[jart]

Mozilla's llamafile project is designed to enable LLMs to be preserved for historical purposes. They ship the weights and all the necessary software in a deterministic dependency-free single-file executable. If you save your llamafiles, you should be able to run them in fifty years and have the outputs be exactly the same as what you'd get today. Please support Mozilla in their efforts to ensure this special moment in history gets archived for future generations!

https://github.com/Mozilla-Ocho/llamafile/

[visarga]

LLMs are much easier to port than software. They are just a big blob of numbers and a few math operations.

[andix]

I think software is rather easy to archive. Emulators are they key. Nearly every platform from the past can be emulated on a modern arm/x86 Linux/windows system. Arm/x86/linux/windows are ubiquitous, even if they might fade away there will be emulators around for a long time. With future compute power it should be no problem to just use nested emulation, to run old emulators on an emulated x86/linux.

[throwaway314155]

> I think software is rather easy to archive.

* assuming someone else already spent tremendous effort to develop an emulator for your binary's target that is 100% accurate...

[andix]

The reality is, that someone else already spent a tremendous effort of building emulators. Or do you know any older platform that can't be emulated? 100% accuracy is not needed, that's not possible. Even current hardware is not 100% accurate and has bugs/flaws.

[refulgentis]

LLMs are much harder, software is just a blob of two numbers.

;)

(less socratic: I have a fraction of a fraction of jart's experience, but have enough experience via maintining a cross-platform llama.cpp wrapper to know there's a ton of ways to interpret that bag o' floats and you need a lot of ancillary information.)

[_ea1k]

Indeed. In 50 years, loading the weights and doing math should be much easier than getting some 50 year old piece of cuda code to work.

Then again, CPUs will be fast enough that you'd probably just emulate amd64 and run it as CPU-only.

[jart]

llamafiles run natively on both amd64 and arm64. It's difficult to imagine both of them not being in play fifty years hence. There's definitely no hope for the cuda module in the future. We have enough difficulties getting it to work today. That's why cpu mode is the default.