AI Recomping Controversy, PC Games on Switch and more (ft. Scarlet) - Tempcast #99
Theme by VinsCool and AlanJohn
The podcast is also available on Spotify!
I'd prefer to run local models when possible.I'm starting to wonder if you're actually an AI, this hallucination about Pandora's Box doesn't help.
Supply and demand is still a thing though, if fewer people used cloud AI the demand/justification for data centers also goes down and fewer data centers likely means lower silicon sales/prices.
You're also assuming that they don't just shift manufacturing over to HBM, which I wouldn't be surprised if that somehow makes things worse.
I'd prefer to run local models when possible.
Yes very likely so because it takes a lot of compute to run models as good as say, Claude Opus 5.5. The good models are out there and quantized, however when you lower their training set and make it smaller, therefore making it able to run on more and more computers, you also make it dumber. So you can try to run Opus 5.5 locally, and if you wanted to run it on say, your gaming laptop, you'd have to run a heavily quantized version. Specifically getting a PC setup outfitted for the biggest, most powerful local models, that is a must for running the more non-quantized models. Source on Quantization if you're interestedLocal models I don't have as much of a problem with since obviously that doesn't contribute to data center demand, at least not as much, that's why per my quote and when I usually complain I specifically say "cloud AI".
However, I have a suspicion that allot of people just use cloud AI.

Even with a lot of VRAM at your disposal (I tried on a Mini PC with 128GB of unified memory) the local models just don’t seem to handle complex tasks. I’ll admit I haven’t revisited it in a while though. I’m sure they will get better.Yes very likely so because it takes a lot of compute to run models as good as say, Claude Opus 5.5. The good models are out there and quantized, however when you lower their training set and make it smaller, therefore making it able to run on more and more computers, you also make it dumber. So you can try to run Opus 5.5 locally, and if you wanted to run it on say, your gaming laptop, you'd have to run a heavily quantized version. Specifically getting a PC setup outfitted for the biggest, most powerful local models, that is a must for running the more non-quantized models. Source on Quantization if you're interested
I personally use Claude for my projects (yes I'm an AI person, hate me if you choose, but at least I admit it). I've tried local models for experiments, learning how to setup local models et cetera and doing demonstrations for classes, and what I can say, is the cloud models are more accessible, and also more powerful than what many people can afford to do locally. And as for my position on AI: I think AI's bad, but I also know it is here to stay. I do believe there is a bubble and even though I use AI, I still want it to pop. I do not support AI for anything that takes artists jobs. Other than that, I'm like neutral to leaning towards AI.