
“Bet. I’m gonna keep messing with a few formula for chatgpt prompts. I’m getting closer and closer to the way I want it.”
I’m working with an artist on another project and he’s using ChatGPT to pipeline Midjourney edits on his hand drawn sketches. The output has been fantastic. He isn’t using ChatGPT to render new stuff, but make adjustments. He’s adjusting light and shadow, lighting positions, and color in his sketches.
The Sky Is the Limit, Don’t Fly Close to the Sun
I’m bullish on AI technology. It’s hard not to get excited during this period rapid innovation. If we can survive through all of the “how to make money with ChatGPT” scams some really helpful products will shake out. AI done right will enhance the human experience.
We have to seriously consider bias in all forms. It’s easy to throw DEI words an AI and push them aside, but there are more than black-white biases. Ageism can be a real problem that should be met. Education and geography biases are also things we should keep an eye on.
GPTs stil hallucinate and make stuff up. Engineers still haven’t figured out how to add provenance and veracity to their models. (I’m working on those)
We are witnessing the beginnings of a new human-computer experience and it’s fun being on the ride. We should consider the ethical and societal impacts AI will inflict on the world. For more on this read Macie Jones’ thought piece on Microsoft firing their AI-ethics team.