Is it AI or not

Exactly!

A woman I met at a family event recently asked me what I thought of AI. I started talking about Leibniz and his speculations, Charles Babbage's Differential Engine, Isaac Asimov's robot books. Well, she let me have the limelight; thank you. Later I found out that what she had in mind was ChatGPT and OpenAI. It's amazed millions, created a rush of books (including ChatGPT for Dummies), and fuelled a whole new debate. "AI" has replaced "algorithms", which replaced "apps", which replaced "programs".

Ed

Every technology needs to be classified.

CharGPT is about as useful as OCR. OCR is about 99% accurate. You've just run 200 pages through the scanner. Now what...

Voluminous output, that must be scrupulously checked.

An "advisor", not a "boss".

A robust source of beer bong pictures.

Paul

So you want Fyzuhr to build a toilet that nags you to get a monthly booster clot-shot?

An algorithm would be programmed by some (presumably) human programmer as, essentially, a list of rules to follow.

AI (old name: neural networks) is trained by being given a huge stack of photographs that radiologists have previously examined and pronounced "yes" or "no". The neural net makes guesses about whether each picture is a "yes" or a "no", and somehow learns from its mistakes, so that over time its accuracy becomes better and better.

While the training is simple in principle: pathways in the neural network that led to a correct result are given a boost and those that led to an incorrect response are depressed -- in practice the result is so complex that nobody can determine what caused the NN to reach a particular decision in a particular case.

Or, at least, that's how I understand it. A NN product was offered by the company I used to work for, and the programmer explained it to me that way. Nothing I've seen has told me it's different in principle now, though I believe much bigger computers are being used, and with most of the Internet as a training set. But does anybody know exactly why an AI would answer questions like "Who holds the world speed record for walking across the English Channel" with specific name, date, and time? I don't think so.

If you really meant TrueCrypt, and you use it, you might want to think about switching to the fork called VeraCrypt.

Humans as a rule are pretty dumb. (Dave Barry famously remarked that the three strongest forces in a human are stupidity, selfishness, and horniness. I've seen nothing to prove him wrong.)

But as for the Turing test, I think that ship has sailed. Case in point: A few months ago, at character.ai, I was chatting with Napoleon. Not really, of course, but the responses were consistent with what I knew of his actions and speeches from my reading of multiple biographies of Talleyrand, his foreign minister and nemesis. Seems to me that it passed the Turing test.

What laypeople often mean when they talk about artificial intelligence is not the Turing test, or not _only_ the Turing test but rather something intangible like a sense of self. They don't want to know if a machine can imitate a human, they want to know if it's "alive", whatever that means.

Add to the mix:

• Daniel Dennett, in /Consciousness Explained/ concluded that our sense of self is essentially a bunch of parallel processses that look like a serial process. (Or maybe it was the other way around. Fascinating book, but it's a long time since I read it. I do remember that he wasn't just speculating in a vacuum: he used evidence in standard medical journals with the results of actual experiments done with human perception and cognition.)

• It may all be moot. If the speculation is correct that we're living in a simulation, humans are all already AI.

GPUs were the breakthrough. The Graphics part was a misnomer although the original intent was to speed up the calculations involved in CG. That made them ideal for crypto mining to the point where a GPU shortage was caused by the miners buying the high end boards. They are also good at the vector manipulations needed for training a neural net.

Using something like PyTorch you can experiment on a PC. Even then it's much faster if you have a GPU that supports CUDA. CUDA is an open standard but afaik only Nvidia supports it in their chips.

When you get to something like Chat you're talking many very expensive GPUs, a lot of power, and millions of dollars. Chat isn't aware of recent events since it was frozen with the data available when the training occurred and retraining is very expensive.

Once the model is trained inference, or using the model, is much less intensive. That's the 'Pre-trained' in GPT.

For me the interesting part is pruning the model developed on a huge system to run locally with limited resources. Cell phones are getting to be power enough to do so. Many people didn't realize that for something like speech recognition the audio was sent off to Google, processed, and the text returned. the light dawned when they realized Alexa has very big ears. The desired outcome is to have more functionality at a local level, including very small processors like the Arduino family that don't require huge amout of power to run.

It's a fascinating development but like all disruptors the potential for bad is just as high as good.

'Stochastic' gets used a lot so 100% accuracy isn't even a realistic goal. 'Good enough' is the criteria. If a simple model can tell a cat from a dog

Nags? No, there will be a matching device similar to an insulin pump.

ChatGPT stings even an old computer programmer. It beats the Turing Test by miles. But, as you've grasped, it's not got nearer the truth; it's got nearer the norms of human interaction, in which fake news and camouflage and outright ignorance play such a part. But if we're ever going to set up Blade Runners to police humanity, they'll look for those qualities rather than truthfulness. This is what's so scary about it. It mimics us so damn closely.

Ed

It's not generalized AI as per the Turing test.

That and language changes. If marketers want to start calling Machine Learning (ML) AI instead, well, we'll need a new term when we have one that passes the Turing test. Probably "Overlord."

It is not generalized AI, which is what AI used to mean, though.

My main concern is not the semantics but corporate needs to release unsafe stuff in a rush to be first, and the sort of people who buy that unsafe stuff. There are a lot of unsafe people that like to play with guns. Nothing against guns, just irresponsible owners.

...and I really don't think welding ChatGPT onto Bing is going to be the success MS thinks it is that will finally usher in an age of Bing and Edge supremacy over Google. Just no. I don't think people want it.

It's already too late. Pandora has opened the box. There's no putting it back in.

Be afraid. Be very afraid.

I wouldn't go that far, but standard human stupidity is gonna make this veeery interesting.

Have to get people to poop in a litterbox first ;)

The Turing has been passed quite a while ago. Plus the test is flawed.

AI has never been properly defined and so people are using to describe all sorts of things.

It's a sad reflection of priorities where the primary concern is about being sued rather than making sure patients get the best treatment.

Some lawyers getting caught being dumb is not the same as using a clinically approved tool. A clinician isn't ever going to diagnose a patient via ChatGPT. If they did they deserve to get the book thrown at them.

Er, hospitals.

No it isn't.

Which is likely illegal. Hence all the suits against google et all.

The risk to jobs from AI is massively overblown. People aren't going to lose jobs to AI, they will lose jobs from other people who use AI.

Radiologists' jobs have the potential of improvement with AI.

Exactly. Perfection is not the target. At a minimum an AI that performs as well as a human that never gets tired, grumpy or distracted and is twice as fast would be enough of an improvement.