They're definitely not the only one. I've been building on relational transformers, which does prediction and classification over relational data (it handles numeric types better). It's validating to see that these small models that do prediction tasks are so useful to the community, but also stings a little that it was so hard for me to communicate how game changing they are.

You can also mine attention from image models, it's a lot of fun and very interesting.

A new transformers architecture based on relational data instead of text. Been exploring search reranking based on graph data and customer prediction.

https://github.com/relativedb/relational-transformers

https://relationaltransformers.com/

You can read more about the architecture here: https://relativedb.com/blog/relational-transformer

I did took a native c++ approach when writing a relational transformers engine (RelativeDB). My journey was pytorch -> c++ -> Triton (lang). While C++ was more performant than Triton, I couldn't afford to optimize on every gpu. I just accepted the ~15% throughput loss for my cloud service, which honestly wasn't bad for the amount of flexibility I got out of it.

But the cpp port of vllm looks great, that'd be great if you'll maintain that. I hit the same limitations with vllm.

The whole graph is walked at inference time to create the context. There is an attention feature, called 'neighbor attention', that propagate signals along primary-foreign key relationships. The task table includes the foreign keys themselves.

I created a prose tool as a claude code plugin to catch these and other AI-isms. It's not really intended for release quite yet, but it has been very useful. https://github.com/MariHQ/mari-cc

If this was a claude plugin with a hook on my dep files, I'd be in.

I've added a hook that can be installed for Claude and Codex.

I'm not sure a game needs transactions. It seems to me that entity streaming concepts are a better fit: reading from an event queue and consistently reconciling the state over time. Like the sorted top-n in the example, if it was a game and the query is generally known, it would be better to just materialize it as a fast rank, which goes beyond the IVM-style materialization they suggested.

Eeeeeh. You can never trust the client, but the server is far away, so most games have speculative execution of actions and then the server can tell them to fuck off and revert/ignore the action.

That smells a bit like a transaction to me.

Many games have obvious inventory dup exploits that this is a very reasonable solution for.

Have you taken a look at the various test harnesses for video encoding/decoding, or consider publishing a standalone one yourself? I get a bit-for-bit rewrite can have some advantage but it's hard to know if the AI implementation actually covers all of the edges cases.

I’m using FFmpeg’s own FATE suite, which is intended to poke at those edge cases, as well as JVT vectors. I’d love to get a more robust set of weird files similar to what symphonia has, though, so if anyone has any resources I’d appreciate it.

I have a legit laminar flow hood (airclean 600, 32 inch) in my home for plant tissue culture. I gotta say, I thought it was going to be easy street to sterile culture- it was not. Sterile protocol matters so much that I think it would have been just as easy to do what I needed to do in a glove box with all the extra precautions I had to take. It's honestly hard to beat an enclosed container with zero airflow.