This is what happens when you make all the h100/h200/equivalent cards exclusive and lock them up in warehouses. We have no way of running these models locally… yet. Keyword is yet, the exclusivity period is going to end just like it did for 3D graphics when 3DFX democratized it with the voodoo cards. They’re only 300gb of memory and a chip ahead. It’ll shrink.
This all is just spotlighting the weakness of NVIDIA, AMD, Apple, Microsoft, etc. They all avoided manufacturing in-house for so long and now they're fighting for fab time. Intel on the other hand is interesting...
Incorrect. The best Intel chip you can get is Panther Lake which is made on Intel 18A node, available globally at the end of this month. Intel has already used EUV machines in Intel 7 and Intel 3 nodes, for the last few years.
The problem is everyone is using a different “level” of AI model. Experiences by those who can only afford or choose not to pay for the advanced reasoning are far worse than those who can and do pay.
The majority of their cars (Y/3 models) have the penthouse (top) of battery pack super easily accessible from under the back seat, no need to drop a pack.
Not to mention Tesla has the best service mode system in their computer of any brand of all time. They also have the best free to owners assembly/disassembly manuals in the service portal https://service.tesla.com/. They have taken self-service literally to the next level compared to anything I've ever driven ICE, Hybrid or EV and I've owned all of them.
+1 for the Tesla service manuals. My wife’s was making a clunk from front suspension. Before my assistant (my kid) had finished taking off the wheel, I found the up-to-date official torque specs on service site. Usually it takes me a while to find torque values and cross check with another source. It was beyond refreshing to see Tesla buck the trend of selling service-manuals-as-a-service.
Service documentation / manufacturer software required for cars I currently wrench:
- Early 20’s: Bookmarked URL to the official online documentation (Tesla). With that said, I haven’t had need beyond checking mechanical connections, flushing brakes, and replacing filters.
- Early 10’s: VM containing a mid-00’s version of windows that runs a cracked copy of the long defunct manufacturer software service manual. Also runs software to interface with car, but simply painful to use. Beginning of era where tasks like replacing the 12v battery require manufacturer software to interface (though simple things still had undocumented secret Contra-like button sequences to do so).
- Early 10’s car: folders of screenshots and pdf exports collected over a decade for various procedures I needed to do. OBD-2 dongle + generic app handled basic things. Not much different than decade prior vehicle.
- Early 00’s: PDF of a seemingly printed-and-scanned copy of a digital version of the service manual. Off by a model year, surprising number of inconsistencies given its German. Computer and K+DCAN connection required for re-coding new parts, flashing, etc. Some fancier OBD-2 scanners could do majority of service related functions (cycle abs, reset airbag light, etc).
- Late 80’s: PDF scans of the dozen+ service books (still trying to luck into a physical copy of the set without paying an absurd sum). Most mechanically complex vehicle I own. No computer necessary, but soldering required.
- First, the submission we've preferenced was actually submitted over 24 hours ago. If you look at the item IDs, you'll see that one is lower. And if you look at the list of submissions from theverge.com [1], you'll see its submitted time is currently shown as "1 day ago". When we re-up an older submission, as we've done here, we give it some "grace time", so that its rank isn't pulled down so much by the age factor in the ranking calculator.
- The guidelines ask us to "Please submit the original source. If a post reports on something found on another site, submit the latter." The article linked from this submission is re-reporting content from that article in The Verge, and it was The Verge who did the interview and broke the story. We consider it important to reward the publication that does the original news-gathering and reporting by directing the traffic to them, and also to reward the submitter who submits the original/best source of the story, even if they submit it later (though in this case they submitted it first anyway).
You need to think hard and understand that it is irreversible before you publish your content under certain licenses.
My problem with this type of gate keeping is that machine learning does open up translations that are accurate to the masses. It is quaint having a real human do your translations though. Kind of like having a real human drive your car or do your housework. Not everyone can afford that luxury. But, on the other hand, having a singular organization own the training data and the model and not publishing the model itself is where the gatekeeping continues.
I haven't had a rolling blackout in our particular grid in several years. The battery energy storage has been a great benefit. Our home battery energy storage system has been fantastic as well for localized unplanned outages.
Very early days on this it seems. The released spec so far only being 64 pages, must just be a high level summary of their goals so far? Maybe we see it in use in 8-10 years? https://www.gccorg.com/article/69/426.html
Once you've had an outage on AWS, Cloudflare, Google Cloud, Akismet. What are you going to do? Host in house? None of them seem to be immune from some outage at some point. Get your refund and carry on. It's less work for the same outcome.
Yeah, just double++ the cost to have a clone of all your systems. Worth it if you need to guarantee uptime. Although, it also doubles your exposure to potential data breaches as well.
++double: spoken as "triple" -> team says that double++ was a joke, we can obviously only double the cost -> embarrassingly you quickly agree -> team laughs -> team approves doubling -> you double the cost -> team goes out for beers -> everyone is happy
double++: spoken as "double" -> team quickly agrees and signs off -> you consequently triple the cost per c precedence rules -> manager goes ballistic -> you blithely recount the history of c precedence in a long monotone style -> job returns EINVAL -> beers = 0
Shouldn't be double in the long term. Think of the second cloud as a cold standby. Depends on the system. Periodic replication of data layer (object storage/database) and CICD configured to be able to build services and VMs on multiple clouds. Have automatic tests weekly/monthly that represent end-to-end functionality, have scaled tests semi-annually.
This is all very, very hand-wavey. And if one says "golly gee, all our config is too cloud specific to do multi-cloud" then you've figured out why cloud blows and that there is no inherent reason not to have API standards for certain mature cloud services like serverless functions, VMs and networks.
Edit to add- ink ow how grossly simplified this is, and that most places have massively complex systems.
And data egress fees just to get the clone set up, right? This doesn’t seem feasible as a macrostrategy. Maybe for a small number of critical services.
If you use something like cockroachdb you can have a multi-master cluster and use regional-by-row tables to locate data close to users. It'll fail over fine to other regions if needed.
Why not host in house? If you have an application with stable resource needs, it can often be the cheaper and more stable option. At a certain scale, you can buy the servers, hire a sysadmin, and still spend less money than relying on AWS.
If you have an app that experiences 1000x demand spikes at unpredictable times then sure, go with the cloud. But there are a lot of companies that would be better off if they seriously considered their options before choosing the cloud for everything.
I totally agree with you. Where I work, we self-host almost everything. Exceptions are we use a CDN for one area where we want lower latency, and we use BigQuery when we need to parse a few billion datapoints into something usable.
It's amazing how few problems we have. Honestly, I don't think we have to worry about configuration issues as often as people who rely on the cloud.
Not GP, but my company also self-hosts. We rent rackspace in a colo. We used to keep my team's research server in the back closet before we went full-remote.
reply