Hacker News Reader

Comment by A_D_E_P_T | Hacker News Reader

But is it really "AI"? Seems like little more than a fancy image processing method that:

- recognizes a face

- removes a few common wrinkles, e.g. nasolabial folds (-- these in particular are the most common marker of an "aged" face, and people who don't have them or who have broad nasiolabial folds are often said to "age gracefully")

- smooths out and brightens the rest.

It's not creating or generating anything, really. Just running a very adroit filter in real time. To what extent is it, then, AI? We should be wary of the way buzzwords are inserted everywhere, I think. (Remember when it was all "nano"?)

View on HN

The is this AI debate is never ending. At this point AI is any technology that is not trivially explainable and works on poorly defined problems.

This seems trivially explainable and clearly well defined, though? No fuzzy edges here.

If it's clearly defined then write code that implements the same filter. No neural networks or ML models allowed. Then I will believe you.

Actually it would be really interesting to see code that attempts to e.g. classify MNIST digits from first principles. No neural nets, no GBDTs, no clustering. And no cheating by trying to translate the above into if-then statements. Seems next to impossible.

I recall Andrej Karpathy saying that people originally tried to do that for image classification (such as cats/dogs problem) and weren't able to get anything close to high accuracy.

Slightly relevant, Viola–Jones Face-Detection Algorithms pretty much does “manual” face recognition.

It's not really worth arguing about the definition of "AI". All the things you listed as "not AI" could just as easily be a list describing what "AI" can do. Sure, there are more technical terms for these particular "filters" and how they actually achieved, but almost no one cares about making that distinction IRL.

"AI" is going to stick as a broad term, get used to it.

It's called AI because there's an ML model in there somewhere doing some heavy lifting

> some heavy lifting

Nice

That's called just ML... ML doesn't imply intelligence.

> But is it really “AI”?

No, nothing is “really AI” except AGI, everything else is only called AI when its new and shiny. Once its established, it stops being AI.

E.g., rule-based expert systems are no longer described as AI by almost anyone, despite being one of the earliest kinds of “AI” systems. In a few years, LLMs and diffusion-based text-to-image and image-to-image systems won’t be “AI” either.

Those are still AI, they are not ML.

It probably is some sort of convolutional filter chain to do that recognizing and blurring, so you'd deploy it as a TFlite with TensorFlow Lite, which kinda makes it AI :p

Shhhh, don’t give away the secret sauce.

I think you're being a bit pedantic about it because they used the term AI.

It's not AI, it's advanced image processing.

The whole history of AI is one of moving goal posts:

X is hard, only people can do it. X requires intelligence. A computer doing X is AI.

AI researcher programs a computer to do X

Oh, X isn't AI, it's just [describe what AI researcher did]. That's not real AI.

https://en.wikipedia.org/wiki/AI_effect

> Researcher Rodney Brooks complains: "Every time we figure out a piece of it, it stops being magical; we say, 'Oh, that's just a computation.'"

Thanks I didn’t realize it had a name.

I don't know. Back in the 80s the goal posts for AI was always a machine that understands. We've not even come close to building that.

Instead, for-profit companies seem to have let their marketing departments move the goal-posts to the 40-yard line. "We can't build real machine intelligence, so we'll call this AI."

For me, I always thought R2D2 or R. Daneel Olivaw were what AI was supposed to be. Not "playing Go is hard, so a computer program that can do that is AI." It's the AI label that has been misapplied.

I know, this is where I get the lecture about words meaning different things over time. It's why the technical term for the stuff we've been talking about is Machine Learning systems, not "AI." The "AI" phrase is marketing fluff.

> I don't know. Back in the 80s the goal posts for AI was always a machine that understands. We've not even come close to building that.

For any reasonable definition of understanding, we have. If you wanted to test my understanding of an article after reading it, what would you ask me? Now submit the same articles to ChatGPT and ask the same questions. I'm not sure who would score better.

No. We haven't. Any reasonable definition of understanding includes knowing what's real and what isn't. We've created software that does a great job of simulating responses. The software is unable to "know" anything, let alone that it's assertion that the current year is 2022 is false. It would be possible to coerce ChatGPT into generating the assertion that 2 + 2 = 5 and it would follow the simulation of the argument that this was so, generating text along the way that seems to brim with confidence.

The ability to generate text from statistical models of text is not understanding. That we can fool vision systems with decals is because those vision systems cannot bring a model of the world to bear on the interpretation of the statistical matches coming in through the detectors.

We've built Chinese rooms able to supply gibberish and excellent translation with equal aplomb.

Ask ChatGPT to reason out what 2 + 2 is, and it will deliver.

Ask a cocky undergraduate philosophy student to BS you about how 2 + 2 = 5, and they will confidently do the same.

What’s your point?

That philosophy undergrad will know that they're spouting BS.

- Hey check out this AI tiktok feature

- You mean advanced image processing?

- blank stare. uh... ok, whatever, it's really cool

- It is advanced image processing, not "AI"

- yea, so anyway, check this out

Reply on news.ycombinator.com