Matt

agi is here but it got an email job

[Matt @Matthewagi]

Personalized Arxiv feed

I built a system that allows you to create a personalized feed or search over recent listings. I use a two tower architecture with a preference model and a re-ranker.

https://t.co/92kDbSIHZu

[Matt @Matthewagi]

Got a nice kick in the butt yesterday to do something I've been putting off.

Re-ranker for Arxiv articles is trained and ~industry standard

[Matt @Matthewagi]

@mattparlmer It's amazing how this skill generalizes across great companies

[Matt @Matthewagi]

@mickeyxfriedman I tend to agree. I've churned in the past from similar when I've found the work/community empty so to speak. But I don't think the community needs to be what brings satisfaction in the work

[Matt @Matthewagi]

@simonsarris This makes it feel like you should be "annoying" and make the editing paid

[Matt @Matthewagi]

@mickeyxfriedman Does it need to be satisfying?

[Matt @Matthewagi]

@callanable this is even better. Where did he get them?

[Matt @Matthewagi]

@simonsarris Unfortunately the friction is the reason to upgrade. In these the core functionality is the loss leader. Is editing the core functionality of the app?

[Matt @Matthewagi]

What's underappreciated in deepinfer is how coupled Deepseeks Arch is to its success. You don't get that throughput without the prefill effects

[Matt @Matthewagi]

@leothecurious There's not a big monetary incentive yet

[Matt @Matthewagi]

@deepfates ayyyy lmao

[Matt @Matthewagi]

@multimodalart This is a good step. I also like how your bench has moved towards prioritizing the private portion

[Matt @Matthewagi]

@multimodalart I may need to play with this later!

[Matt @Matthewagi]

@seconds_0 lmao there's one for everything

[Matt @Matthewagi]

Most of what they tested was almost certainly in the lexical regime (which is basically BM25 with data governed weighting rules).

True Semantic search has never been tried

[Zed @zeddotdev]

We tested 25 approaches to MCP tool search, including decision models, LLMs, and embeddings. Our benchmark runs 150 search queries against 536 tools.

We settled on weighted BM25 with a few normalization rules. Fast, local, no model or API calls.

[Matt @Matthewagi]

@dejavucoder Brutal results (great work!)

[Matt @Matthewagi]

Jev's pricing matches really well with a serving cost of 0.01255 (or half) which puts margin at 50 - 75% assuming 60% operational load across GPUs.

For opensource to be competitive it needs to hit < $0.01/mtok. Luckily this is easy

[Matt @Matthewagi]

@voooooogel this reads clear!

[Matt @Matthewagi]

@hopes_revenge he does it again

[Matt @Matthewagi]

@shakoistsLog this is good

[Matt @Matthewagi]

@Duderichy