It sounds interesting, but what are the actual advantages compared to a standard HTTP client? At what data volume does performance parity begin?
32 karma · joined September 2, 2026
- Right now, I can only accept or reject the suggested response. Adding or editing it requires a lot of steps.
- It would be great to have a neat autocomplete feature. That's because prompts always need to be detailed, and that means typing a lot of text that’s essentially already in the context (and that's exactly what we should be using).
- It would be better to have a menu with action options and not have to type anything at all. Previously, Claude did a good job displaying a UI with different options (but for some reason, it stopped working for me specifically - probably because I specify everything in as much detail as possible).
- Again, providing details means the user explicitly agrees to the conditions, and the prompt is more effective on every level—plus, it caches better. Otherwise, users usually write something like: "Do everything, but switch the left one to the right and flip the one on the far end twice" (in response to three blocks of lists with task IDs) - even other people wouldn’t understand that.
NPU performance will be twice as high, while power consumption will be four times lower. No additional fan noise (if you know what I mean).
I'll wait another month until the first phase of the =battle royale= among models of this kind wraps up, put together a solution for the NPU/iGPU, and post it on HF.
But honestly, I'd love for the entire ceiling to be one controllable light source. That would be perfect.
When comparing Bun and Node.js for my tasks, I keep finding more advantages in Bun’s optimizations. I'll handle GPU processes through BFFI. Node.js is also prone to unstable memory behavior with native processes, so it doesn’t make sense to use it as an extra layer.
For those looking for stability, Node.js is definitely the better choice.
For example, Linux — how fascist, communist, or monarchist is it? I'd also be interested to know which NVIDIA drivers are more supportive of the LGBT community—the open-source ones or the proprietary ones?
I had never even thought about this until your comment.
LLM memory.. let's say, architecturally, its suitable for storing and processing queries - is questionable, but okay. But what real-world use case would generate such queries? It seems to me that the context for formulating and processing the results would be orders of magnitude greater than simply writing down what's needed in plain text. If my notebook took the form of a temporal multiverse of entities, I'd lose the ability to use it properly.
But maybe I'm just too old to quickly get the hang of every new type of Pokémon.
P.S. I'm using the built-in SQLite to store a graph with up to 10 million edges in an MCP app I'm developing for working with code.
At first, I tried all those popular vector databases and was disappointed with their performance. In the end, the best and fastest solution turned out to be building a multi-database system on SQLite, compiled with everything related to multi-client operations removed. Only exclusive mode was left. Everything is as binary as possible. The index is completely separate — an IVF with pre-training — and is built on the GPU (250K vectors are built, processed, and saved in 4 seconds). Right now, my biggest problem is frequent data changes, and I need to implement optimizations to reduce recalculations.
So far, I haven’t seen any vector database implementations that are heading in the right direction. Maybe only Lancedb looks promising, but it’s too heavy for my needs.