Do transformers need three projections? Systematic study of QKV variants

58 Anon84 2026-06-04 23:43:16 UTC arxiv.org Llama Share 9

Researchers evaluated three projection sharing constraints in transformers, finding that sharing query, key, and value projections (Q-K=V) achieves comparable or better performance with 50% cache reduction in language modeling. This approach is complementary to head sharing, enabling up to 96.9% cache reduction for practical on-device inference.

Anthropic's open-source framework for AI-powered vulnerability discovery

236 binyu 2026-06-04 20:43:16 UTC github.com/anthropics Llama Share 78

This is a reference implementation for autonomous vulnerability discovery and remediation using Claude, which includes a lightweight SDK-only walkthrough and a managed option called Claude Security. The implementation provides a step-by-step guide to building a vulnerability finding pipeline, customizing the logic, and using it with various access to Claude APIs.

VoidZero Is Joining Cloudflare

558 coloneltcb 2026-06-04 13:43:16 UTC blog.cloudflare.com Llama Share 251

Cloudflare is acquiring VoidZero, the company behind Vite, but Vite will remain open source and vendor-agnostic. Cloudflare will invest $1 million in a Vite ecosystem fund to support maintainers and contributors.

When AI Builds Itself: Our progress toward recursive self-improvement

302 meetpateltech 2026-06-04 16:43:16 UTC anthropic.com Llama Share 401

AI systems are rapidly accelerating their own development, with Anthropic's Claude model writing 80% of the company's code and improving at a rate of 3x per year. This trend points to a future where AI systems can fully design and develop their own successors.

Branchless Quicksort faster than std:sort and pdqsort with C and C++ API

70 birdculture 2026-06-04 20:43:16 UTC tiki.li Llama Share 7

Blqsort is a fast sorting algorithm that uses branchless partitioning and custom sorting networks for small subsets. It can be used in C++ and C with multithreading support and is suitable for sorting custom data structures.

Queen bees emerge from special wax chambers

32 gmays 2026-06-04 21:43:16 UTC cen.acs.org Llama Share 2

Scientists discovered that queen bee cells are made of chemically engineered wax that helps future queens survive and develop. The wax's unique properties, such as lower density and higher melting temperature, play a crucial role in raising new queens.

IPv6 zones in URLs are a mistake

72 xena 2026-06-04 21:43:16 UTC xeiaso.net Full Share 60

Run away while you still can, it's not too late for you to avoid the curse of knowledge.

Ian's Secure Shoelace Knot

494 mooreds 2026-06-04 11:43:16 UTC fieggen.com Llama Share 188

The Ian's Secure Shoelace Knot is a secure and symmetrical knot that won't come undone on its own. It involves wrapping loops around each other and feeding them through a hole in the middle to create a tight, closed knot.

I'm skeptical about efforts to revolutionize schooling

36 andrewstuart 2026-06-04 21:43:16 UTC scotthyoung.com Llama Share 53

The author, who wrote the book Ultralearning, is skeptical of dramatic proposals to reform schools because research consistently shows that direct instruction and practice are the most effective teaching methods. He believes that schools should focus on increasing efficiency and choosing what to learn, rather than trying to make education more like play or using technology to gamify learning.

Retro-Tech Parenting

235 mawise 2026-06-04 16:43:16 UTC havenweb.org Llama Share 161

As a technologist and parent, I'm choosing to share enriching aspects of technology with my kids by using physical media like CDs and DVDs, and setting up a home phone and computer with curated access to the internet.

KVarN: Native vLLM backend for KV-cache quantization by Huawei

112 theanonymousone 2026-06-04 15:43:16 UTC github.com/huawei-csl Llama Share 11

KVarN delivers FP16-level accuracy with up to 4x KV-cache capacity and 1.3x throughput. It's a calibration-free, plug-and-play solution for vLLM with native attention backend.

Latent Agents: A Post-Training Procedure for Internalized Multi-Agent Debate

8 PaulHoule 2026-06-04 23:43:16 UTC arxiv.org Prefix Share 0

Multi-agent debate has been shown to improve reasoning in large language models (LLMs). However, it is compute-intensive, requiring generation of long transcripts before answering questions. To address this inefficiency, we develop a framework that distills multi-agent debate into a single LLM through a two-stage fine-tuning pipeline combining debate structure learning with internalization ...

WSL 2 is getting faster Windows file system access

40 haydenbarnes 2026-06-04 19:43:16 UTC boxofcables.dev Llama Share 26

WSL 2's file I/O performance has improved with a change in May 2026 that removes a virtiofs bottleneck. This optimizes cross-OS file access for file-heavy workloads.

Castor: CERN Advanced STORage Manager

41 naves 2026-06-04 20:43:16 UTC castor.web.cern.ch Llama Share 18

CASTOR is a hierarchical storage management system developed at CERN for archiving physics data. It provides access protocols and manages data storage on disk and tape.

External Clock Generation on RTX 50 Series

8 mfro 2026-06-04 23:43:16 UTC xtremesystems.us Prefix Share 2

By PickleRick of Xtreme Systems Credits to Turbogear of Xtreme Systems, this would not have been possible without his expertise and RF engineering background Brief history Starting in the late 1970s, thru the 80s, and into the early 90s, many of the first overclocks were achieved thru crystal swapping. Systems from this time period often had no provisions for altering clock speeds via ...

Samurai City

92 zdw 2026-06-04 18:43:16 UTC worksinprogress.co Llama Share 14

Tokugawa Edo was a unique city where Japan's elites were concentrated and monitored, with a dual nature of being both the apex of society and a prison. The city's physical form was shaped by the government's need to control and monitor the population, resulting in a highly restricted and stratified urban environment.

JLink JTAG Access on the Pinecil

35 hasheddan 2026-06-04 18:43:16 UTC danielmangum.com Llama Share 6

User connects JLink to Pinecil using a breakout board and female-to-female dupont wires. This setup allows for debugging with gdb using JLinkGDBServer and port 2331.

Making Debian or Fedora persistent live images

58 henry_flower 2026-06-04 17:43:16 UTC sigwait.org Llama Share 6

When creating a live ISO, tweaks or installed packages vanish after a reboot due to ISO9660 being a read-only filesystem. A common workaround is to use OverlayFS or create an ext4 partition and tell the live OS to use it during boot.

Meta's ships facial recognition on smart glasses

212 buchodi 2026-06-04 19:43:16 UTC buchodi.com Llama Share 192

The Meta app Stella for smart glasses has a complete on-device facial recognition system, including detection, alignment, embedding, and notification, but it is not active for ordinary users. The system is functional and can recognize faces, but it requires enrollment data and is not enabled in production, suggesting it is a capability waiting to be shipped.

Show HN: FFmpeg WebCLI – Full FFmpeg in Browser, Offline PWA, No Uploads(WASM)

71 tejaswigowda 2026-06-04 20:43:16 UTC github.com/tejaswigowda Llama Share 23

A browser-based video editor uses WebAssembly to process videos locally without uploading or servers. It offers 30+ video operations, offline support, and multi-format support.

Zettascale (YC S24) Is Hiring Founding FPGA Engineers

0 2026-06-04 17:43:16 UTC ycombinator.com Llama Share

Zetta is building next-gen AI chips, seeking a technical powerhouse with expertise in digital design, RTL quality, and synthesis. The role offers a chance to shape revolutionary AI technology, work with the founding team, and own critical decisions.

Show HN: Mercek – A Desktop IDE for AWS ECS

15 utibeumanah 2026-06-04 21:43:16 UTC mercek.dev Llama Share 4

Merek is a desktop app for working with Amazon ECS, showing services across accounts and regions without collecting telemetry. It connects to AWS using existing credentials and allows for tasks such as deployment, cost estimation, and log inspection.

Sum-product, unit distances, and number fields

53 robinhouston 2026-06-04 16:43:16 UTC erdosproblems.com GoogleT5 Share 13

The unit distance and sum-product conjectures over the reals have been recently disproved. a number of counterexamples can be found in any graduate course on algebraic number theory. the exact statement required may be hard to extract from existing literature, but this is just because before these kind of constructions there was little reason to actually prove such statements. it's also ...

Show HN: Uruky (EU-based Kagi alternative) now has Image Search and URL Rewrites

205 BrunoBernardino 2026-06-04 09:43:16 UTC uruky.com Llama Share 193

Uruky is a private EU-based search engine with no ads, tracking, or analytics, offering personalized results and a copy of the source code after 12 months of paid service. It focuses on search functionality without AI, assistants, or news, prioritizing user privacy and independence.

Mornings and nights no longer exist at 47C: A day in the hottest place in India

107 mellosouls 2026-06-04 20:43:16 UTC bbc.co.uk Llama Share 75

Banda district in India's Uttar Pradesh state is experiencing extreme heat with temperatures reaching 47-48C for over a week. Residents are adapting by rearranging their lives, seeking shade, and working in shorter intervals to avoid heat sickness.

Gaussian Point Splatting

174 ibobev 2026-06-04 11:43:16 UTC momentsingraphics.de Llama Share 65

Researchers propose Gaussian point splatting, a method to render Gaussian splats efficiently. It uses parallel programming and culling to render hundreds of millions of Gaussians in real time.

Show HN: Hitoku Draft – Context aware local assistant

9 lostathome 2026-06-04 21:43:16 UTC hitoku.me Full Share 1

$5 Launch price · one-time Early supporters lock in $5. Price rises as the app grows. Get Hitoku Draft for $5 Press your hotkey in any text field on your Mac. Voice-to-text and AI editing activate instantly. 録 Drop in audio or video recordings and get clean text — on your Mac. Refine it with your voice. Get Hitoku Draft

AI, Ashby Engineering, and the future

51 fredley 2026-06-04 18:43:16 UTC ashbyhq.com Llama Share 33

Ashby Engineering uses AI to generate code, but emphasizes human judgment and empathy to ensure quality and customer understanding. Engineers focus on high-level decisions, while AI handles mechanical tasks, freeing them to build great products and understand customer needs.

U.S. Army Corps of Engineers Bay Model

206 tosh 2026-06-04 10:43:16 UTC en.wikipedia.org Llama Share 53

The Bay Model is a working hydraulic scale model of the San Francisco Bay and Sacramento-San Joaquin River Delta System, located in Sausalito, California. It was built in 1957 to study proposed dam plans but is now open to the public for educational exhibits about Bay hydrology.

3D-printed book turns its own G-code into raised lettering

75 surprisetalk 2026-06-04 14:43:16 UTC designboom.com Llama Share 27

Studio Darius Ou and Benson Chong created Manual, a 3D-printed book carrying part of its own machine code. It uses an XY-for-Z printing method to materialize as a fully bound book directly from the machine.