I'm curious if it's really possible to develop Artificial Superintelligence (ASI) or Artificial General Intelligence (AGI) without relying on simulations. For instance, while DeepMind made some strides in mathematics through calculated analysis, other scientific fields like biology and chemistry necessitate rigorous experimentation that can't simply be theorized. So, how can we expect a large language model (LLM) to test its ideas and learn effectively?
DeepMind had to compute thousands of calculations daily for even minor improvements, which makes me skeptical about achieving anything meaningful that requires real-world testing. It seems to me that we need to craft an incredibly accurate simulated environment where LLMs can safely experiment and iterate on their ideas much faster than humans can. However, simulating every detail down to the quantum level feels far beyond our current capabilities. Instead, I wonder if we could develop specialized simulations for specific fields over time that the LLMs could utilize to innovate and compare results across these simulations. What does everyone think? Could this approach work?
3 Answers
It's definitely possible, but it might take a lot longer than we hope. The complexity involved, especially with real-world scenarios that need testing, can't really be bypassed easily without simulations.
Exactly! The idea of needing a separate external interface to test real-world outcomes seems essential for any language model to truly ground its innovations.
We need to remember that intelligence has its limits. It often takes years to devise and prove new theories, sometimes requiring extensive simulations to get there. Take antennas, for example, engineers have to simulate complex physics to create new designs, without which progress would be stalled.
I believe we might hit some significant hurdles that AI will struggle to overcome, like complex combinatorial problems or chaotic systems that multiply errors. Even if we could simulate conditions perfectly, the sheer volume of variables involved could still make achieving desired outcomes a real challenge. No shortcuts here!
Yeah, and even the best AI would need time and effort to create real solutions. The complexity of everything—from proteins to chaotic systems—means we can't overlook the lengthy process of testing ideas through simulation.
I see your point, but looking at self-driving cars, there's still a long way to go despite tons of investment. If we need these kinds of simulations to even make progress with simpler tasks, how can we expect AGI to roll out without similar methods? Makes you wonder if we're in a simulation ourselves!