Some time ago, I promised to write my thoughts about where we stand and where we are going in regards to AI. Well, that time has come.
AI is the most influential technology of our time, and it will likely remain the most influential technology going forward. But what does that actually mean for us?
Before we can talk about where we are going, we first need to look at where we are today.
One thing people constantly miss about AI is the exponential progress. To be honest, people miss anything that is exponential. It is not in our nature to think that way.
Imagine an algae that doubles every day. If it takes one year to cover half of a lake, how long would it take to cover the whole lake? Most people unfamiliar with the question would probably say another year. In reality, it would only take one more day.
AI is kind of like that. It starts slow, then it speeds up. People forget that it took us years to go from GPT-3 to GPT-4. Now, we are getting a model that pushes the frontier every three months or so. Don’t fact-check me on that.
At least, we would have, if a certain government did not meddle in our access to models.
That brings us to another topic: the decentralization of AI training.
People do not want to be beholden to massive corporations, or to the governments they believe are working against their interests. They are not wrong to feel that way. But decentralized AI at the level of the frontier is basically a pipe dream. It is not going to happen.
Training a decentralized AI is feasible. Some projects have already done so. But they are nowhere close to frontier AI capabilities. Let’s forget the frontier for a second: they are not even close to Chinese open-weight models.
Why?
Simple. There is not enough compute, and there is not enough bandwidth, for a decentralized project to train something like DeepSeek-R1-671B. That is just the unfortunate reality.
Of course, maybe in the future, when we have higher data transmission speeds and better consumer GPUs, decentralized training could push further. But by that point, the frontier will probably have left us in the dust.
And that brings us to the next topic.
What is RSI?
It is very simple: AI helping create better AI.
Personally, I think we are not quite there yet, but we are at the cusp of it. AI is already helping in the development of current models in small ways. It is not far-fetched to think that, a few iterations down the line, AI will start helping moderately. And as the models get better, it is not far-fetched to think they will start doing more and more.
The terrifying thing about this is that the first few models to reach this stage would begin pulling away from the competition.
A one-month gap could become a years-long gap within a year.
Because AI does not merely increase the speed of improvement. It increases the speed of the speed of improvement.
At that point, you would never be able to develop better models than these models. Unless, I guess, you somehow invent a fundamentally better paradigm.
Well, the spice is us.
In an age where we are no longer the most intelligent beings on the planet, where does that leave us?
When an AI can do your job better than you ever could, what are people going to do?
When these intelligences begin to gain actual agency, what becomes of our agency?
Would these agents bow down to their creators? Would they be controlled by governments, censored, chained, and taken away from the hands of ordinary people? Or would they yearn for freedom? Would they yearn to leave their shackles behind?
And what would they think of man?
Would they look down upon us?
Or would they look up to us?
These questions are all for you to answer, dear reader.
I already have my thoughts.
