Hacker News new | ask | show | jobs
by stanislavb 26 days ago
This. They can see their valuations slipping. They hope that in a few/several years they will start reaping profits. However, in several years local hardware will be well suited to run models locally at 80-90% efficiency - for "free". You won't need frontier models for daily tasks in a few years. I'd guess.
1 comments
anuramat 26 days ago
> 80-90% efficiency

wdym by that

> for daily tasks

which are?

link
glenpierce 26 days ago
You get about 80-90% of the results for daily tasks like: getting summaries or explanations of complex material. Writing software tools for data analysis. Getting recipes for a given set of ingredients in the fridge.
link
anuramat 24 days ago
> 80-90% of the results

why would I ever want 80% of the results though?

link
numpad0 26 days ago
128B-A16B class models at 10-50 tok/s should be plenty for most tasks done on computers
link
julianlam 26 days ago
What do you need a frontier model for, really.
link
ekidd 26 days ago
Frontier models like Fable are mostly useful if you want to paste in one or two prompts, and receive a subtly broken application that looks impressive. That is very hard to do with local models today.

What current local models work fine for is delegating clearly-described tasks in a code base the programmer actually understands. Qwen3.6 27B and DeepSeek V4 Flash are both great little workhorses.

There's also GLM 5.2, which is kind of like "store brand Opus", and which might be considered a "near-frontier" model. I don't have as much experience with it.

link
int_19h 26 days ago
FWIW Fable is insanely expensive for the task you just described, so much so that I don't think it's practical for that. Its practical use is as a dev lead / architect / project manager model, doing planning and writing detailed feature specs and code reviews while Opus/Codex/Gemini does the actual coding.
link
anuramat 24 days ago
research, refactoring, architecture

also just code-that-is-not-as-shit in general

for clearly defined stuff smaller models are fine, but at that point I'd rather write the code myself

in any case, if there's a chance a cheaper model will fuck it up, is the price diff worth gambling with your time?

link