Quoting: g000hYou can fit a small local model in 500MB, although naturally its accuracy will suffer.

For me, I only want to run models using my own choices of software. For instance, I would not be happy with Ubuntu or some corporation pushing some agentic features alongside the LLM. The agentic side is where A.I. gets into mischief - It exposes security holes, e.g. prompt attacks which can affect backend systems.

Also, agentic features enable Big Tech corporations (Microsoft, Google, OpenAI) to conceal telemetry and other self-serving capabilities (backdoors, etc). Naturally, any built-in agentic capabilities would be the first place hackers will be aiming to exploit.

For camera focus, which is one of the features evoked by Seager, Canon, Sony, Nikon, Olympus, Panasonic and Pentax had integrated neural network trained using Deep Learning for autofocus in their DSLR and Miroless for around 15 years. This is notably how they greatly increased the number of species on which focusing using the "Focus on eyes" feature works.

So "AI powered" camera focus can use very small models depending of what you call AI powered.

View this comment - View article - View full comments

IMO there are only three features using LLM which should be installed by default (on a desktop setup): LLM based text-to-speech, LLM based speech-to-text/voice commands and image description tied to tts.

And only the two first features should be active by default and opt-out with the first page of the installer asking if you want to keep them active or not. And the third features being opt-in with no model pre-uploaded (but uploading a local model being easy).

Because they are accessibility features.

View this comment - View article - View full comments

You forgot about the incoming Oculink Dev Kit. To summarize this is a module you add in the expansion bay to give a Oculink 8i port to the Framework 16 plus an Oculink 8i PCIe Dock size for normal GPU + corresponding Oculink cable plus a Oculink 8i dock to use the Framework 16 GPU module externaly. (A bit less useful).

The fact that it uses Oculink 8i means that this is PCIe x8 with 128 Gbps.

View this comment - View article - View full comments

IMO Valve is experimenting to find the minimal viable product allowing customers to install Windows 11 if they aren't happy with SteamOS limitations (*cough*anti-cheat*cough*). Once you have the minimal viable product, you can estimate the minimal production cost and start looking at what kind of compromise you can do to offer more than the mvp while keeping the cost low.

Plus we have idiots like EA rolling out games using Secure boot and TPM based anticheat which are defeated after one day but the idiots will still keep requiring Secure Boot and TPM as implemented on Windows 11 to launch the game. So until Valve brings up their own server side anticheat to stuff and convince corporates to switch to it, the best Valve can do is pay developper to improve Secure Boot support and TPM 2.0 on Linux so they can implement it on SteamOS. However you cannot test potential OS features which haven't been developped yet on present hardware. Leaving only Windows 11 to test if their hardware is compatible with EA's Secure Boot and TPM based anticheat.

This is my opinion on why Valve.would test a Steam Machine running Windows 11 with only 8Gb of RAM for the CPU (the GPU on the prototype seemingly having its own separate VRAM).

View this comment - View article - View full comments

This makes it a viable alternative to the Gamesir Super Nova in Dual Shock 4 mode (where the Gyro is usable on Steam while keeping analog triggers), with even an advantage for the back buttons. (The Gamesir SN is also a "back buttons can only mapped to other physical button. But you can do it without the Gamesir software.)

View this comment - View article - View full comments

@TrainDoc I think that the main issue is the name. Because the Framework Desktop is more a high end mini-PC or a NUC than a standard desktop. The fact that they put it in concurrence to the Mac Studio (above $4000 for 128Gb RAM) and Nvidia Digit (price unknown but suspected to be above $3000) shows it. (And AMD announced they will using the FD as dev kits to distribute to the ROCm community make the comparaison to the Nvidia DIGIT even more true.) And indeed the Framework Desktop is more modular and repairable than those two, and that half of other mini-PC. Because not only most mini-PC use custom motherboard formats unlike the Framework Desktop mini-ITX and are usually no swappable and not usable into an ATX, but half of them also use mobile, soldered CPU, and Apple and Mac are far from the only ones to use soldered RAM.(And aren't made to be run in cluster either.)

So in the world of mini-PC the FD is above average in repairability and upgradability, even if half of the Windows mini-PC beat them in RAM upgradability. Morever when Windows mini-PC are still dominated by Intel, so even with socketed CPU, the fact you cannot swap motherboard mean you can only at most update the CPU by one gen if you are lucky.

And we need to keep watch for future gene of motherboard, because on the LTT video, Nirav Patel talked in a bit more detail about Framework and AMD's failed efforts to get the Strix Halo APU to work with socketed RAM. He mentioned that AMD assigned one of their technical architects on the subject but determinated that the way the current 256 gb/s memory bus work on Strix Halo, the signal integrity isn't good enough outside soldered RAM. So hopefully this is more a Strix Halo flaw, and future APU on this category might work with CAMM2 RAM.

View this comment - View article - View full comments

Quoting: pb

Quoting: basedthe access to their stores being SOLD on Steam

My point exactly. If it wasn't taken down by Valve, it would have been taken down by someone else. It's was the only right move to nip it in the bud.

Valve's legal Team probably barged in the office of the Steam admins screaming "Stop this folly."

View this comment - View article - View full comments