@adiix_official Same lunchbox runs MTP qwen 27B at 1000tps eval and 60 tps output. It's a thing of beauty for sure, but long contexts crash hard. A 80k token context starts eval at 1k and drops to 300, some patience is required. That being said, more than usable. Team red is goated.
@precisox We've had one of these for a while and it's nice to play with but it's hardly making a dent in the local AI scene.
Most small LLMs run but generation is slow. Interesting models are heavily invested in NVIDIA tooling and you have to patch things to get them work.
@cmuratori@Mk2Sektor Programmer here. We've been extensively using someone else's work (libraries, frameworks, APIs) since forever. AI churned code isn't that much different. Licensing is the pain point.
@VaibhavSisinty How the hell do you guys do it? I just asked it to review a 100 line PR and waited 20 minutes for "only a minor issue on line 70".
I'm so skill gapped it's not even funny...
@heyvasos@Blender@threejs It does survive it. It's pretty amazing, meshes require little to no cleaning, based on my relatively simple tests. It also does wonders with tiny image inputs (150x50px).
@TBashscripts@BrodieOnLinux I'm my experience, such legacy systems run ubuntu 14 and have apt-get aliased to "echo 'DO NOT TOUCH THIS SYSTEM!!!1!'" and are as-is ๐
@MetalBlakedowns My savior! This song has been stuck in my head for two days and I couldn't remember the name. This turned up with a Google search ๐
@iced_coffee_dev Can't wait to actually get this! Simon, your yt videos pushed me to dive into one of my childhood passions and I can't thank you enough!