Hacker News

Article URL: https://tracker.archiveteam.org/

Comments URL: https://news.ycombinator.com/item?id=49605182

Points: 1

# Comments: 0

This tool started as a research for me trying to answer the question "Does AI agents drift from your architecture conventions or if you need enforcement instead?". So I started out building a benchmark to test that premise and on the scope I tested, the agents did not drift and I had to stop that research line. So to reuse what I have learnt from that research, I started out trying to catch engineer's regressions through review and CI and create those conventions codebases already follows with the evidence instead of leaving them in someone's head.

Before I started building this tool, I also mostly write the rules by hand and linter's documentations for every linter I wanted to use. But after my research, the tool automates the process and I'm excited to see how it can help other engineers too. I ran the full package workflow (scan, recommend) across 92,861 real public TypeScript repositories on github and it completed on every one without a single crash and the reversible install/uninstall round-trip held. I will appreciates any feedback you have on the rules it infers for your repo and how it can help you enforce them or not.

The study is public here https://github.com/Tommkruix/agentrulebench and Archprint is public here https://github.com/Tommkruix/archprint also.

Comments URL: https://news.ycombinator.com/item?id=49605147

Points: 1

# Comments: 0

Article URL: https://albertjoseph0.github.io/nyc-maptap/

Comments URL: https://news.ycombinator.com/item?id=49605122

Points: 2

# Comments: 1

People use embedding models all the time for rag/semantic retrieval. However, when a newer, more desireable model comes out, there is an expensive (both in time and computational) cost of re-embedding every document in the database.

However, I figured out an interesting way to forgo that upfront embedding cost.

algo:

old model/index -> retrieve top-K docs -> score those docs with the new model -> cache/materialize the new embeddings

so instead of rebuilding the entire vector store upfront, the old index keeps getting retrieved from, while the new model reranks those candidates.

This works surprisingly well for some model pairs, (i tested 63 source-> target migrations on h100s, on upto 1M documents).

For example, on a 1M document Natural Questions dataset,

native Qwen3-Embedding-8B: 0.6812 nDCG@10 Qwen3-4B -> Qwen3-8B, K=50: 0.6816 Qwen3-0.6B -> Qwen3-8B, K=50: 0.6638 MiniLM -> Qwen3-8B, K=50: 0.6486

(the hard part is determining k, I held the k constant above to give some sense of migratability).

You can install it with pip

pip install embedflow

and the code is on github

https://github.com/arnsri33/embedflow

Comments URL: https://news.ycombinator.com/item?id=49605110

Points: 1

# Comments: 0

Article URL: https://medium.com/freedomofthought/defeating-antisemitism-through-zionism-6c648c34c408

Comments URL: https://news.ycombinator.com/item?id=49605086

Points: 1

# Comments: 0

I’ve been working on Caveat, a self-hosted publishing tool for people who want control over their newsletter’s design, subscriber data, and infrastructure.

It includes a rich-text editor, autosave, reusable content blocks, publication designs, typography and colour controls, subscriber management, exports, and Resend email delivery. It includes both the editor and public website.

Repo: https://github.com/CaveatJS/site Examples/docs: https://caveat-home.vercel.app/

I built the first version over a few hours across three sessions with GPT-6 Astra. I chose the features, tested the workflows, and refined the design and implementation.

The sessions processed roughly 43.6M tokens, about 97% cached input.

Caveat is still early: paid subscriptions, scheduling, and multi-author workflows are not implemented yet.

If you publish a newsletter, what part of your workflow would prevent you from using it today?

Comments URL: https://news.ycombinator.com/item?id=49605072

Points: 1

# Comments: 0

Article URL: https://www.scmp.com/news/china/diplomacy/article/3366477/canada-and-china-thaw-continues-militaries-hold-first-talks-8-years-amid-us-pressure

Comments URL: https://news.ycombinator.com/item?id=49605046

Points: 6

# Comments: 2

Article URL: https://en.wikipedia.org/wiki/Thames_Tunnel

Comments URL: https://news.ycombinator.com/item?id=49605036

Points: 1

# Comments: 0

Article URL: https://www.passkeycentral.org/design-guidelines/optional-patterns/create-passkey-after-account-recovery-due-to-forgotten-password

Comments URL: https://news.ycombinator.com/item?id=49605024

Points: 1

# Comments: 0

Article URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC11585964/

Comments URL: https://news.ycombinator.com/item?id=49605011

Points: 2

# Comments: 0

AI pink doesn’t Columbus get eleven this, yellow do beach you? Do more advanced model understand the above or is it only for humans?

Comments URL: https://news.ycombinator.com/item?id=49604630

Points: 1

# Comments: 0

Article URL: https://github.com/agentdesktop-dev/agentdesktop

Comments URL: https://news.ycombinator.com/item?id=49604626

Points: 1

# Comments: 1

Article URL: https://blog.elevarq.com/elevarq-1-0

Comments URL: https://news.ycombinator.com/item?id=49604624

Points: 1

# Comments: 0

Article URL: https://dol.praxostudio.com/

Comments URL: https://news.ycombinator.com/item?id=49604594

Points: 1

# Comments: 0

This is one of those projects that's been sitting in the back of my mind for years. I made a ThreeJS UI for browsing (read: re-living) the ticket stubs I've been collecting since I was 9 years old.

I was going to create a whole form so that people could share their own collections but landed on just opening up a version you can use to DIY with your agent.

That part is here: https://github.com/billybillymc/stubs

Enjoy! Notes and feedback welcome and appreciated.

Comments URL: https://news.ycombinator.com/item?id=49604590

Points: 1

# Comments: 0

The origin of this project was me, as a casual farmers' market attendee, grossly misremembering the start date of early girl tomatoes. Ironically, I still haven't achieved that goal, because I couldn't find a suitable source with varietal level granularity online. I'm half dreading a "Why didn't you check X?" comment and discover there's some perfect local varietal data source, but that would really be the best case scenario.

The poster is customizable per crop, and the pictures reposition heuristically. There's only two regions, SF and NY, with very different source data for each. See https://vertumnus.fyi/about/sources for methodology.

A static poster is not the greatest form factor for a schedule with annual variation, and this whole project was a non serious side exercise. See https://vertumnus.fyi/about for more.

The project was coded in a single 731 prompt Claude Opus 5 session over 22 days within a 46 day period, from July 20th to September 2nd, 2026. I've published the entire transcript, searchable and organized by chapter, at https://vertumnus.fyi/about/transcript. A write up, with the stuff that Claude was good at and its repeated code quality shortcomings, is at https://vertumnus.fyi/about/ai.

Comments URL: https://news.ycombinator.com/item?id=49604581

Points: 1

# Comments: 0

Article URL: https://huggingface.co/gdiamos/amx-reasoning-v1-instruct/blob/main/paper.pdf

Comments URL: https://news.ycombinator.com/item?id=49604579

Points: 2

# Comments: 1

Article URL: https://huggingface.co/spaces/nanotron/ultrascale-playbook

Comments URL: https://news.ycombinator.com/item?id=49604557

Points: 1

# Comments: 0

Article URL: https://www.latent.space/p/aeo

Comments URL: https://news.ycombinator.com/item?id=49604551

Points: 1

# Comments: 0

Article URL: https://lileks.com/motels/index.html

Comments URL: https://news.ycombinator.com/item?id=49604550

Points: 1

# Comments: 1