drittich

Perhaps that's a false dichotomy?

Also exists in LM Studio.

I played with gemma-3-4b-it-qat recently using a mid-tier graphics card and a few things stood out to me:

1. It was very fast, between 35 and 70 tokens per second, with initial response in under 200ms. That kind of speed is a feature.

2. It was very useful. I had a brainstorming session with it that was both fluid and fruitful

3. I can't wrap my head around so much knowledge being contained in about 3GB of data. It seems to know something about everything. Imperfect, but very useful.

False dichotomy. An alternate position is to do it in a measured, planned way, not under duress as the economy tanks and international relations are soured.

I think the main point stands, though, which is that you can't undo to the previous state. E.g., rolling back all tariffs/deportations/firings/budget cuts would not undo the damage done.

It's like when you are in a traffic jam. It's the other drivers fault, not you!

I am going to hazard a guess that it's because the closer the services are to your data, the easier they are to implement, and you often get great speed too. FTS in Postgres has been fantastic for me, and combining it with vector search and RAG gives you a pretty sweet deal for low effort.

See also:

* An Introduction to the Delia Derbyshire Archive [video](https://www.youtube.com/watch?v=pPIqq7_RjnA)

A GREAT documentary.

Thanks, sounds amazing! I've checked it out from local library.

I'm wondering if any glass plates were recovered? I'm not sure they were all salvaged at the time, but I don't know how they would fare in the saltwater.

+1, there are definitely some sad parts to this story.

Ah, but where were the survivors buried?

Thanks, much appreciated!

How do you weight results between vector search and bm25? Do you fall back to bm25 when vector similarity is below a threshold, or maybe you tweak the weights by hand for each data set?

I must have missed that feature. I don't see it under Data Source or Integrations. Can you point me to it? Use case is mostly call an API and show one or more property values from the JSON object that comes back, or use it as a tabular data source.

Edit: Maybe I'm misunderstanding the UI and this is something you would do directly through Python?

Will there be the ability to call an API and parse the JSON (e.g., to flatten) to use as a data source?

While your original comment might be taken as funny, this one quickly devolved into a strawman argument ending in a personal attack. Was that your intention?

I like the idea! A few things I noted: 1) despite the paged UI, the API call /rest/v1/arxiv_papers?select=*&order=published.desc is downloading all 708 articles. You will find the UI more snappy if you do server-side paging. b) most of the javascript is cached, but not all of it, e.g., page-script.js

I tend to agree. I'm interested in the concept, but the opening line seems to justify its need as a reaction to s-expression syntax in Lisp. Knowing nothing about that, I fear I'm going to miss the context of the whole article, and I also can't determine whether this is a straw man or not. And, as another commenter mentioned, it makes this whole thing feel like it's serving a very niche need. Which doesn't jive with the title, which is very generalized, and could be quite a compelling concept.

Someone else mentioned babylon.js in this thread. Did you evaluate that and choose three.js over it?

Please explain what ANY expands to...

I just got interested in Mega Man recently - do you recommend a playthrough in order, or is there a best version to play for a good initial experience?

I completely get your take on this and don't disagree. IMO it reads as a flippant response to a potentially serious question, and I am going to assume this is unintentional, which does not mean it is unfair to call it out.

Stepping back, this is the kind of discourse that Twitter can sometimes reinforce: short responses that, lacking context, can be interpreted in polarizing ways. After a bit of reading on the participants (because both of them are working in interesting areas), my belief is that the "pencil" response is actually shorthand for a whole set of beliefs that provide a lot of context and nuance to the discussion, but if you don't know that, it sounds like saying "AI is as dangerous as a new brand of chewing gum".

In addition, without defining what risks we're talking about, it's really hard to know the scope the answer is addressing. E.g., societal risk due to AI in general? vs. say, the risks of affecting the quality an existing model with fine-tuning?

So, I am chalking this up to a misunderstanding due to the limitations of the platform as well as the short-hand used in the response. And I could be completely wrong :)

I wrote a C# library to do this, which is similar to other chunking approaches that are common, like the way langchain does it: https://github.com/drittich/SemanticSlicer

Given a list of separators (regexes), it goes through them in order and keeps splitting the text by them until the chunk fits within the desired size. By putting the higher level separators first (e.g., for HTML split by <h1> before <h2>), it's a pretty good proxy for maintaining context.

Which chunk size you decide on largely depends on your data, so I typically eyeball a sample of the results to determine if the splitting is satisfactory.

You can see the separators here: https://github.com/drittich/SemanticSlicer/blob/main/Semanti...

Isn't "give it a big chunk" just the same problem at a higher level? How do you handle, say, a book?

How does this work when there is a limited context window. You do some pre-chunking?

I really wanted to like it, and the visuals were fantastic. But, I had no emotional reaction to the characters and feel like the plot was just flowing along without a reason for me to care. I gave it 30 minutes.

Vector similarity searches can be done with simple SQL statements, and while the performance will not match a vector db, it's often good enough. For a db like SQLite that can run in-memory, I suspect they would be reasonably fast.

And it's not just the models. E.g., the Assistants API is not available yet in Azure, and there is no expected ship date for it. But I'm confident it's coming.