How can we lock down a TV, still use it but trust we aren't being spied on?

Don’t give your TV internet access, and use an Apple TV.

At what point do we need to contend with an out-of-band mesh network? Sure you didn't connect your TV to your LAN, but did anyone within LoRa range, on any device from the manufacturer?

Create an open Xfinity Wi-Fi network, as well as an AT&T one. See if the TV connects to them.

This is probably obvious to most people, but I will mention just the same: do not install the antenna that the TV comes with.

> This is probably obvious to most people, but I will mention just the same: do not install the antenna that the TV comes with.

Wifi antennas would typically be internal and unrelated to the one F-Type coaxial connector one would typically think of as the "antenna" input.

Can't really stop the internal module from scanning and connecting to open access points unless you control the software.

We do that but use a Kubuntu Linux laptop behind the TV and a wireless mouse and keyboard (when we want to pause a movie in Chrome and search for something, like details about an actor in a movie).

I tried doing this and the video stutter on a raspberry pi made it too annoying. Probably better off using an N100

> Probably better off using an N100

Exactly this. Search for a mini PC and it'll have a cheap Intel SoC in it that performs way better than a Raspberry Pi 5 but cheaper (comes with everything for ~$300), and even consumes a similar amount of power.

That's too bad! Which Pi version and RAM configuration did you try?

A heads up: my CachyOS gaming PC has been doubling as my media box. But sadly, I've noticed that modern streaming DRM limits the quality because the OS+hardware is not "trusted". There's probably a way to hack around it, but I haven't gotten into it

There it is. Not a single TV at my house ever seen a WiFi password. I add cheap Amazon FireStick and be done. Sure Amazon also spies on me but as far as I know original FireStick does not have built-in microphone and probably is not scanning/copying my LAN content to some cloud/back home. And at any time I can just throw it in the trash, get another $29 device. Once I/you plug WiFi password to your TV, there is now way of coming back.

Leave it offline.

Thankfully, there are still people willing to jump on the R9700 bandwagon and get a vLLM fork working. If you have an RDNA4 card check out https://hub.docker.com/r/stilldeadcode/vllm-radiance

Deadcode is currently working on INT4 right now on his R4D kernel.

The MXFP4 fork is excellent too. Its my daily driver right now. https://codeberg.org/ggz14/radiance-vllm-mxfp4

Also has PARO quant support there too (early stage)

Also speedups in both repos for 4x R9700s

The way it does tensor splitting without all-reduce cost over PCIe bus wasn't something I thought was possible.

What kind of performance are you getting with 4x R9700s--what do you do with all the VRAM (batching, concurrent requests, etc)?

personally? i have 2x gpus.. but i get bursts of ~200tok/s generation, and around 4500-5000tok/s prefil

Yeah the R4D Kernel rules imho.

Similar here peak ~250 and down to ~120 as it gets close to 128k (which is where I set DSH compaction) though it can readily do 256k.

I just got DeepSeek Harness (DSH) set up with 2x R9700 and it's rather mind blowing that these can do actual work and quickly. Up until now I've always been evaluating and searching for better hardware/model/tweaks. This is much more than I even hoped for and considered getting extra 3090/4090. Now I can stop looking/tweaking and start using it for all the different things I've yet to discover it's good for. I do plan to also try/use Hermes and Pi. DSH is annoying that every plugin install/remove requires a restart--given that "everything's a plugin".

Thanks! Didn't expect to see this here. Exactly what I needed to run Qwen3.8-27B-Quark-AWQ-MXFP4-native.gguf as well as other experiments on one or 2x R9700's (I hope).

That MXFP4 is an excellent project. But I have difficulties on reading that README. Is it intentionally generated like that with LLMs?

Feed the setup and run scripts to your LLM.

I think messing up teaches people to straighten up. But hey I’m a Millenial, maybe I’m out of touch now.

Buy yourself 6 AMD r9700's an older 5965wx thread ripper with mob, 128GB of ddr4 and run deepseek v4 flash lossles. Then pocket the other $6000

You're going to need some kind of PCI switch or bus bifurcation card and a way to house those cards.

How many toks/s could you get this way?

Between 20 and 50 depending on context length. Power consumption between 900 and 1500 W for the whole machine.

Omg C, so insecure! You should be ashamed of my insecurity!

I'd really like to see a 45B-ish dense model ready for a dual GPU setup. Something with a little more intelligence while still within the range of some higher end local setups.

There is definitely an under-served target memory size of 48GB - almost everything aims for: 12, 16, 24, 32, 64, ...) But most dual-gpu setups, 3090/4090 (and some mac configs afaik) have 48GB, and most 64GB systems would do well with the extra 16gb of overhead saved. 48GB is also moderately common in PC memory configurations since 24gb DIMMs are a thing.

I wish we could stop sensationalizing this about the AI and really just understand the incompetence of the labs disabling an internet connection in a sandbox.

As written it sounds like you're saying that it was incompetent of the labs to disable the sandbox internet access?

They tried to disable open internet access but the models zero-day'd their Artifactory package registry and got internet access anyway.

No sensation... that's just what happened.

If you really wanted to sandbox a machine you’d offline cache the packages and not give it any physical route to the internet, not via a jump box, not via a proxy, nothing.

This was poorly executed.

I don’t really know how these training runs operate in reality. But I assume it’s using a lot of raw GPU power directly. It’s hard for me to visualize how exactly you’d go about completely cutting off these datacenter and cloud resources from the internet without actually going there, unplugging the WAN connection, and physically typing out what you need to happen on the cluster.

It seems like whatever virtualized sandboxes they have are not enough. But it’s equally hard to imagine their SWEs jumping on a plane to a data center to do this work locally

They literally gave it a proxy to the internet (artifactory). The only thing between the model and the internet was Artifactory.

You can take far greater measures to lock down external traffic than just that.

An offline package cache (aka artifactory WITHOUT its own internet access) likely would have precluded this whole thing.

This is a company with insane amounts of money, they can afford to fly techs wherever they need to for as long as they need to be there.

Air gapped environments are nothing new and they're standard practice for sensitive applications.

DMZ - https://en.wikipedia.org/wiki/DMZ_(computing)

Unplug the ethernet cable leading to the outside world, then?

Are you suggesting that training agents to have the sole goal of exploiting security vulnerabilities isn't the incompetent part of this, but that the sandbox wasn't secure enough?

Would we apply this logic to literally any other technology?

As AIs become more capable, the level of competence required to avoid disaster likewise goes up over time.

The whole site is suffering from AI psychosis.

So I always make sure I take a crack at doing what I want first. Then I ask for an AI review and it usually has a more efficient way to get the job done.

For example I had a working linear decay velocity boost function added to a godot scene and it's suggested refactoring actually reduced lines of code just today. So kind of happy when that works out, but yes I also applied the changes manually so I would grok them and hopefully remember it better.

I wonder if this would work with the oculus pro?

So I'd like to see a Nemotron3 Ultra converted to a 1.58bit format then have them retrain on the open dataset.