nicolas

This is just an intermediate state

Considering they have a Tesla factory in Germany, how many of those are sold to employees?

Grok3, Claude 3.7 and GPT 4.5 all within a couple days. It’s crazy how consistent the progress is that the labs are making.

YouTube video by TJ DeVries The Only Video You Need to Get Started with Neovim

youtu.be/m8C0Cq9Uv9o?...

How is grok3 so fast if it is so huge? Any ideas?

Quality content

Openwebui is really a nice package and somewhat of a UX sweetspot. It’s great for daily tasks and for experimenting with LLM use cases (especially with its tool and function calling features), but the codebase feels very stitched together.

It’s a convenient way to get so much memory in that small of a form factor

Critical information to keep in mind when interpreting the o3 results on the ARC-AGI benchmark (vs. o1):

1. o3-Inference cost a staggering $1.15M.
2. o3 was trained on a public ARC-AGI dataset, giving it a massive advantage over o1 which had no access to that data.

It’s not a fair comparison.

LogReg is the true hero of data science

Pie chart: 74% hydrogen 25% helium 1% other (in tiny font)

Atoms are binary. They are either intended to be hydrogen or helium. We can't just scrap this worldview just because of a handful of exceptions

There is also llm.datasette.io/en/stable/

Research like this is so important. A lot of decision making when deploying RAG systems is based on vibes-based eval using dated models and datasets. Many techniques simply don't transfer to modern models like you'd hope. Query expansion seems to fall into this category.

Production logs could be replaced/extended with "Tracing", which is actually a structured way to look at what the system does in real live and the user interactions, very popular in LLM land.

Would you call it open weights if it is a restricted license that requires a separate license agreement for commercial use?