AI is still pretty dumb

We're a long ways away from actual artificial intelligence

Having worked for an “AI” company, I’ve seen how the sausage gets made. It’s thousands of lines of python code preparing the data, then one line of actual ML that actually uses the GPU to train the model which might take hours or days to execute, then thousands of other lines of python that apply that model.

In the entire codebase, there are only 3-4 actual lines of ML. Everything else is data manipulation. The actual ML itself is applying an ML library coded by someone else.

In the AI/ML courses I’ve taken, this is exactly what they have taught. Much more work is necessary to prepare the data than actually apply ML and then much more work is necessary to apply the models that come out into meaningful results. Although a lot of horsepower is necessary for computing that one very lengthy and detailed line of Python.

Not much is taught about what actually happens when crunching those numbers but from my relatively simplistic understanding I can tell you it’s really nothing magical. It’s a lot of relatively simple algorithms applied in a massively parallel manner over and over again, usually against large training datasets. And mostly, it’s just linear regression.

That’s all AI today, which should really be called ML, is today. Sure, it’s impressive and has a lot of applications. But it’s very, very far away from anything resembling the power of the human brain.

Neural networks are something else entirely and frankly, I don’t know a lot about them so I cannot speak much to their applications or how they work other than they are really black boxes.

All the rage right now is ChatGPT and GPT-4. First, let me say I am very impressed. However, I am not overly impressed nor do I think it’s revolutionary, especially with ChatGPT. For the most part, it’s been shown to be able to quickly and easily regurgitate information it has collected from the Internet in an easily digestible form. Congratulations, you’ve created a more efficient search engine by combining Watson with more better NLP. That’s not a very impressive achievement.

The more impressive side is its creativity. I am impressed it can do things like write cover letters or come up with names for companies. However, that’s excused away simply by having a very good ML system (GPT-4) that has been trained well. It’s impressive, but not even what I would call AI (depending upon your definition of intelligence).

Is it useful? Absolutely. Is it valuable? 100%. GPT-4 will change the world. The corporate applications are immeasurable. It’ll 100% replace a lot of jobs. But so did the ATM. Was that a world changing invention? No. It just replaced low skilled jobs. Like ChatGPT and GPT-4 will.

I’m impressed with the progress that ChatGPT represents and a little worried about some of the moral questions it brings up. ChatGPT refuses to answer a lot of questions on moral grounds while answering similar questions with no issue, showing the clear bias of its trainers, both socially and politically. That is definitely something that should be discussed more. But all I see it as is an about right on time representation of where we are with ML, not a groundbreaking leap forward. It combines improvements with ML training, NLP, and data ingestion to create something fun to show off. Yay! But I’m not surprised by it.

Show me an AI smarter than me and then we’ll have a real conversation. AGI is all I really care about. Until then, everything is just incremental improvements on ML. I’m happy for OpenAI and am sure they will make billions off of GPT-4. But it’s still rather simple statistics and math.

When will we see real, groundbreaking advances in AGI? Something that will make my jaw drop? Something that will pass the Turing test and even more.

Until that day, don’t tell me AI is anything more than just a little fun with math and still pretty dumb. It has its applications and it’ll continue to improve but I’ll continue to yawn.