This is potentially one of the biggest value propositions of AI: The ability to iterate very fast on something is arguably it’s main value propositions, and iterability is probably going to be important for AI.
This is also why uploads would be huge for the economy, due to their ability to copy and iterate at a vastly higher level on vastly larger scales.
This also implies existential risk or catastrophic risk , depending on how amenable it is to iterability, would be a significant issue. Indeed, I think a lot of debates on pessimism vs optimism should focus on how much particular risks are iterable.
Yes. I sometimes reflect on how semiconductors, engines—many real world products are trapped in local minima.
For example semiconductors are forced to explore small tweaks to the current sota process and no other. This is because asml only supplies equipment to handle silicon, of a particular wafer size, and a whole suite of processes you are allowed to perform.
Diamond circuits? Not happening. Even silicon carbide took a long time to transition to for FETs.
Engines the same idea—I am sure you have seen monopistons and microturbines and various rotary ideas. But again, every idea is not as good right now as just tweaking a production scale design made of basically the same system of valves and cams and pistons as the model T, with only many small tweaks, is cheaper.
AI could be capable of developing these “significantly different” ideas from scratch, iterating through prototypes with automated research all the way from idea to a form that is actually polished enough to be competitive. High reliability high efficiency rotary engines or diamond wafer semiconductors that are 3 nanometer and run at 5ghz.
And this wouldn’t theoretically take the 30+ years it took us, since each AI model working on it has reviewed the results of all experiments, and can make decisions—thinking fast AND deeply—for the next step in under a minute.
Also I expect plenty of failure, maybe diamond doesn’t work at scale, maybe all the rotary engine ideas are bad, but what’s the harm—you just wasted an AI and some robots time. Plenty more where that came from.
This of course also applies to AI itself. No reason to be stuck with transformers and micro tweaks.
This is potentially one of the biggest value propositions of AI: The ability to iterate very fast on something is arguably it’s main value propositions, and iterability is probably going to be important for AI.
This is also why uploads would be huge for the economy, due to their ability to copy and iterate at a vastly higher level on vastly larger scales.
This also implies existential risk or catastrophic risk , depending on how amenable it is to iterability, would be a significant issue. Indeed, I think a lot of debates on pessimism vs optimism should focus on how much particular risks are iterable.
Yes. I sometimes reflect on how semiconductors, engines—many real world products are trapped in local minima.
For example semiconductors are forced to explore small tweaks to the current sota process and no other. This is because asml only supplies equipment to handle silicon, of a particular wafer size, and a whole suite of processes you are allowed to perform.
Diamond circuits? Not happening. Even silicon carbide took a long time to transition to for FETs.
Engines the same idea—I am sure you have seen monopistons and microturbines and various rotary ideas. But again, every idea is not as good right now as just tweaking a production scale design made of basically the same system of valves and cams and pistons as the model T, with only many small tweaks, is cheaper.
AI could be capable of developing these “significantly different” ideas from scratch, iterating through prototypes with automated research all the way from idea to a form that is actually polished enough to be competitive. High reliability high efficiency rotary engines or diamond wafer semiconductors that are 3 nanometer and run at 5ghz.
And this wouldn’t theoretically take the 30+ years it took us, since each AI model working on it has reviewed the results of all experiments, and can make decisions—thinking fast AND deeply—for the next step in under a minute.
Also I expect plenty of failure, maybe diamond doesn’t work at scale, maybe all the rotary engine ideas are bad, but what’s the harm—you just wasted an AI and some robots time. Plenty more where that came from.
This of course also applies to AI itself. No reason to be stuck with transformers and micro tweaks.