The Rise of Whatever

Broadly agreed with all the points outlined in there.

But for me the biggest issue with all this — that I don't see covered in here, or maybe just a little bit in passing — is what all of this is doing to beginners, and the learning pipeline.

> There are people I once respected who, apparently, don’t actually enjoy doing the thing. They would like to describe what they want and receive Whatever — some beige sludge that vaguely resembles it. That isn’t programming, though.

> I glimpsed someone on Twitter a few days ago, also scoffing at the idea that anyone would decide not to use the Whatever machine. I can’t remember exactly what they said, but it was something like: “I created a whole album, complete with album art, in 3.5 hours. Why wouldn’t I use the make it easier machine?”

When you're a beginner, it's totally normal to not really want to put in the hard work. You try drawing a picture, and it sucks. You try playing the guitar, and you can't even get simple notes right. Of course a machine where you can just say "a picture in the style of Pokémon, but of my cat" and get a perfect result out is much more tempting to a 12 year old kid than the prospect of having to grind for 5 years before being kind of good.

But up until now, you had no choice and to keep making crappy pictures and playing crappy songs until you actually start to develop a taste for the effort, and a few years later you find yourself actually pretty darn competent at the thing. That's a pretty virtuous cycle.

I shudder to think where we'll be if the corporate-media machine keeps hammering the message "you don't have to bother learning how to draw, drawing is hard, just get ChatGPT to draw pictures for you" to young people for years to come.

People will write lengthy and convoluted explanation on why LLM isn't like calculator or microwave oven or other technology before. (Like OP's article) But it really is. Humans have been looking for easier and lazier ways to do things since the dawn of civilization.

Tech never ever prevents people who really want to hone their skills from doing so. World record of 100m sprint kept improving even since car was invented. World record of how many digits of pi memorized kept improving even when a computer does that indefinitely times better.

It's ridiculous to think drawing will become a lost art because of LLM/Diffusal models when we live in a reality where powerlifting is a thing.

>LLM isn't like calculator or microwave oven or other technology before. (Like OP's article) But it really is.

I would not buy a calculator that hallucinated wrong answers part of the time. Or a microwave oven that told you it grilled the chicken but it didn't and you have to die from Salmonella poisoning.

You would if you were able to do basic mental maths and you learned to engage and run basic sanity checks. That's still much faster than grabbing the slide rule. (And it's not like people are infallible)

Obviously if one product hallucinated and one doesn't it's a no brainer (cough Intel FPUs). But in a world where the only calculators were available hallucinated at the 0.5% level you'd probably have one in your pocket still.

And obviously if the calculator hallucinated at the 90% of the time for a task which could otherwise be automated you'd just use that approach.

I've seen my accountant's fingers flawlessly fly using a calculator to track expenses down to the penny. Few people have those mental skills even in the days before calculators - either mechanical or digital.

Slide rule are good for only a couple of digits of precision. That's why shopkeepers used abacuses not slide rules.

I have a hard time understanding your hypothetical. What does it mean to hallucinate at the 0.5% level? That repeating the same question has a 0.5% chance of giving the wrong answer but otherwise it's precise? In that case you can repeat the calculation a few times to get high certainty. Or that even if you repeat the same calculation 100 times and choose the most frequent response then there's still a 0.5% chance of it being the wrong one?

Or that values can be consistently off by within 0.5% (like you might get from linear interpolation)? In that case you are a bit better than a slide rule for estimating, but not accurate enough for accounting purposes, to name one.

Does this hypothetical calculator handle just plus, minus, multiply, and divide? Or everything that a TI 84 can handle? Or everything that WolframAlpha can handle?

If you had a slide rule and knew how to use it, when would you pay $40/month for that calculator service?