kurthr
Probably shouldn't be having those Kenyans (Kalenjin) competing in marathons either. Completely unfair advantage.
All models are wrong, some are useful.
My books went from 2 slit experiments, to particle in a box (1d and then 2d harmonic oscillators) with tunneling, and then the Hydrogen atom potential with electron orbitals as an extension, to perturbations for He, to Hartree-Fock and Born-Oppenheimer for large atoms and molecular models. Learning the <bra/ket> formalism can be useful for thinking about it, but it has to be motivated by the math.
I guess black body and scattering were pretty early too as motivation for the developments (as they historically were). Whereas EPR/Bell, Decoherence, and "Interpretations" were mostly discussed as interesting flavor and philosophical arguments.
To add a bit more almost off topic substance from the enshitification of human work. I really do think the idea of hooking these "AI" models up to humanoid robots is a bad idea not for all the mundane reasons, but just because "alignment" seems impossible. All of the world's stories about humanoid robots basically involve them attacking humans and behaving in nefarious destructive ways. It's fundamentally embedded in the models, and there's no reason post training would pull it out, because it's basically in the key value vector meaning "malicious human destroyer". Even to find counter examples like C3PO, Daneel Olivaw, Data, or Wall-ee they come after droid/robot driven wars and apocolypses. It seems unwise.
Notably, neutrons are quite unstable outside of a nucleus and decay generating a proton, electron, and a neutrino. There are many theories on the discrepancy.
https://journals.aps.org/prd/abstract/10.1103/PhysRevD.110.0...
I think the challenge here is that other than a CCTA there's no good way to estimate individual (as opposed to hereditary) ASCVD risk. Sure LDL, ApoB, Lp(a), and BP, are statistical proxies, but tell you little about the individual risk. CAC isn't much better, but if you've been taking EPA and VitK2 you have a much higher chance of more stable plaques, which are also more visible on CAC without much larger X-ray exposure and iodine contrast effects.
Already, the big capabilities are becoming multi-layer/die integration. Whether it's BSPDN, stacked die, optical interconnects, etc it won't be as simple as Denard scaling. It will be dependent on experience and supply chains, which are harder to transfer than equipment or IP.
I do think the market for inference will drive a lot of innovation, and a lot of that will be at larger nodes. A single reticle die masked ROM with CIM can do 7B at 6nm. At 2-3nm you can do 27-32B which is a single expert or router. At that point there's not much reason to shrink for AI since it's mostly 1-hot die per token (except for the router). Update the parameters every 3-6mo when a new frontier comes out, and take the cost/power savings to the bank.
If they had a solution to get to >90% yields on sub 5nm they'd be doing it now. They don't, even the 7-10nm isn't there. Best hope is they get to buy the entire world's supply of EUV in 2027, then they install all of them, then they bring them up, then they build processes around them and make masks, then they ramp the yield on 3-4nm by copying exact TSMC.
The hilarious thing is that TSMC couldn't do this at a new fab in Taiwan, with full access to equipment, supply chain, and people. So, nope.
You're not going to make money buying a new 14nm bitcoin miner, and that is where the China foundries are for cost effective silicon. They're great at 28nm+ and good at 22nm pricing.
Most new design modern sub 7nm ASICs are made in Taiwan/Korea by TSMC and SLSI. That is where miners are profitable.
And what we have in the US right now is not a free market, nor a well regulated and efficient one. You can call it crony capitalism, if you want, but when banks are bailed out to the tune of $Ts, Federal deficit spending is responsible for "positive" economic growth, and speculative companies are shoved into the indexes that make up what savings the working class has, it barely makes the cut.
Plenty of more "socialist" places with with free healthcare and rule of law, are more market capitalist. Enshitification happens where it is allowed to, because it is allowed to.
I'd argue that it's the anti-conspiracy theories, that are just as common and almost as insidious. For example, the idea that the intelligence agencies aren't aware of what their government representatives are doing in the background, and can't/don't influence it. There are literally thousands of individuals who work in the same town with the same people with, their very jobs are to discover and keep secrets, and they are almost all presumably capable. Means, motive, and opportunity. And yet, supposedly of the many thousands, none of them has ever done it without express permission from a judge. It would take a massive conspiracy of group think for it not to happen.
More seriously, you only need to convince 20-30% of the population with a well sub 100 IQ, who think about survival, tiddies, and beer (not in that order). It won't be a long con, probably mercifully short.
Get into the notation and tensor formalism of raising and lowering operators, etc. It really helped make some of the results more clear to me, just as <bra/ket> helped with QM.
Or are you doing high end astronomy or building a back yard Wave detector? That would get truly expensive!
Maybe somebody will use them, but not the original owners.
The other hyperscalers are even further delayed. CoreWeave has some VeraRubin, but not in volume (as I understand).
That said, getting a point solution and not knowing the path to get there, is certainly much less useful, and generalization is much harder. Again, more work for mathematicians, but maybe without the easy "I solved it" proof metric for rewards. Goodhart's law strikes again.
I saw this once in a pre-production cell phone where tact-switch IO inputs were being spuriously activated. Figuring out that part of the UI required pushing physical buttons, suddenly made sense of "why do your batteries only last an hour".
Embedded touch is a whole'nuther issue. One could certainly remotely measure the radiated emissions of a cell-phone to get finger proximity and position in at least one if not both XY dimensions.
That said, there are some implementations that stand the test of time, that are fit for purpose, and do not have to be changed, debugged, or updated. I'd prefer to copy/call one of these rather than write my own.
Of course, I'm a very mediocre programmer, so for certain I should not try to be original.
Sure, it depends on what you're trying to do. Fit an equation or understand it. Yes, you can use a reduced mass to simplify the equations and make it more accurate, which usually isn't necessary when the new center is at less than 1% of the radius of the Sun (note ellipses have 2 foci from Kepler/Newton). Even Jupiter's center of rotation is at the surface of the Sun, and both have eccentricities <1%.
Of course, if you're using epicycles, you're never going to get Einstein's elliptical precession correction to the orbit of Mercury. Simpler, is usually better for models, even when you need greater and greater accuracy, overfitting is the enemy of explanation (to coin a phrase).
All models are wrong, some are useful. - Box