Dran

“Open source” in ML is a really bad description for what it is. “Free binary with a bit of metadata” would be more accurate. The code used to create deepseek is not open source, nor is the training datasets. 99% of “open source” models are this way. The only interesting part of the open sourcing is the architecture used to run the models, as it lends a lot of insight into the training process, and allows for derivatives via post-training

It’s a little deeper than that, a lot of advertising works on engagement -based heuristics. Today, most people would call it “AI” but it’s fundamentally just a reinforcement learning network that trains itself constantly on user interactions. It’s difficult-to-impossible to determine why input X is associated with output Y, but we can measure in aggregate how subtle changes propagate across engagement metrics.

It is absolutely truthful to say we don’t know how a modern reinforcement learning network got to the state it’s in today, because transactions on the network usually aren’t journaled, just periodically snapshot for A/B testing.

To be clear, that’s not an excuse for undesirable heuristic behavior. Somebody somewhere made the choice to do this, and they should be liable for the output of their code.

I don’t know of any off the top of my head, but with a cheap digital caliper and tinkercad, I assume you’d be able to model one fairly trivially. You could friction-fit two halves around the cable, and secure it with some simple adhesive, or some kind of simple bolt/nut fastener mount if you wanted to get clever.

Never not learn a new skill!

I’ve always wondered why board partners didn’t just raise to scalper prices and take a $2200 profit per card sold.

And tbh, it’s Nvidia’s fault that the partners don’t have enough dies, I’d much rather a partner take the margin than an unnecessary middleman.

Why?

There’s nothing preventing you from forking a Lemmy client or server to prototype this. Depending on how you implement the activitypub backend, you might be able to make it transparent to a user if you present an algorithm as an array of cross posts via a /c/ of a server.

Anything more might require forking a client, which might be easier to implement but may be harder to convince a large userbase to migrate to.

I use ansible on one of my side projects; I use puppet at work. It’s the same reason I use raw docker and not rancher+rke2… it’s not about learning the abstractions; it’s about learning the fundamentals. If I wanted a simple abstraction I’d have deployed truenas and Linuxsserver containers instead of Taco Bell programming everything myself.

Sure. I have an r630 that is configured as an NFS server and a docker host called vacuum. There is a script called install_vacuum.sh that with a single command, can build the server to my spec from a base install of Ubuntu 24.04. it has functions to install base packages from repositories, add new repositories, set up users, create config files for NFS, smb, fstab, crontab, etc… once an NFS server exists on my network, any other server could be my docker host. My docker host is set up from a script install_containers.sh. as with before, it does all the things to get me a basic docker host, firewalled, and configured for persistence via my NFS server. It also has functions to create and start docker containers for all of my workflows (Plex, webserver, CA, etc), and if those containers don’t exist, it will build a docker image for said workflow based on a standardized format (you guessed it) bash build script for the containers. There is automation via cron on whatever host runs docker to build and update the containers once a week, bare-metal servers update themselves nightly, rebooting when necessary via unattended-upgrades.

Basically, you break everything down into the simplest function possible, have everything defined via variables in shared configurations that everything sources before running, and you have higher and higher level functions call other functions until you have a single function that cascades into a functioning system. Does that make sense?

Have you started collecting your notes into scripts?

Not sure if many people do what I do, but instead of taking notes I make commented functions in bash. My philosophy is: If I can’t automate it; I don’t understand it. After a while you build enough automation to build your workstations, your servers, all of your vms and containers, your workflows, etc, and can automate duplicating / redeploying them whenever required. One tarball and like 6 commands and I can build my entire home + homelab.

I think you’ve correctly identified their self-interest over altruism, but you’ve misidentified the internal value of discouraging clickbait. YouTube is a treasure trove for building training datasets, and its value increases when metadata like thumbnails, descriptions, titles, and tags can be trusted.

It’s the AI gold rush; notice how this coincides with options to limit or disable third-party training but not first-party training? It coincides but is definitely not a coincidence.

Nothing beats the bang/buck ratio of used enterprise hardware (always buy new drives though if you care about the data)

https://www.theserverstore.com/ https://www.serversupply.com/ https://www.servermonkey.com/

I’ve bought from all of these in the past, personally I’m a fan of dells but there are arguments for just about any of the major 3 (dell, hp, sueprmicro)

Personally my main server right now is an r630. 96 threads, 768gb of ram. With that many memory channels, not only can you run all of what you listed, you can even do medium-sized inferencing/diffusion if you’re interested in that sort of thing.

Idk, this was kind of a rare combination of “write secure function; proceed to ignore secure function and rawdog strings instead” + “it can be exploited by entering a string with a semicolon”. Neither of those are anything near as egregious as a use after free or buffer overflow. I get programming is hard but like, yikes. It should have been caught on both ends

Because that bug was so egregious, it demonstrates a rare level of incompetence.

Because $350 couldn’t possibly buy enough hardware to run a modern operating system!

• Microsoft, probably

This is probably the play they’re making; the only thing that makes me think it might be something else is that they also announced ditching proprietary code in favor of kvm in workstation. Makes me wonder if they instead are deciding to slowly kill the product line, and instead of just stopping development entirely, they’re giving it out as if it’s some huge gift to try and “buy” good will before it becomes an inferior product?

Either way, support costs for the product are now $0 (because you can’t buy it) and development costs are about to be near-zero if they’re forking upstream kvm.

https://www.phoronix.com/news/VMware-Workstation-KVM

The thing about rational actors, is when given the same information they should make the same choices. I would argue that they’re most likely, instead, just at the peak of mt. stupid

We shouldn’t blame the victims that society failed to properly educate. You’re right that if people intimately understood apple the way you probably do, they’d never buy an apple product. I would argue, however, that it’s a failing of education not an informed choice to be corporately cucked.

I don’t think anyone should expect a battery replacement to be free after 10 years, but it shouldn’t cost $100,000

Just because you can’t use it doesn’t mean a hacker can’t. If someone discovered a vulnerability in the 3g handshake or encryption protocol, it could be an avenue for an RCE.