AI Is A Car That Everyone Expects To Be A Spaceship

There have been multiple waves of AI hype in the past. This one feels different. AI has amazed us before in tasks with very strict boundaries such as playing games like chess. It has never before replicated tasks that we thought were the purview of humans.

That's changed with generative AI. Chatbots have existed for a while, but have never come this close to sounding like real humans. AI art has never been this good. Parsing unstructured questions have never been this good. AI written articles have never been this good. True self driving cars inch ever closer to reality.

Never before have so many people considered whether AI may replace their job. Many are even saying that anyone who doesn't understand how to work with AI won't have a job.

I think the improvements AI are great. I think they're going to be extremely useful tools and I look forward to seeing what they can do.

At the same time, I think the hype over AI is still extremely overblown.

Let me make some admissions: I am not an AI expert. I don't have a Phd in computer science. The closest I've come to working with AI is putting algorithms other people made into production environments. There are people with more impressive qualifications than me who would disagree with everything I'm about to say.

I'm also not a rocket scientist or an automative engineer. Yet, if an automotive engineer told me they can make a car so fast it'll turn into a spaceship, I'm going to be very very skeptical.

An AI generated image.

Let's start with the basics: when everyone talks about AI they often talk about machine learning. You give an algorithm a large amount of data and have it "trained" to do certain things with it. To get an algorithm to recognize dogs, you give it millions of pictures of dogs followed by millions of pictures without dogs. Each of those images need to be tagged so that the algorithm understands the pattern.

The important thing to note is that the algorithm is not really looking at the picture as a whole and trying to understand what is in it. It is looking at every single pixel and matching the pattern. It results in some interesting differences with human interpretation of pictures. Example: in the dog image, you can change the color values of the pixels just so slightly that the human eye won't notice a difference, but an algorithm may now think that dog is a school bus or a house.

Being able to parse large amounts of data is not the equivalent of actually understanding the content. IBM made a huge mistake here with Watson back in 2011. At the time, Watson had won a game of Jeopardy against top players. This was a pretty impressive feat as previously, AI could primarily win games such as chess which have strict boundaries and rules. Jeopardy is a game where anything in the English language could be said.

Yet, Jeopardy is still a game with strict rules even if the boundaries are less so. There's an answer and a question must be formed. With enough training data using English texts, an algorithm could find patterns to provide the correct responses. The algorithm does not actually understand the content.

Hot on the heels of Watson's win, IBM started thinking "what business could we use this in?". They (or at least the executives) failed to understand that they had built an amazing natural language processor and not human intelligence. They simply looked at industries with large dollar amounts attached to them and decided to pick healthcare. If Watson can "read" millions of books, surely it can read every medical text ever created and gain an understanding of healthcare that humans can only dream of.

Has Watson replaced your doctor yet? Was it even involved the last time you needed medical care?

IBM struggled to make Watson a major part of healthcare for a decade before selling the division. It's not that Watson wasn't a great tool, but there was a failure to understand its capabilities. Parsing text is not the same as understanding text. It could match patterns with words that appeared in those medical texts, but it did not understand medicine.

That's machine learning. It is capable of doing some incredible things far beyond the scope of humans, but it is far from intelligent and can't do many things that humans consider basic.

That leads us to the second type of AI: Artificial General Intelligence (AGI). This is human level intelligence. This is the type of AI that could be what you see in movies, from Skynet to C-3PO. This is the type of AI that could do any job a human could do, but better since it'd have increased processing power. This is the type of AI many companies *talk* about moving towards. This is the type of AI that would justify all the hype and all the fear.

Saying we can turn the AI we have today, which is built on the fundamentals of machine learning, into AGI runs on a pretty big assumption: that intelligence can be boiled down to patterns.

The problem with that assumption is that we don't even understand how the human brain really works. We're still trying to understand how intelligence is created in nature. We're trying to simulate the output without fully understanding the fundamentals that went into it.

Every AI project today relies on a huge amount of training data. There is work on reducing the amount of training data needed, but they all will ultimately need some amount of training data.

Humans on the other hand can learn from a single data point. A child only needs to see one dog to start recognizing other animals as dogs or could be dogs. There will be some error. Maybe they'll think a horse is a big dog. They won't however think a snake is a dog. They don't need to see a million pictures of dogs, or even a thousand pictures, to start recognizing dogs.

What about generative AI you may ask? It's not just recognizing things, but creating them. It's writing software code. It's writing blog posts. It's making images.

Generative AI does do all those things. At your request. It does what you tell it to do. It doesn't have dreams or feelings. It doesn't create goals for itself. It can recognize the pattern of what you asked. It can use the data it has of similar things *other humans* have created, mix and match the data, and provide you the result. That by itself is pretty amazing. It will be able to do things we have humans do in a fraction of the time. I look forward to having it replace many of my more menial coding tasks. Yet, it can only do things similar to things that humans have done before and have been included in the training data.

AI will also create errors no human would ever make. When Google's Gemini created an image of a black pope, it did not understand what it was doing. It may have parsed history books, but it did not understand history. People talk about fixing these "hallucinations". I think AI will improve and the instances of hallucinations will decrease, but they will never be eliminated. Only truly understanding the content in the training data could AI do that.

I titled this post saying AI is a car that everyone expects to be a spaceship because I think it's an easy way to understand the root problem here with everyone's assumption that AGI is around the corner. We're basically iterating on making a car faster and faster. Yet, no matter how fast that car is, it won't turn into a viable spaceship. There's a lot of fundamentals missing in the construction of a car that would prevent it from going into space.

Given our limited knowledge of how the human brain even works or what makes intelligence, I'd argue that we are missing some serious fundamentals as well when saying we can iterate on today's versions of AI and turn it into AGI. I could be wrong. Maybe intelligence can be boiled down to matching patterns. Somehow I doubt that.