HN Explorer

Ask HN: How close are we to local LLMs being useful? What's the impact?

7 points by AbstractH24 ↗ HN
Feels to me like local models are an under-covered aspect of this whole AI boom.

If everything improves over time, at some point a good chunk of tasks won’t need to be done in data centers or be subject to the whims of a few frontier AI labs.

How close are we to that? Or is my thinking flawed?

9 comments

[ 3.9 ms ] story [ 20.6 ms ] thread
buffer_overlord
until its cheaper to train and infer than 100k gpu data centers...i doubt it will ever compete.
PaulHoule
I do classification with SLMs and for my tasks when I have a few thousand samples the frontier models in zero-shot and few-shot modes are embarassingly bad in comparison.
david927
I think we're past that point; they're absolutely useful already for a lot of tasks. I think it's about costs, convenience, and benefits of a frontier model for what you're doing.
AbstractH24
Which ones are most useful? Any suggestions on where to go to start exploring this world?
segmondy
Local LLMs have been useful since 2024. If you don't know this then you are just far behind. Catch up!
montfort
Take a look at the AMD Ryzen AI Halo Developer Platform with a Ryzen AI Max+ 395 processor. These systems, with 128GB of unified memory and their specialized processors, deliver greater performance and inference power than Apple's Mac Studio. This already allows you to run fairly decent models for personal classification and coding tasks. I think for complex design needs you'll still require frontier models that can only be run in large data centers, but much of the underlying work could already be done on-premises.