‘The discourse is unhinged’: how the media gets AI alarmingly wrong | AI

AI is incredibly over-hyped, the media is the main culprit for this, and this is very dangerous for the field – it has, in the past, caused several AI winters significantly hampering the progress for long periods of time.

Modern AI is very much limited. Even though there are many instances of things it can do, as you’ve listed, the actual holy grail of AI – higher level, common sense reasoning, is still as distant as it was 30, 40 years ago. And unfortunately all the focus seems to be on the statistical, pattern-recognition, big-data types of approaches. These work very well for some problems, but not at all for others.

The media’s fault in this is that they have no clue about what is in modern deep learning’s reach (can be done within a year if a team worked on it) and what is completely impossible.

Also, among the things you’ve listed:

Predict the weather

This is not an AI thing, the weather is a chaotic system that you can predict up to a certain degree if you have enough data and processing power. Intelligence does not help. You might use some approaches to try to somehow compress the data you have, find some patterns and then work with these patterns to simplify the dataset. But the predictive power is in computational power.

Teach itself

Nothing new really happened here, we’ve had “self-teaching” algorithms for many decades now.

Walk

You’re mostly right on this one, some algorithms have come pretty close to preforming like a motor unit controlling the actuators like our legs.

Talk

Here you should have used “text to speech”, because it’s not the same as “talking”.

See

Human vision is still vastly superior in most vision-related tasks. Again, media often over-hype the challenges of computer vision as being representative of the whole vision problem itself. But no – even though a net can get a lower error rate on a dataset than a human does, it doesn’t mean the problem is solved. Humans still “see” vastly better and these algorithms will need significant advances to come close (it’s not just a matter of scaling up the net and giving it more data).

Write

Again – ELIZA could write in the same way 50 years ago. If I asked any AI “What happens if you turn a bucket full of water upside down?”, it would not know how to “write” back an answer.

The others I have no comment on, I mostly agree with them.