Will AI be designing nuclear reactors of the future? I put ChatGPT to the ultimate test and the results are… surprising.Ask Me About Thorium Mug - https://at...
TL;DR: It suggests several methods and makes a few mistakes which he had to point out to which it suggests even more absurd solutions to.
The AI recommends doing things in long and hard ways and does not conceive of new or novel technologies; it just mashes together existing ones despite their implementation being difficult or impossible by simply waving away these issues by saying things like “Much research and development would be needed but…”
More the latter. Neural networks have been used in biomed for about a decade now fairly successfully. Look into their use of genetic algorithms, where we are effectively using the power of evolution to discover new therapies, in many cases even new uses for existing (approved) drugs.
But ChatGPT has no way to test or improve any “designs”, it simply uses existing indexed data to infer what you want to hear as best it can. The goal is to sound smart, not be smart.
That’s actually a decently good analogy, though a random redditor is still smarter than ChatGPT because they can actually analyze google results, not just match situations and put them together.
It’s a system that has some notion of context and a huge database of information and is really good at guessing what words to put on screen based on the provided input.
It can’t think of anything new or novel, but can generate “new” output based on multiple sources of data.
As such, it will never be able to design a fusion reactor, unless it’s been trained on input from someone who actually did.
TL;DR: It suggests several methods and makes a few mistakes which he had to point out to which it suggests even more absurd solutions to.
The AI recommends doing things in long and hard ways and does not conceive of new or novel technologies; it just mashes together existing ones despite their implementation being difficult or impossible by simply waving away these issues by saying things like “Much research and development would be needed but…”
so similar to say, a redditor trying to sound smart by googling and debating another while both has no qualification on that topic, got it.
I wonder how much of that is just an inherent part of how neural networks behave, or if LLMs only do it because they learned it from humans.
More the latter. Neural networks have been used in biomed for about a decade now fairly successfully. Look into their use of genetic algorithms, where we are effectively using the power of evolution to discover new therapies, in many cases even new uses for existing (approved) drugs.
But ChatGPT has no way to test or improve any “designs”, it simply uses existing indexed data to infer what you want to hear as best it can. The goal is to sound smart, not be smart.
That’s actually a decently good analogy, though a random redditor is still smarter than ChatGPT because they can actually analyze google results, not just match situations and put them together.
ChatGPT is, despite popular consensus, not an AI.
It’s a system that has some notion of context and a huge database of information and is really good at guessing what words to put on screen based on the provided input.
It can’t think of anything new or novel, but can generate “new” output based on multiple sources of data.
As such, it will never be able to design a fusion reactor, unless it’s been trained on input from someone who actually did.
And even then it’s likely to screw it up.