That’s mostly true. But if you have a GPU to play video games on a PC running Linux, you can easily use Ollama and run llama 3 with 7 billion parameters locally without any real overhead.
Azure/AWS/other cloud computing services that host these models are absolutely going to continue to make money hand over fist. But if the bottleneck is the infrastructure, then what’s the point of paying an entire team of engineers 650K a year each to recreate a model that’s qualitatively equivalent to an open-source model?
The engineers can generally also do other things, the security will likely be better, and its fully possible API costs will exceed that sum if you need that much expertise inhouse to match your API usage.
Just an off-the-cuff prediction: I fully anticipate AI bros are gonna put their full focus on local models post-bubble, for two main reasons:
-
Power efficiency - whilst local models are hardly power-sippers, they don’t require the planet-killing money-burning server farms that the likes of ChatGPT require (and which have helped define AI’s public image, now that I think about it). As such, they won’t need VC billions to keep them going - just some dipshit with cash to spare and a GPU to abuse (and there’s plenty of those out in the wild).
-
Freedom/Control - Compared to ChatGPT, DALL-E, et al, which are pretty locked down in an attempt to keep users from embarrassing their parent corps or inviting public scrutiny, any local model will answer whatever dumbshit question you ask for make whatever godawful slop you want, no questions asked, no prompt injection/jailbreaking needed. For the kind of weird TESCREAL nerd which AI attracts, the benefits are somewhat obvious.
I guess it depends how you define what an “ai bro” is. I would define them as the front men of startups with VC funding who like to use big buzz words and will try to milk as much money as they can.
These types of people don’t care about power efficiency or freedom at all unless they can profit off of it.
But if you just mean anyone that uses a model at home then yeah you might be right. But I’m not understanding all the harsh wording around someone running a model locally.
you almost always get better efficiency at scale. If the same work is done by lots of different machines instead of one datacenter, they’d be using more energy overall. You’d be doing the same work, but not on chips specifically designed for the task. If it’s already really inefficient at scale, then you’re just sol.
The whole point of using these things (besides helping summon the Acausal Robot God) is for non-technical people to get immediate results without doing any of the hard stuff, such as, I don’t know, personally maintaining and optimizing an LLM server on their llinux gaming(!) rig. And that’s before you realize how slow inference gets as the context window fills up or how complicated summarizing stuff gets past a threshold of length, and so on and so forth.