Saturday, February 18, 2023

You Really Need to Check in on AI Right About Now

I have some links to offer for cutting-edge artificial intelligence (AI) news. If you follow the field closely, these links are incrementally interesting. And if not, they're a good jumping-in point to touch base with our current standing in a very fast-moving field.

When it comes to ChatGPT and other text-generating AI, the top-line takeaway for non-experts is "sometimes it comes up with infuriatingly false results."
The philosophical angle on this is: how did we ever imagine artificial intelligence would be reliable? The very nature of human-style intelligence is its spotty inconsistency. You want 100% assurance? Direct cold silicon to blindly crunch numbers according to narrow rules. But if you want a computer to think more like a human, be prepared for it to get stuff wrong. Duh.
A friend recently used ChatGPT to query a very large data set drawn from a popular food web site (a good one, I'm told!) for the best Sichuan restaurants in Queens. It did a remarkable job diving into terrabytes of noisy discussion and coming up with a top five. One of those five, however, was a totally-not-Thai catering service in Ohio. Whoops!

If you tried to accomplish this via raw computing on some supercomputer, you'd need to construct so many rules - and do so much human filtration on the results - that it wouldn't be worth it. ChatGPT, let loose with minimal guidance to just sort of wing it, did pretty well...aside from the horrendous wrongness.

But here's the thing that's not discussed much outside comp-sci circles. You can stave off some of the wrongness via slightly more guideful guidance. Reverting to the dawn of computing, we're discovering all over again that it matters how you ask the question.

And, delightfully, work on this issue isn't being done with nerdy math formulas or COBALT commands. It's pretty damned liberal artsy, actually. "The Art of ChatGPT Prompting: A Guide to Crafting Clear and Effective Prompts" is so friendly and armchair-readable that you'd never imagine it represents cutting-edge tech/computer research.

But attention right this moment is not on ChatGPT so much as on, believe it or not, Bing. Microsoft just added god-level AI chat to their otherwise anodyne search engine (here's the announcement). Even if you're uninterested in AI, it's still an important development, because this is the moment when Bing became a real competitor to Google.

Here's a good quick treatment of how that's going, from Daring Fireball. The links therein are particularly dandy, so you may want to dive in for solid nutrition on this Bing move and on the AI chat field, generally. It's easily worth your attention given that we may be on the brink of a tech advance as significant as the moon landing.


Machines will never be conscious. Consciousness is not an emergent quality. We are in consciousness; consciousness is not in us. But, clearly, there's a whole lot you can do via computer intelligence - and also plenty of creepy risk (that last Daring Fireball link offers an eerie taste), with or without the mysterious spark of Awareness.

No comments:

Post a Comment