What you wrote is true when considered for low-risk and isolated simple problems, but it will be insufficient for more complex real-world dynamics.
Most of the points you provided are about there being no difference between an almost perfect result and a result that is closer to perfect.
"Diminishing Returns" is a failed idea in scenarios that contain asymmetric risk. The margin of error being a "few notches" higher does not always mean the result we get is acceptable. It is not acceptable for a mission critical problem, such as human lives being in danger, to be "good enough"; compute power and reasoning must continue until the risk gets closest to zero.
Also, collecting new data in high-risk areas is a more costly, inconclusive, and dangerous process compared to reasoning. As long as the data at hand is not completely exhausted, collecting new data is a risky move.
In the development of artificial intelligence systems, it is necessary to push reasoning to the extreme limits, otherwise, with the "good enough" level, they cannot go beyond the current statistical parrot LLMs.
I'm sorry if I fell short, I just wanted to share my ideas.
This is thoughtful, thank you! What I took from it is: high-risk applications (for example where lives are at stake) demand the very best judgement - so even if it takes a huge amount of effort, diminishing returns can be worth it. Does that sound right?
Separately you're saying that, while diminishing returns to reasoning might suggest that getting more information is much more effective, sometimes that itself is hazardous or costly. Indeed!
I agree substantially with those points. This is one reason that I think AI systems are not yet ready for many critical and high risk applications.
I don't think this is evidence against my central expectation here, which is that returns to reasoning alone are often quite diminishing.
What you wrote is true when considered for low-risk and isolated simple problems, but it will be insufficient for more complex real-world dynamics.
Most of the points you provided are about there being no difference between an almost perfect result and a result that is closer to perfect.
"Diminishing Returns" is a failed idea in scenarios that contain asymmetric risk. The margin of error being a "few notches" higher does not always mean the result we get is acceptable. It is not acceptable for a mission critical problem, such as human lives being in danger, to be "good enough"; compute power and reasoning must continue until the risk gets closest to zero.
Also, collecting new data in high-risk areas is a more costly, inconclusive, and dangerous process compared to reasoning. As long as the data at hand is not completely exhausted, collecting new data is a risky move.
In the development of artificial intelligence systems, it is necessary to push reasoning to the extreme limits, otherwise, with the "good enough" level, they cannot go beyond the current statistical parrot LLMs.
I'm sorry if I fell short, I just wanted to share my ideas.
This is thoughtful, thank you! What I took from it is: high-risk applications (for example where lives are at stake) demand the very best judgement - so even if it takes a huge amount of effort, diminishing returns can be worth it. Does that sound right?
Separately you're saying that, while diminishing returns to reasoning might suggest that getting more information is much more effective, sometimes that itself is hazardous or costly. Indeed!
I agree substantially with those points. This is one reason that I think AI systems are not yet ready for many critical and high risk applications.
I don't think this is evidence against my central expectation here, which is that returns to reasoning alone are often quite diminishing.
I did my best to contribute somethings. I'm sorry for mistakes, thank you for your message I'm really learning a lot.