Last week, I wrote about false messengers in the IT industry. I think with the success of Artificial Intelligence, we are also facing the false messengers of Artificial intelligence. If some topics are popular, such as Football, many advocates arise! The same happens now with Artificial Intelligence.
In the past, many of them advertised how they leverage AI while they presented showcases from the '70s with some statistics spiced with Linear Regression and sold as AI.
Now, the same happens with Large Language Models (LLMs); while most companies are trying to build more sophisticated models, they still don't understand the basics of how neural networks and even the human brain work.
Like explaining financial problem solution is independent of language!
This is what we must accept, we must divide problems into smaller pieces. Yes, we can build a 100s of trillion parameters generic model - but as it is a generic model, it will create generic answers to particular problems.
We can divide all problems into fields, where for example, we can realize the problem belongs to a physics problem, but at this point, the language barrier stops. We can solve the problem in the same way in Spanish and English language. The problem-solving is language-independent! So training a model for a particular language such as Hungarian is a lame solution because I'm sure most of the physics-related documents are not in Hungarian language.
So in an ideal world the problem solving looks like the following sequence:
Understanding the problem and field (Language dependent - Transform the problem to the language where we solve problems with most references)
Solve the problem with a specific model (such as Physics with such language as English)
Translate the solution to the origin language.
This way, I don't understand why we focus on a trillion generic models; I'm sure billion parameter-specific domain models and basic language translation models can solve even bigger problems than any company can build in their home country language, such as UAE build a large Arabian model for generic solutions.
But we can listen to the false messengers who promise the opposite, and work on a large problem which can take two to three years until they can show anything!
They always buy time for themselves! Every day is a gift, and in the end we can still blame vendors or slow hardware for not having the result that we promised!
Be honest! How many times have you seen a multimodal model built from small perfections?
We still not working on perfection, but more on a mass mess that we can sell everyone!
I believe the future is open source, and built on these small perfections for the more about thousands of intelligence types that exist today in Humanity!
AGI will be built on these perfections and I hope we will find the point where we can collaborate instead of compete!