Yes, AI is coming for solution formulation, absolutely, but not all of it, because it is actually a statistical machine with context limit.
Until the day LLMs are not statistical machine with a context limit, this will hold. Someone need to make something that has intent and purpose, and evidently now not by adding another 10T to the LLM parameter count.
> because it is actually a statistical machine with context limit.
So are humans.
Machines have surpassed humans by magnitudes in many capabilities already (how many billion multiplications can you do per second?)
And I argue that current LLMs have surpassed many of my capabilities already.
For example GPT/Opus can understand and document some ancient legacy project I never saw before in minutes. I would take a week+ to do the same and my report would probably have more mistakes and oversights than the one generated by the LLM.
AI advocates are _way_ too confident about the nature human cognition. Questions that have been debated by philosophers and cognitive scientists for decades are now "obvious" according to you people, though you never provide any argument to support your statements.
>Machines have surpassed humans by magnitudes in many capabilities already (how many billion multiplications can you do per second?)
Do you have any idea how many calculations it takes for a human to put a ball through a hoop while running across a court?
It could be millions or billions in a second. Manifesting consciousness, coordinating body movements, and everything else all at the same time takes calculations.
You may not be aware that your brain is doing multiplication, or any other kinds of math, constantly, but it is.
I agree we do some marvelous things in sports but if we extrapolate from this table tennis robot, it's clear machines can/will do just as well there too:
That table tennis robot is not conscious. That table tennis robot does one thing well. A human is capable of far more. There is far more going on for a human playing table tennis than a robot. It doesn't matter if the table tennis robot plays table tennis better, it can't also play hocket, soccer, football, basketball, chess, polo, baseball, or many other things one human can do.
The human condition is nothing but a massive amount of calculations under the hood. You don't feel it, or understand it, but it's there. Everything in nature is math, every physical phenomena has a cause and effect rooted in mathematics, and it's no surprise that humans are great at subconsciously calculating myriad things on-the-fly, as life is happening around us.
We are not pre-trained using the summary of all human knowledge over all of history. Yet we make certain decisions with much more ease.
We are much more limited, but we fundamentally work differently. Hence adding more parameter like certain companies are doing isn't necessarily going to help. We need to rethink how LLM work, or how it work in tandem with something that's completely different.
I think it's doable, I just don't believe it's LLM, and I don't think anyone now knows what it is.
That is not what the education system does. That's an obvious distortion of reality. People train over billions of documents to statistically predict the next word to gain and understanding of language. LLMs do this statistical processing in order to mimic humans natural language learning ability. And there has been continued evidence of the limitations of this approach to accurately mimic the totality of human cognition.
Yours is a “God of the gaps” argument. You will remain technically correct (the best kind of correct!) long after the statistical machine has subsumed your practical argument, context limit and all.
I fall into the "pessimistic heavy user" camp, I burn thousands of $ worth of SOTA tokens monthly but it just makes me more acutely aware of the limitation and amount of work I need to do to work around them and what kind of decision that I should reserve to myself instead of trusting the LLMs to do.
I can give you the exact mathematical formula used to statistically optimize the output of a neural network from input examples. Can you do the same for the brain?
That's where we fundamentally disagree about.
Yes, AI is coming for solution formulation, absolutely, but not all of it, because it is actually a statistical machine with context limit.
Until the day LLMs are not statistical machine with a context limit, this will hold. Someone need to make something that has intent and purpose, and evidently now not by adding another 10T to the LLM parameter count.