If you’re a technolgy user, you might have noticed that everything is AI these days. I got an email from Microsoft today that shows the new name of one of their partner programs: Microsoft AI Cloud… The financial press is calling this the “shot in the arm” the IT industry needed. Some analysts are crediting AI for averting a recession. But what’s real and what isn’t?
Current AI Models Are Based on Probability
I asked ChatGPT (chat.openai.com) the following: “Is this statement true: current AI models are based on probability?” Here’s the response (after I asked ChatGPT to reduce it to 3 sentences):
Yes, the statement “Current AI models are based on probability” is true. Many contemporary AI models, including language models like GPT-3.5, rely on probabilistic methods to estimate the likelihood of different outcomes and generate contextually relevant responses. These models use large-scale datasets to infer probabilities and produce human-like text.
Actually, I could have written those sentences, but the entire interaction with ChatGPT took about 10 seconds. The sentences are basically well-structured (at least they would pass Mrs. Crawford’s red pencil basically unscathed). The longer, four paragraph version included some repetition.
So at least I would say that the current AI models are useful for generating text describing relatively widely known information.
Limited by their Training
The last paragraph of the “long version” read as follows
However, it’s essential to note that the field of AI is continually evolving, and new techniques and models might have emerged since my last update. Therefore, it’s always a good idea to verify the latest information from more up-to-date sources.
ChatGPT wasn’t always so forthcoming. It’s training ended in September 2021; that just means that it was fed data until that date. It can’t possibly (in this version) know about anything that happend after that date.
To show you the difference, take a look at this conversation about the maximum contribution to a retirement plan from February of 2023:
And from today
You can see the evolution, or as one author put it, the ChatGPT “lobotomy.” In the first example, ChatGPT asserts (for example) that the 2023 IRA limit is $6000. The problem is that Congress raised the IRA limit to $6500 and the 401(k) limit to $22,500.
The Current State For Business
A spate of AI applications is coming online. These include AI apps to generate photos and video from prompts. Asking DALL-E to generate a picture of someone fly fishing in a river gives the following:
Blow those up a little and have a look, particularly at the face of the 2nd fisherman. Or is it a woman? Android? Hard to tell, really.
Other applications listen to Zoom, Teams, etc., meetings and provide summaries. I haven’t tried them yet; I wonder if they work better than Siri? Those will be useful if the accuracy is good enough.
You can build Power BI scripts with ChatGPT, have it write Excel macros, use Git Copilot to help you write programming code, and a host of other items. But (as I’ll demonstrate in later posts) you need to know at least enough to give the answer a proofing. Otherwise, you may accept something that just doesn’t work quite right.
People Make Mistakes Too
This is one item that most writers that are pro-AI, but just a bit skeptical finally get around to. If I ask my doctor to look at a rash, what are the chances that I get a correct diagnosis? Suppose he glances at it and says, “That’s poison ivy.” What are the chances he is wrong? Let’s suppose he gives me some steroid cream and an antibiotic just for good measure. If the rash goes away, was he right?
Or what if you ask your attorney a question and get the wrong answer? Lawsuits against professionals (attorneys, doctors, CPAs, etc.) are filed every day.
AI may not always be right, but at some point it will likely be good enough. You will be able to give an AI a copy of your financial statement and ask, “What could we do to increase sales?” The AI will have the ability to reply.
But how will you validate the answer? That’s the real question for business.