zbendefy

Similiarly I wonder why we dont run our own email server despite the sensitive data there.

Dont have much experience how well they perform but quantized models can run on ~4 H100 or A100 which should be true for kimi k3 and qwen 3.8 as well.

I think they didnt have the amount of debt /negative cashflow as they do now.

I dont have estimates on the cost of running models, but I think openai and anthropic are running on subsidized prices. At actual prices it might be worth it in the future.

What harness are you using?

Maybe we would get a microphone on mars. Just kidding i know air pressure is vastly different, but still it would be cool to listen to ambient sound from there

Good point!

There is also the case for Markov chains being theoretically able to do these if tuned well. Or even SAT problem.

This reminds me of stuff like interns wiping the production servers.

The solution in the comments there is always "have backups" and "why can an intern do that stuff? why arent policies set".

But I dont think these can be applied to something like openclaw agents

I did the same with an rpi3, not sure if I used this guide but it seems good:

https://www.raspberrypi.com/news/printing-at-home-from-your-...

i think so. The biggest hurdle with new languages is that you are cut off from a 3rdparty library ecosystem. Being compatible with C 3rd party libraries is a big win.

Tangential, but Flash had a nice side effect that the "app" could be exported in a self contained way via SWF.

Exporting this site for example in a future proof way is not that obvious. (Exporting as pdf wont work with the webgl applets, exporting the html page might work but is error prone depending in the website structure)

50 years from now, flash emulators will still work on swf files, but these sites might be lost. Or is there a way to archive sites like this?

>The user writes the data to CPU mapped GPU memory first and then issues a copy command, which transforms the data to optimal compressed format.

Wouldnt this mean double gpu memory usage for uploading a potentially large image? (Even if just for the time the copy is finished)

Vulkan lets the user copy from cpu (host_visible) memory to gpu (device_local) memory without an intermediate gpu buffer, afaik there is no double vram usage there but i might be wrong on that.

Great article btw. I hope something comes out of this!

What has changed now in the memory landscape/ai workload in the recent months compared to summer or spring?

there is also a vr mod for HL1 as well

My takeaway is that sandboxing should be more readily available, and integrated into the OS.

I used sandboxie a while ago for stuff like this, but afaik windows has some sandbox built into it since a few years which I didnt think about until now.

Why not get the wifi enabled fridge and just not hook it up to your router?

Genuinely asking because I plan to do this once I have to get new appliances, is there something missing that way?

Thing is (correct.me if Im wrong), that if you use modules, all of your code need to use modules (e.g. you cant have mixed #include <vector> and import <vector>; in your project). Which rules out a lot of 3rd party code you might want to depend on.

How different is CBOR compared to BSON? Both seem to be binary json-like representations.

Edit: BSON seems to contain more data types than JSON, and as such it is more complex, whereas CBOR doesn't add to JSON's existing structure.

its 12 years not 22.

An embedded device bought today may be easily in use 12 years from now.

offtopic: why is the Copyright © icon shake like crazy at the bottom of the page?

Edit: Oh I guess it seems to be intentional, I clicked around and I like the rgbcube site map.

This is such a good analogy. Awereness about social media shluld be like awereness about junk food you consume.

Is it wrong?

Correction, I didnt remember correctly as those are actually there in the code release, here is the r200 (radeon 8500) renderer specific code:

https://github.com/id-Software/DOOM-3/blob/master/neo/render...

Not just that, they had specific renderer backends, one for GeForce, one for GeForce3, one for Radeon 8500 that they had to cut out as they used proprietary information or code perhaps.

Aren't registers fixed by x86_64, while cache is a CPU hardware specific thing (e.g.: newer cpus have more cache than older ones, bit register count is fixed 8 on x86 and 16 on x86_64)?

So I think the compiler can work with registers at compile time but cant work with an unknown structure of cache

No, the full R1 model is ~650GB. There are quantized version that quantize it down to ~150GB.

What you can run locally are the distilled models, that is actually LLama and Qwen weights further trained on R1's output

I dont think they rent gpus for $5million because its cool and want to show the world...

Note: you are probably running a distilled version of R1, which is actually LLama or Qwen further trained on the input/output of R1.

The full R1 is huge (~700GB), altough there are still quantized versions, the smallest one is around 150gb (1.58bit)

Also without the "attention is all you need" paper from google

I think some state is also being given (or if its not, it could be given) to the network, like 3d world position/orientation of the player, that could help the neural network anchor the player in the world.