Attention is valuable these days, so making people go to their websites for people to check if something is real is good for them, its people they can try to sell more cameras (or phones) and all that.
My sister has a Windows 10 laptop she used for her accounting business. One day it decided not to boot, saying there was no boot device. I took the laptop home, took the SSD out (Samsung 1TB), put it in an external USB case, plugged it into another Windows laptop, and it showed up in Explorer. Weird.
I had another brand-new, identical Samsung SSD, so I hooked both the old and new drive up to a Linux laptop (with USB cases) and tried to dd the old drive to the new drive. That mostly worked, but VERY VERY slowly: it would run fast for 5 seconds and then have no activity for 30 seconds. I had a fan blowing on the old drive to keep it cool because it was running very hot.
The dd copy would eventually fail and then I'd restart it with appropriate iseek and oseek values. I also did a cmp /dev/zero with the new disk to verify that it was all zeroes (it was brand new), and that allowed me to use conv=sparse on the dd. The reason for that was to avoid writing to ever sector of the new disk; I didn't want to copy sectors from the old drive that had never been accessed (she only used about 250GB of the 1TB).
It took a couple of days and about 5 restarts to finish the copy, but it did work, and as a precaution, I made another copy of the drive and ran a cmp of the original drive and the 2nd copy (also having to restart cmp several times). Since that compare worked, I knew that all 3 drives had identical content. The new drive worked fine in her laptop and she was mighty glad to see her Windows login screen.
The thing that made this work, IMO, is that Linux has a longer timeout for errors than Windows apparently does, especially during the boot sequence. Plus Linux allows adjusting the drive timeout, so if the device is doing error recover, which is sometimes slow, it gives it time to finish rather than reporting an error.
One of my theories was that the bad SSD was overheating, but if that was the case, a cold boot should have worked, with the failures only coming later.
The other theory is that one of the chips on the SSD failed, so the drive was having to use the ECC codes to correct for the missing information, and the correction process was taking longer than Windows boot would tolerate.
Next time you have a disk where you need to do repeated dd runs over different ranges, or suspect that you might need to, use ddrescue. It tracks which sectors have been recovered (and has lots of useful options).
You can also get 'partclone' to generate a list (in ddrescue format) of sectors containing data, so you don't need to try to read unused areas of the disk. For the partclone trick to work, the FS does need to be at least somewhat readable.
Checkout the insane amount of money Goodwill makes because of people getting rid of their "junk". There are 151 independent Goodwill organizations and all of them have a CEO, usually making 6 figures a year.
"Goodwill Industries was established in 1902 and is widely known across the country as the place where we all donate clothing and household goods to help others."
That's the first sentence from your link. Clearly people don't treat this org, literally called "good will", the same as they treat freakin eBay.
I don't think this is true, because people are often willing to spend a bit of extra time to do something good, like make a donation, but wouldn't be willing to take that same time to make $10.
I don't know much about the AI field, but it seems to me that trying to train any model to be all things to all people is a really dumb idea. It requires huge financial resources and is causing extreme shortages/market distortions in
any resource used by an AI company - RAM, SSDs, data centers, etc.
In the real world, you don't hire a plumber and expect him to also do your landscaping, fix your car, and tailor your clothes. It would seem like a much better use of resources if I could download an app that specialized in shell, Python, and C coding for example, or maybe even that would be 3 apps that communicated. Maybe I could even run them on a regular machine with 16GB of RAM. I don't need one huge model that can do that and code in Fortran, COBOL, and Lisp.
As humans, we've done pretty well by specializing. I hope this gets explored more with smaller, focused AI models vs the current path of one model to rule them all that can only be run in a data center the size of a country.
> I could download an app that specialized in shell, Python, and C coding for example, or maybe even that would be 3 apps that communicated. Maybe I could even run them on a regular machine with 16GB of RAM. I don't need one huge model that can do that and code in Fortran, COBOL, and Lisp.
I would daresay for "coding tasks", you actually _want_ a model that can code "in all languages".
Sure, it might be that outdated language XYZ is really useless to you or the task you want, but being exposed to their limits, philosophy and concerns across environment, framework and organization, among other things, means for example you get insights of your problems from other areas and points of view.
That's afterall how we got Newtonian physics and calculus, right? A person studying physics someday noticed how the "math of the day" wasn't able to calculate some results without a lot of elbow grease. He then "found" the "missing math" and with it was able to generalize what at the time was considered a bunch of isolated phenomena into a cohesive corpus of knowledge.
So for example, I want my code to have mechanical sympathy like Fortran; well defined input/output interfaces, and not-interweaved control structures, like COBOL; stateless, side-effects-free business logic like Lisp.
Similarly, it's better (for me) to use the right ring finger for Backspace rather than the pinky. The pinky requires moving the whole arm whereas the ring finger just needs a wrist flick to reach Backspace.
As a musician (piano), I've never been able to listen to music while working: it's too distracting, even without lyrics. That makes sense to me because for musicians, hearing just the music still makes your brain want to focus on the structure of it, time signature, rhythm patterns, interesting chords, key changes, etc. - things that a non-musician isn't so intellectually aware of, even if they like them in the music.
Interesting. Guitarist (as a hobby) here. I find familiar, favorite music good for concentration, with or without lyrics. But new music is what gets me. I give it too much attention, similar to what you describe.
I mostly listen to metal and all its various subgenres.
>That makes sense to me because for musicians, hearing just the music still makes your brain want to focus on the structure of it, time signature, rhythm patterns, interesting chords, key changes, etc
There's special music for focus that tries to keep all those to a minimum. 4/4 beats, no fancy rhythms, no changes, basic chords and repeating melodies etc. After a while, even if you're Mozart, you can ignore it just fine, just get the vibe and the driving pulse.
> So, back way before ChatGPT era, the folks over at AI safety/X-risk think sphere worked out a pretty compelling argument that two AGIs never need to fight, because they are transparent to each other (can read each other's goal functions off the source code), so they can perfectly predict each other's behavior in what-if scenarios, which means they can't lie to each other. This means each can independently arrive at the same mathematically optimal solution to a conflict, which AFAIR most likely involves just merging into a single AI with a blended goal set, representing each of the competing AIs original values in proportion to their relative strength. Both AIs, the argument goes, can work this out with math, so they'll arrive straight at the peace treaty without exchanging a single shot. In such case, your plan just doesn't work.
That price at Vultr gets you 1GB of RAM, and 25GB of relatively slow SSD.
The KV cache of your Claude context is:
- Potentially much larger than 25GB. (The KV cache sizes you see people quoting for local models are for smaller models.)
- While it's being used, it's all in RAM.
- Actually it's held in special high-performance GPU RAM, precision-bonded directly to the silicon of ludicrously expensive, state of the art GPUs.
- The KV state memory has to be many thousands of times faster than your 25GB state.
- It's much more expensive per GB than the CPU memory used by a VM. And that in turn is much more expensive than the SSD storage of your 25GB.
- Because Claude is used by far more people (and their agents) than rent VMs, far more people are competing to use that expensive memory at the same time
There is a lot going on to move KV cache state between GPU memory and dedicated, cheaper storage, on demand as different users need different state. But the KV cache data is so large, and used in its entirety when the context is active, that moving it around is expensive too.
Now check out the cost difference in 25GB of computer RAM vs GPU RAM.
And yes, this is also why computer RAM has jumped the shark in costs.
The bandwidth differences in total data transferred per hour aren't even in the same 5 orders of magnitude between your server and the workloads LLMs are doing. And this is why the compute and power markets are totally screwed.
I never got a surprise bill myself, but reading a few cases like this motivated me to cancel my GCS account and remove my CC. Now if I try to use it it fails immediately with an error.
As author of HashBackup, I know people are using it with GCS, and I'd like to be able to test against it, but not enough to swallow a large surprise Google bill.
