No, it doesn't take years and the testing is performed, roughly, as I just described it.
They build a test rig with 10,000 drives which they run continuously producing a total "drive hours" count and then wait for the first one or two or ten drives to fail.
X failures / 10,000,000 drive-hours will give you your MTBF, etc. ... and only takes ~40 days given a 10k drive test rig.
“They can't realistically do anything else to estimate it …”
Of course they can.
How do you think they get the MTBF for hard drives?
What they do is deploy thousands of them and wait for the first one to fail. Then they wait for the second and third ones to fail… And then they can construct a statistical abstraction for the entire population based on the very first degradations … and that doesn’t take long at all.
"GLM 5.2 is "almost Opus," and it needs at least 8xH200s for comfortable inference ..."
What is the behavior if one were to run GLM 5.2 with only a single H200 ?
Would it fail to run at all, or would it just run so slowly as to be unusable ?
I would like to prove out the build, and concept, of a SOTA model locally, but then backfill the rest of the GPUs in 18-24 months when they cost significantly less ...
See above where I frame this as a classical arms race and the equilibria (stalemate) between white/black hats is a very safe and stable place to be.
You and I don't get to choose who the white hats are and those actors could change dramatically as US administrations change, etc.
It's better to rapidly reestablish a new stalemate ... and I say that as someone who has spent the last six months rapidly triaging vulnerabilities and newly discovered attack surface in an attempt to safeguard the livelihoods of everyone who depends on my business.
"Mythos found 1000 zero days in a few weeks - if I had asked your thoughts on this a few years ago, I'm sure it would've been "that is a super-weapon"."
No, my take would have been that a very strong new move had been made in a long-standing, classical arms race.
Since I rather like stable equilibria in areas like this, my thoughts would have been that white hats need to catch up immediately to reestablish the stalemate ... which was probably the course we were on, prior to today.
"We need a Second Amendment for AI: the right to keep and bear strong AI shall not be infringed."
It is already robustly protected by the first amendment.
If I had legal possession of the fable source repository, etc., and printed it on paper, it would be legal for me to possess and distribute, discuss, annotate, etc.
"I know some people who are adamantly against solar and wind ..."
Ignore them because it doesn't matter.
Physics causes finance, which causes politics ... and their politics will immediately change when the finance crosses whatever threshold they happen to be anchored to.
That's different for different people and different situations but you can be sure it will happen. Those people will not pay markedly higher electrical bills or have a (relative) doubling of their cars TCO for their politics.
I am rarely in Atherton and have no connections to it.
However, I really appreciate both dense city centers and nicely organized single family home towns and neighborhoods.
So it makes perfect sense to me that one would advocate for the network effects of upzoning and building in dense, transit and service connected job centers ... and simultaneously be against breaking up an existing equilibrium in a smaller town.
Urban and sub-urban sprawl is an aesthetic and environmental disaster and it disappoints me that what was once a progressive imperative (working against sprawl) was jettisoned the moment it ran counter to economic self interest.
"... a stock is like if you went to some second hand/thrift store and bought a brand of clothing that was reviled for some reason or another, i.e., use of child sex slave labor, you giving a thrift store money to wear the second hand clothing not only does not benefit the reviled company ..."
This is incorrect and, frankly, ridiculous especially in light of your own scoffing at the "foolishness" of others.
The secondary market value for company stock has a direct impact on current and sustained operations in areas including, but not limited to:
- the ability to sell debt and the interest rates at which it can be sold
- the ability to attract and retain executive "talent" with stock compensation
- the ability to attract - or ward off - takeovers and buyouts from other firms
- the ability to expand operations, or development, through follow-up offerings
Your observation of this basic truth (that company shares purchased by at-large market participants don't yield funds directly to the firm) is, of course, correct.
However it is not as profound a factor as you think it is.
All people want is an electric Audi allroad. Instead, we get an e-tron.
All people want is an electric V90 wagon. Instead we get a polestar.
All people want is an electric Jeep Wrangler. Instead we get "Recon EV".
The reason for this is that the incumbent manufacturers understand clearly that the electric versions would completely eclipse the ICE models and their existing investments in design and tooling would rapidly diminish.
... and so, all of the eInitiative, iMobile, TronCars ... it's all a desperate (and lame) attempt to continue selling the ICE line and grow marketshare with the addition of the electric car consumers.
