RexZhang

Building @knowra_app: agent-friendly arXiv search. I share notes on what new papers actually tested.

[RexZhang @Reeexopia]

dsh-plugin-knowra is out. Search arXiv by topic, author or exact dates from DeepSeek Harness, then let Scout read the full paper — method, evidence, limitations — so your agent can tell you which papers are worth a close read.

https://t.co/GlKfj5H7C3

[RexZhang @Reeexopia]

@Hi_Mrinal oh my, white page + lights off. your eyes must be screaming. btw how do u find papers?

[RexZhang @Reeexopia]

@mert_ozcetin_ @X Still at $0 made online, me either! lol

[RexZhang @Reeexopia]

@Ram5721 Hi, me Rex, building Knowra. Scout reads the full paper so your agent can tell a strong paper from a good abstract. It all sits on a daily index of every paper arXiv announces.

[RexZhang @Reeexopia]

@rubana_kh F) no users for a solo dev

[RexZhang @Reeexopia]

@itz_komal01 both if possible, u can feel the difference

[RexZhang @Reeexopia]

@benabmathew @sreenidhb Hi, lets connect!

[RexZhang @Reeexopia]

@itz_komal01 I need em both!

[RexZhang @Reeexopia]

@bycampelo Thanks! Live in beta: https://t.co/YpiZFtCyHb — Scout reads the full paper, not the abstract, so your agent can tell a strong paper from a good abstract.

[RexZhang @Reeexopia]

Building a weekly arXiv digest for Knowra users. It's a great way to catch what's about to blow up.

Launching soon. Would love early feedback.

[RexZhang @Reeexopia]

@MiaAI_lab @NVIDIAAI crazy.

[RexZhang @Reeexopia]

9-to-5, then 5-to-9. So tired. Especially when a whole evening goes by and I realize I made no real progress. Exhausting. How do you all deal with this?

[RexZhang @Reeexopia]

@MiaAI_lab hi, ive asked u before about ds-v4.1-flash, u said fp4 version not suit for 2x spark, and whats ur recommended setup for like a self-hosted small team? is 4x enough for around 4 ppl daily use?

[RexZhang @Reeexopia]

@RaziaAliani For paper-heavy work: don’t make Claude read whole PDFs to decide what’s worth reading. A full-text brief is ~1k tokens vs 9–21k for the paper. I’m building an MCP tool for this on arXiv, and some users asked for PubMed. Would love your take: does it make sense as a main source?

[RexZhang @Reeexopia]

@JulienBlanchon well, but it's still expensive tho. I created tools for a quick review, especially for agents, that reduce token usage, and only the rest needs more attention

[RexZhang @Reeexopia]

Paper: https://t.co/1sScFh4f44

Authors are upfront: each refresh policy was trained once, and detector timing isn't yet separated from a fixed schedule. Seed replication is on their list.

Found and read via Knowra, which I built (@knowra_app).

[RexZhang @Reeexopia]

Self-improvement work asks what to train on next. A-RSI asks when current tasks stop teaching.

It refreshes the task pool when fewer prompts yield nonzero group advantage. No val scores or human labels: 57.77 → 61.05 on 11 benchmarks.

Picking by difficulty didn't beat random.