Building an email-to-calendar-feed service for all the mails from the multitude of services and attachments that I get related to my kindergartener.

My read is that no customers will leave since they are much more interested in news coverage -- and this helps the AP focus more on news.

This is a tangent, but I wonder if they feel that they are just creating LLM training data and that few readers (even of Sunday papers) will actually read their reviews.

The base model is Qwen2.5-VL-3B and the announcement says a limitation is "Model can suffer from hallucination"

Seems a bit scary that the "source" text from the pdfs could actually be hallucinated.

Given that input is image and not raw pdf, its not completely unexpected

I wish there were some explain plans in either post, since I don't get what's going on.

If the query uses the index, then the on the fly tsvector rechecks are only on the matches and the benchmark queries have LIMIT 10, so few rechecks right?

Edit: yes but the query predicates have conditions on 2 gin indexes, so I guess the planner chooses to recheck all the matches for one index first even though it could avoid lots of work by rechecking row-wise

Is pre-training in FP8 new?

Also, 10M input token context is insane!

EDIT: https://huggingface.co/meta-llama/Llama-3.1-405B is BF16 so yes, it seems training in FP8 is new.

Deepseek v3 was FP8

S3 Tables is designed for storing and optimizing tabular data in S3 using Apache Iceberg, offering features like automatic optimization and fast query performance. SimpleDB is a NoSQL database service focused on providing simple indexing and querying capabilities without requiring a schema.

This is very cool. Kuzu has a ton of great blog content on all the ways they make Kuzu light and fast. WebLMM (or in the future chrome.ai.* etc) + embedded graph could make for some great UXes

At one time I thought I read that there was a project to embed Kuzu into DuckDB, but bringing a vector store natively into kuzu sounds even better.

Great point! Several years ago there was a project GRainDB, which along with GraphflowDB (a purely in-memory graph database) formed the ideas of what is now Kuzu :)

https://graindb.github.io/ https://github.com/graphflow/graphflow-columnar-techniques

500 errors for me

Doesn't compression make any minification gains negligible?

Depends on what you’re serving up. Blog? Yes. Video game? No.