Some say it will be the savior of mankind, others the cause of its demise. It's what's keeping the economy afloat, and also what's going to cause it to collapse. It will have a major impact on you whether you are a businessperson or a student, a techie or a luddite, a retiree or an entrepreneur. And it's already embedded in almost everything we do, from checking the weather to getting directions, from choosing which is the best vacuum cleaner to suggesting a vacation itinerary. The most recent two-letter combination to become part of our everyday speech, to GE and ET, VW and UN, add AI.
The term "artificial intelligence" itself isn't particularly new, coined in 1956 by John McCarthy for a workshop on the topic at Dartmouth. It evolved in fits and starts, and until relatively recently was strictly the province of backroom nerds. Then just three years ago ChatGPT was demoed, attracting over a million users in 5 days. And now it seems that every search, every customer interaction, every shopping experience, every routine online engagement has an AI overlay.
But while AI is most definitely prime time it is not always ready. It's hard to remember a time when products were endlessly tested before they were released, so that the version we encountered was virtually glitch free. We then moved into the "always beta" era, where what we are using is very much "of the moment," with bugs and glitches assumed, and patches and updates expected. And nowhere is that truer than with ChapGPT, Copilot, Gemini and their ilk.
Much has been written about AI "hallucinations," where references are made up out of whole cloth. But each version makes mistakes when the answers are in plain site as well. If you've played with any of it you've likely seen a summary of a conversation that is mostly correct, yet screws up some obvious facts, such as a daughter who is actually a wife. And the more you mess around the more you see random glitches.
As a "for instance," when a friend was getting a new boat I asked Gemini to help me create a logo. I gave it the prospective name and a style, then asked it to create something. The first attempt was impressive on its face, but needed refinement. I typed back my suggestions, and a new iteration appeared. Not quite right, so I tried again: the exact same thing came back. When I noted that non-response, I got, "You are absolutely right, and I am truly sorry." It repeated back what I wanted and said it would make the fix. Next output: no change again. I flagged it. "You are absolutely correct, and my apologies are not enough. I am failing to follow your instructions, and I deeply regret the ongoing frustration this is causing." It rebuilt it once more, this time making some incremental changes. Still not what I wanted, I asked for tweaks and tried again. That cycle repeated, some 25 times in all, until I got even close. Had it been a probationary employee, I would have cut him/her/it loose around output 14.
In that same vein, a presenter at a recent conference I attended told how he liked to quiz ChatGPT as he commuted in on the train. A passionate Pittsburgh Steelers fan, he asked it a staple of sports radio: which was the best football franchise in history? As might have been expected it returned that the New England Patriots get that nod. But wait, he asked: have you considered turnovers? Rushing yards? First downs? He gave it a bunch of metrics where he knew the Steelers excelled. The AI went silent for a while. When he asked what was happening, it said it was still looking at the data set. Radio silence again. As his commute was coming to an end, he asked one more time. "Sorry, I don't have access to the correct statistics," was the response. As the presenter put it, "I think I created the first AI teenager."
Make no mistake: what AI can do is nothing short of amazing, and we've seen it get better and better with each new release. The trick will be in using it where appropriate, providing the appropriate checks and guardrails, and making sure there is a disconnect switch. Looked at another way, unless it grows opposable thumbs, at least for now, I think we're OK.
-END-
Marc Wollin of Bedford is learning to use the new tools where he can. His column appears weekly via email and online on Substack and Blogspot as well as Facebook, LinkedIn and X.
No comments:
Post a Comment