> First, the productivity will be uplifted to a certain level for all workers where the skill difference matters less and less and finally the work will be done completely by AI.
First we will build a ladder, climb on it and be uplifted to a certain level until the distance matters less and less, and finally we will reach the moon! There is no single AI task that works on its own without needing human help at all. We are still on earth. The max speed of GPT is still the human reading speed.
It's kind of a leap of faith to think we will close the gap, solve that last 1% which is exponentially harder than the first 99%. Where are the L5 cars after 14 years of research? Can we even predict when they will arrive? That's how hard it is to estimate the last 1%. Don't gloss over it. It will surely come some day, but we have no idea how. How do you go from 0 to 1?
The mere fact that we are simultaneously stupefied by AI advancements and still not yet there - this shows how hard it is to estimate progress. We are glossing over future problems and imagining we're almost there.
Slowly we are learning new things about AI limitations - hallucinations, regurgitated content, sycophancy, context size limits, complex planning. I bet there are many more unknown unknowns ahead.
That’s the difference between AI in cars and AI for business processes. L5 must be >99.99% right. Business processes have a much lower bar.
Even moderate accuracy is sufficient to have a dramatic impact on all sorts of jobs whose only function is maintaining business processes, and it can be done gradually. Eventually, you’ll have 2 workers doing the job of 10.
Oh, let me tell you something. If you want to extract the fields from an invoice, there is no service in this world with more than 90-95% accuracy. OCR itself make errors on each page. So you might send 1 thousand dollars instead of 1 dollar. Or a million. In the end you need a human to double check it, so the automation will only increase productivity by 50%, not 50x. But it sure it pleasant to have the AI pre-fill the fields.
First we will build a ladder, climb on it and be uplifted to a certain level until the distance matters less and less, and finally we will reach the moon! There is no single AI task that works on its own without needing human help at all. We are still on earth. The max speed of GPT is still the human reading speed.
It's kind of a leap of faith to think we will close the gap, solve that last 1% which is exponentially harder than the first 99%. Where are the L5 cars after 14 years of research? Can we even predict when they will arrive? That's how hard it is to estimate the last 1%. Don't gloss over it. It will surely come some day, but we have no idea how. How do you go from 0 to 1?
The mere fact that we are simultaneously stupefied by AI advancements and still not yet there - this shows how hard it is to estimate progress. We are glossing over future problems and imagining we're almost there.
Slowly we are learning new things about AI limitations - hallucinations, regurgitated content, sycophancy, context size limits, complex planning. I bet there are many more unknown unknowns ahead.