How would AGI even work to discover new technologies? Take for example something ridiculously complex like EUV, which requires the highly proprietary tacit knowledge of half a century's R&D. I struggle to see how AGI would even approach problems like these besides brute forcing and consuming many nuclear power plants worth of energy, only to produce garbage that doesn't work.

>>106127329 (OP)
The whole idea is that AGI would just replace the humans developing it. Probably just 2x the speed of research at most because you still need to test everything in real life.

>>106127329 (OP)
AGI won't be a thing.
Not with current AI's paradigm which are being hyped, at least.
And not with the shitty repurposed hardware we are currently using.
Source : Nvidia's salesmen.

And it took decade to get to that point. So don't expect another pardigm shift before at least several decade. Especially now all those megacorp are so deep into the sunken-cost pitfall.

>>106127524
>you still need to test everything in real life
Not exactly true though, with a sufficiently advanced intelligence you could use quantum computing levels of speed to fast track an incredible amount of work.
Many things are well enough understood that we can simulate them with near 100% accuracy already, we just lack the processing power to do it at the mind boggling speed we (or AGI) would need to truly go crazy.

>>106127329 (OP)
It only seems to work in closed deductive domains like programming. Material engineering is more capital intensive and would require some kind of fabrication suite that would be hard to build.

You would have to put it in a feedback loop with reality where it can conduct experiments autonomously. Still, AI could solve math and computer science without any feedback from the real world, which could lead to huge breakthroughs with applications in the real world.

it would need a rich world model; an internal representation about space and time instead of semantics that LLMs use.

>>106129024
>Many things are well enough understood that we can simulate them with near 100% accuracy already,
When we make such simulations, we still need to validate them very carefully to make sure that we haven't fucked up. It's *VERY* easy to get it wrong in a subtle way, like a slight error in the rounding mode on one code path.

>>106130023
>which could lead to huge breakthroughs with applications in the real world
or could just vanish up its own asshole with crazier and crazier axioms; math is quite good at that

>>106127329 (OP)
Fast feedback loop. Thats why you need an interface to the world to interact with. Hook up an AI to a telescope that can control the telescope, look at any space for it to observe, make hypothesis, probe, fix theory on bad results, etc. Similarly, hook up AI to a factory that it can control.

And same with driving (tesla/waymo), chat bot, drawing, etc

>>106130143
Human skill issue. GPT5 already doesn't make these mistakes, it's just a matter of scale and speed now.

>>106129420
>It only seems to work in closed deductive domains like programming.
Because it's still just a text generator. Real AGI won't be based on a fucking LLM

>>106129024
>codetranny nonsense
go talk to a chemist about all of the reactions that "should" work but don't.

>>106130263
depends of the simulation accuracy.
if you use mechanistic simulation it's not that accurate.

if you now simulate down to the quantum field, it works extremelly well, but we get choked on compute pretty quickly.

>>106127329 (OP)
Assuming you have AGIs that can apply themselves and learn whatever a super smart motivated human can, and are as creative and imaginative at problem-solving, posing problems to solve etc as those very best humans (this is deepminds/hassabis's definition of AGI) It's not hard to imagine how groups of them could make research progress in many areas. They could be embedded/employed in orgs, companies, universities, ask questions of their human colleagues etc.
Of course that's kind of a ridiculous, scifi assumption and probably will never happen

>>106129024
is this why no spaceship ever fails anymore or something
after 4chan stopped gatekeeping, it's /b/tier everywhere. what the fuck are you even talking about, go medically assist yourself

>>106127329 (OP)
ai can only regurgitate what has already been created by humans. it cannot create new ideas, therefore it will never replace humans. To create new thoughts and ideas you require a soul