389
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 01 Dec 2025
389 points (100.0% liked)
Fuck AI
4683 readers
1755 users here now
"We did it, Patrick! We made a technological breakthrough!"
A place for all those who loathe AI to discuss things, post articles, and ridicule the AI hype. Proud supporter of working people. And proud booer of SXSW 2024.
AI, in this case, refers to LLMs, GPT technology, and anything listed as "AI" meant to increase market valuations.
founded 2 years ago
MODERATORS
Gotta be honest, though, a locally hosted 70B model with basic RAG functionality isn't exactly playing in the same league as the market leaders, which can be bigger by two to three orders of magnitude. And a model that size is already around the limit of what a beefy gaming PC can do with reasonable performance. We're unlikely to ever beat the big players on quality with local models.
What might happen is that the market collapses, the big players all go bankrupt, further LLM development ceases, and locally hosted Qwen3-80B will be the pinnacle of available text generation for the next thirty years.
I wrote "they are becoming much more powerful". The point isn't to beat big players, the point is that we will be able to run models that are just capable of what need, not the super smartest models available. Your last sentence I agree with, that may very well be what happens, except by then we'll have Qwen4 and it'll be even more efficient and more powerful.
I predict incremental quality increases. Qwen4 will probably be a somewhat better Qwen3 (and a dud if we're unlucky). I do agree that it'll probably come out; there's not enough life left in this AI boom for a Qwen5, though.
The biggest change will probably come from figuring out where LLM use will actually benefit us. Right now the industry zeros to answer that with "everywhere" and concludes that it's prudent to spend money equivalent to the GDP of an industrial nation on compute-only data centers.
For example, I expect the use case for coding to be more like "autocomplete a code block based on known patterns" rather than "build a public-facing web application from a prompt".