Georgi's relevant comment: https://github.com/ggml-org/llama.cpp/pull/19324#issuecommen...

Can someone add some context as to what that diff is showing?

and use the original llama.cpp directly. Its infinitely more easy to setup and use now

Setting up ollama is 2 steps:

1. yay -S ollama

2. sytemctl enable --now ollama

How is llama.cpp infinitely more easy to set up?

infinitely more easy relative to what it used to be

This is Yellow Pages type thinking in the age of the internet. No one is going to own an agentic layer (list any of the multitude of platforms already irrelevant like OpenAI Agent SDK, Google A2A) . No one is going to own a new app store (GPTs are already dead). No one is going to foundation models (FOSS models are extremely capable today). No one is going to own inference (Data centers will never be as cost effective as that old MacBook collecting dust that is plenty capable of running a 1B model that can compete with ChatGPT 3.5 and all the use cases that it already was good at like writing high school essays, recipes etc.) The only thing that is sticking is Markdown (SKILLS.md, AGENTS.md)

This is because the simple reality of this new technology is that this is not the local maxima. Any supposed wall you attempt to put up will fail - real estate website closes its API? Fine, a CUA+VLM will make it trivial to navigate/extract/use. We will finally get back to the right solution of protocols over platforms, file over app, local over cloud or you know the way things were when tech was good.

P.S: You should immediately call BS when you see outrageous and patently untrue claims like "Mac minis are sold out all over.." - I checked my Best Buy in the heart of SF and they have stock. Or "that its all over Reddit, HN" - the only thing that is all over Reddit is unanimous derision towards OpenClaw and its security nightmares.

Utterly hate the old world mentality in this post. Looked up the author and ofcourse, he's from VC.

> No one is going to own an agentic layer

Don't underestimate the capitalists. We've seen this many times in the past--most recently the commercialization of the Internet. Before that, phones, radio and television.

Fair enough. This is my war time rhetoric. Its up to us to effect the future we want.

> terms of statistics. You might want to read something called gauge R&R

I think the first thing to focus on is the stats portion - do you have appropriate FAI/SPC/OQC with Cpk requirements defined? Gauge R&R plays a much smaller role, especially in something that is relative

Good, important idea. Unfortunately bad, low effort vibe coded execution

Still a shipped idea, driven by someone. The author has some other interesting ideas.

does it matter if it shipped if no one uses it?

Claude Code has over 5000 open issues. And this is after issues that are inactive for 60 days being auto closed. Such a policy is facetious to say the least. What is more perplexing is why they don't use Claude to triage the issues?

I think the idea is that nobody will be using CC in 5 years. If anthropic loses, nobody will use it. If anthropic wins, still nobody will use it! The value is in the model solving problems, and CC is just the hacky vessel for that, not the end goal. If they believe in themselves, polishing the product is a waste of time.

Exactly. I listened all the speeches at the WEF about AI. Anthropic believes an AI God is coming.

> What is more perplexing is why they don't use Claude to triage the issues?

I wonder if it's because of cost.

There is an AI bubble.

Can drop the extra words

Reminder that this comes from from the founder that got rightly lambasted for his comments about work life balance and then doubled down when called out.

"Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something."

https://news.ycombinator.com/newsguidelines.html

You can see that Qwen3 does worse than Qwen2.5 on our benchmark. Reason is it's never been pretrained for FIM / autocomplete.