C'mon... First, there's no infrastructure in the US to process the gas and purify to semi manufacturing levels. Second, as soon as the epstein war is done (either by Iran allowing navigation through Hormuz or the arab countries developing alternative routes), Helium orices will go back to be so low it doesn't make financial sense to purify unless you have the volumes they have. And by then, still no infrastructure would be ready for the task.
The oil reserve is so that when there's a war in the Middle East you have some period of time before you run out and encounter price shocks.
The helium reserve is so that when there's a war in the Middle East...
The US has drawn down on the first one substantially and sold the second one to the private sector a decade ago that had no interest in maintaining it.
I have to disagree here. While it's been proven that a lot of things can be done by just searching a very big solution space very fast (either by brute compute or by a sophisticated algo or both), hardware solution spaces are not that easily searched.
For a lot of applications where signal integrity is not critical, high frequency switching is not required and power density is moderate, yeah. Anyone can do that, even an AI. But you don't even need to make a pcb for those cases, just throw the components in a protoboard and fire away.
When you start adding variables to the search space (emi interference, matching routing impedances, heat management, mechanical integrity, etc) the search space grows so big that any non-expert will become lost in it.
Will future cheap compute power be able to overcome this? I'm not betting on it.
I'd say that balancing those aspects are exactly what an AI excels at. And an AI can iterate in a way a human wont. Completely redesign something because an edge-case could be slightly better.
AI isn't brute-forcing, or it would have been completely unusable for code.
We can probably find people saying the search space for Go is too large and that we'll never be able to do protein folding and if we can we'll never manage protein interactions and then well never manage to simulate a whole cell (tbc). I just think at some point these machines become smarter than us, if it's 2 years or 20 years the self improvement part will happen on both the hardware and software side. Judgement day is inevitable.
> When you start adding variables to the search space (emi interference, matching routing impedances, heat management, mechanical integrity, etc) the search space grows so big that any non-expert will become lost in it.
Have you tried Astra?
I'm designing an analog ASIC. It will suggest e.g. transistor matching. You can also give it hints for layout.
You still need to validate with SPICE simulation (and obviously silicon).
I'm also looking into non-LLM scripts that use brute force aided by heuristics (e.g. lookup tables) to efficiently search the solution space.
A lot of stuff one learn along the way never gets put formally on a document, and every iteration has to be checked in the lab.
With code or math or simple text, or games, you can have swaths of documents easy to digest and readily available, you can run thousands of tests per second and you can check results right a way.
With a power converter, a ddr4 module, an fpga, you have to build the pcb, assemble it, test stuff little by little, fix thing as they break, increase power, etc.
You can take notes on stuff that has been changed , the reason for the change is usually a few words note if any. And then you find something weird and you see that there is absolutely nothing on internet about it and you have to find out on your own. At least that's my experience with hardware design.
For the important stuff that makes you lose a board or a month of work you make sure to document the reason and the solution, but it usually is in a internal document no one will (regretably) read.
If you then add mechanical stuff, thermal management, failure modes, etc. I just don't think there's enough compute power to bruteforce such a huge space of solutions, even if you don't take into account that you can't really teach the machine about hardware design apart from some open source stuff and development kits where documentation, part lists and gerbers for the pcb are available.
Also did you not see the quoted post at the end, they are already training ai on running material science labs so we are quickly moving away from the non-expert.
You would beat anyone if you could calculate a few thousands moves per second, even more so if you had a neural network that discards suboptimal moves early.
Disclaimer for stockfish in particular: Even if it's mostly bruteforce nowadays, stockfish usually has an edge because it can bruteforce with "taste" and can manage the clock very well knowing when to ponder deeper a position.
Yes, but note the central claim of your original comment:
> ... search space grows so big that any non-expert will become lost in it.
And I am saying that's a fallacious argument: chess has an enormous search space too, and that does not imply AI can't beat humans.
Also
> You would beat anyone if you could calculate a few thousands moves per second, even more so if you had a neural network that discards suboptimal moves early.
I say that hardware design has a way bigger solution space.
In chess, being able to check hundreds of lines per second makes you a superhuman. But in hardware design you get lost if you try to search the solution space for a circuit, schematic, component selection, layout, routing, stiffener and thermal solution.
There are a few like positioning and routing where the solution space can just be considered infinite as there are infinite copper geometries (even if you stick with straight line polygons) that can connect a number of infinitely differently positioned components.
Then, for the mechanical stiffener, again you have an infinite number of geometries that don't clash with your components, don't run over your copper traces and make the board resist the mechanical tests (which the AI could check with a lengthy FEM analysis per iteration).
After all that, or in parallel with the mechanical analysis you can run the thermal ones and tweak positions of the hot stuff and thickness of stiffener walls to get everything fine.
I just can't see an LLM do that with a board with a few hundred components, even if most are decoupling capacitors.
All this emi stuff is easily caught in a few rules that you use over and over again. Exactly the sort of thing an LLM could easily do.
Also, Claude can already write a multiphysics simulator for you. Doing EM analyses on the fly during design is something that will make LLMs much better designers than humans.
You could do that if you knew enough about how it works in detail. In practice they're just gonna make as big a blob as they can muster and apply RL on it or something, and nobody will be any the wiser as to what the thing's feeling.
I'm sure they know more about brains than I do but I'm also sure they cannot derive from an abstract description of a computational process whether that process implements suffering or self-awareness or whatever, which is what I think we'd need to do this human-neurons-as-lego-bricks type research safely.
So many questions! Since we still don’t understand consciousness how can we draw a line to stay on the right side of? Is consciousness correlated with intelligence or independent of it? What if we keep growing these organoids until they are near-human-brain sized. We know that such things can host a consciousness yet not always eg. “brain dead”.
What if the only reason we don’t believe they are conscious is because they have no mouth to scream with?
Also, it occurs to me that a right-to-lifer would have to be against such biotech to remain consistent.
I find it all incredibly fascinating but ethically highly fraught.
reply