I use a Reolink NVR instead of my own server, along with SD cards in the cameras themselves as a backup. My switch has VLAN capabilities, but I haven’t configured my cameras to use it, mainly because I have Wi-Fi cameras outdoors (2 solar-powered and 1 doorbell camera connected to doorbell power). It’s a proprietary local setup that “just works”. I understand not trusting proprietary and wanting self-hosted open source all the way because I certainly don’t trust them to not collect my data and don’t allow Reolink devices to have internet access.

I use Reolink cameras and recently got a Reolink NVR. I do block all Reolink devices from the internet in my OpenWRT firewall. I didn’t want to host a NVR on my server that’s already hosting a bunch of other stuff and further increase the blast radius of my server as a single point of failure. Reolink is all proprietary, but at least it’s local and works without internet access.

LLMs probably created a lot of these bugs in the first place.

Open weight is like getting a compiled binary with no source code. Without the training data and pipeline code, once the original developer stops releasing new versions, it’ll be difficult for the community to pick up and keep going.

For example, if Alibaba decides to only release trillion parameter models from now on, Qwen 3.6 may be the most intelligent Qwen model that will ever run on consumer hardware.

I’m still glad we have Wikipedia as a foundation instead of a publicly traded corporation. The fact that there are way worse offenders doesn’t exempt them from scrutiny, though.

I wouldn’t trust them enough to run anything important.

Qwen 3.6 still works fine on my gaming PC that didn’t cost anywhere near that. ¯\(ツ)/¯

This opinion piece from the BMJ entitled “The illusion of evidence based medicine” sets the stage: https://www.bmj.com/content/376/bmj.o702.

There is so much corruption and industry capture that it can be genuinely difficult to separate the real, unbiased science from “Big Pharma make number go up”. Telling people they are uneducated dipshits and to just “follow the science” ignores and further degrades the fundamental lack of trust.

Eh, I might wait and see how the benchmarks stack up before I sell my house to buy 8 B200s.

On a serious note, what the hell is going on at Alibaba? Now that now that Lin Junyang is gone, 3.6 was the last release we can actually run and now nobody is releasing billion parameter models anymore? Hope the community gets really good at distillation. ☹️

“OpenAI” signed the open weight pledge, which only makes sense given the company name and original values. They’ve had gpt-oss for quite a while, but it feels like a pretty half-assed effort for the purposes of virtue signaling. I’m not holding my breath for them to release an open weight model on par with GLM, Kimi, or Qwen.

deleted by creator

Agreed, DI in moderation can be useful. I generally like to target dependencies with I/O so they can be substituted with test doubles to allow the domain logic to be tested with fast, deterministic tests. On the other hand, I’ve seen codebases that go way overboard with SOLID and create unmaintainable messes that are difficult to reason about.

I see all the tests are passing on this PR, so I won’t even review the code. LGTM.

🙄😖

I took his stance to mean use a LLM if you dare, but make sure Linus can’t tell you did.

I’m also surprised they compare themselves to Steam/Epic when DRM-free is GoGs whole thing

I get the feeling that a lot of the motivation for developing something like this is to play “DRM-free” games that weren’t originally that way. 😉

I have a saved bookmark to old Reddit for the one subreddit I follow. When I saw that, I immediately yelled out loud, “fuck you, Spez!” Startled my cat.

This looks fantastic! I’m currently installing games manually on a few household gaming machines and this would help to simplify that. It looks like parental controls are a TODO, which would be an important feature in my case.

Sorry to be pedantic, but I think it’s important not to call these models “open source”. Open weight is the equivalent of a compiled binary that can be run locally, so you’re basically getting Windows where it’s legal to reverse engineer, not Linux. The weights are permissively licensed, so the community is still free to quantize, fine-tune, reap, etc., which is analogous to something like Tiny11. Without the training data and the pipeline code, though, if Alibaba or Moonshot decide to go closed, the community won’t be able to fork their models and meaningfully continue development.

It’s better than what we are getting from “OpenAI” or Anthropic for the moment, but these are Chinese, for-profit companies that are no more altruistic than Google or Meta, and their open weight LLMs are just as likely to go closed as Llama or Gemma if the companies deem that to be in their best interest.

Good point, that should theoretically be double the decode speed, though not sure how the prefill would differ since it’s mainly compute-bound.