Weekly Bookmarks

These are some things I’ve wandered across on the web this week.

🔖

Graph database-ball! Exploring the Game with the graph capabilities of LadybugDB, DuckDB and PostgreSQL

This article presents a comparison of graph capabilities in three different databases: DuckDB (v1.4.4 with duckpgq), LadybugDB (0.16.1), and PostgreSQL (19devel). We will load a large volume of records (5,635,972 rows of baseball data covering people, parks, team records, and game play-by-plays) into each database, define the entities and relationships, and write a variety of queries that take full advantage of the graph structure.

🔖 Ambient Church

Ambient Church transforms architecturally stunning spaces into immersive audio-visual environments. Our events feature pioneering artists presenting vibrant works in a context that elevates both the music and the space.

Founded in Brooklyn in 2016, we facilitate collective peak experiences through the soundscapes of modern contemplative music. With an emphasis on education and environment, we seek to illuminate an underacknowledged lineage of sonic exploration.

🔖

Language models transmit behavioural traits through hidden signals in data

Large language models (LLMs) are increasingly used to generate data to train improved models1,2,3, but it remains unclear what properties are transmitted in this model distillation4,5. Here we show that distillation can lead to subliminal learning—the transmission of behavioural traits through semantically unrelated data. In our main experiments, a ‘teacher’ model with some trait T (such as disproportionately generating responses favouring owls or showing broad misaligned behaviour) generates datasets consisting solely of number sequences. Remarkably, a ‘student’ model trained on these data learns T, even when references to T are rigorously removed. More realistically, we observe the same effect when the teacher generates math reasoning traces or code. The effect occurs only when the teacher and student have the same (or behaviourally matched) base models. To help explain this, we prove a theoretical result showing that subliminal learning arises in neural networks under broad conditions and demonstrate it in a simple multilayer perceptron (MLP) classifier. As artificial intelligence systems are increasingly trained on the outputs of one another, they may inherit properties not visible in the data. Safety evaluations may therefore need to examine not just behaviour, but the origins of models and training data and the processes used to create them.

🔖

The Archive in Art Art in the Archive

In this essay we will attempt to look at both the archive of art as well as the archive as art. When we draw a distinction between those materials that we treat as documents with a ‘factual’ historical significance (those which offer themselves in the service of scholarship), and the uses which artists make of the archive as one of the media of expression that intersect with their documentary value, we ask ourselves: which theories about the archive’s nature and function are applicable to Syrian art? What are the roles adopted by ‘the document’ and ‘the archivist’? To what extent do these roles alternate and intersect?

🔖

Cave of Forgotten Dreams

Cave of Forgotten Dreams is a 2010 3D documentary film by Werner Herzog about the Chauvet Cave in Southern France, which contains some of the oldest human-painted images yet discovered—some of them were crafted around 32,000 years ago. It consists of footage from inside the cave, as well as of the nearby Pont d’Arc natural bridge, alongside interviews with various scientists and historians. The film premiered on 13 September 2010 at the Toronto International Film Festival.

🔖

5 Reasons I Wear A Mask (and you can too)

a short zine about masking & disability justice

🔖

Starbucks ditches AI inventory system after just 9 months

Starbucks is saying goodbye to its artificial intelligence inventory management system about nine months after its debut, Reuters reported Thursday. The tool, which used computer vision to track some parts of the chain’s inventory, was announced in September as a method to simplify inventory record-keeping and prevent stockouts.

🔖 “Renaud &

Axelle Red” - Manhattan kaboul - (english subtitles)

🔖

FediRoster

FediRoster is a slightly more heavyweight alternative to David Adler’s Sociologists on Mastodon software. It is intended to function as a public list of Mastodon and other fediverse accounts, geared primarily towards academic communities, but suitable for others as well. It offers functions for following listed accounts individually or in bulk. The main novelty here is that you can add yourself to the list through an authentication process instead of all the work falling on a list maintainer. You can sign in through your Mastodon account or send a message to the list’s bot to verify your account ownership. This also means that the hosting process for new lists is a bit more involved (it’s a Python/WSGI application).

🔖

Rust for Python Programmers: Complete Training Guide

A comprehensive guide to learning Rust for developers with Python experience. This guide covers everything from basic syntax to advanced patterns, focusing on the conceptual shifts required when moving from a dynamically-typed, garbage-collected language to a statically-typed systems language with compile-time memory safety.

🔖 CS336: Language

Modeling from Scratch

Language models serve as the cornerstone of modern natural language processing (NLP) applications and open up a new paradigm of having a single general purpose system address a range of downstream tasks. As the field of artificial intelligence (AI), machine learning (ML), and NLP continues to grow, possessing a deep understanding of language models becomes essential for scientists and engineers alike. This course is designed to provide students with a comprehensive understanding of language models by walking them through the entire process of developing their own. Drawing inspiration from operating systems courses that create an entire operating system from scratch, we will lead students through every aspect of language model creation, including data collection and cleaning for pre-training, transformer model construction, model training, and evaluation before deployment.

🔖

wasteback-machine

Wasteback Machine is a JavaScript library for analysing archived web pages, measuring their size and composition to enable retrospective, quantitative web research.

🔖

No, Artificial Intelligence Is Not Conscious

The primary difference between deepfake photos and LLM conversations is that the people who generate the former are deliberately trying to fool others, and many of the people who elicit the latter from LLMs have inadvertently fooled themselves.

🔖

A Visual Guide to Gemma 4 12B

The removal of the encoders, which are typically in charge of making sense of the multimodal inputs, places the burden of making sense of all outputs on the LLM. Although the model is encoder-free, all modalities are now unified within the LLM. Instead of the model having to wait for the encoders to finish processing the audio and image inputs, the LLM can get started earlier processing the input and generating output!

In this guide, I want to showcase what it took to remove the vision and audio encoders and replace them with something much faster. The result, a 12B model that can handle audio and image inputs but without the need for encoders.

🔖

LiteRT-LM Python API

The Python API of LiteRT-LM for Linux, macOS and Windows. Features like multi-modality, tools use, and GPU and NPU acceleration are supported.

🔖

LiteRT-LM

LiteRT-LM is Google’s production-ready, high-performance, open-source inference framework for deploying Large Language Models on edge devices

🔖

Google AI Edge Gallery

AI Edge Gallery is the premier destination for running the world’s most powerful open-source Large Language Models (LLMs) on your mobile device. Experience high-performance Generative AI directly on your hardware—fully offline, private, and lightning-fast.

🔖

Introducing Gemma 4 12B: a unified, encoder-free multimodal model

Today, we are introducing Gemma 4 12B, our latest model designed to bring agentic multimodal intelligence directly to laptops. Bridging the gap between our edge-friendly E4B and our more advanced 26B Mixture of Experts (MoE), Gemma 4 12B packages powerful capabilities inside a reduced memory footprint. It is also our first mid-sized model to feature native audio inputs

🔖 Solid

State Book

Solid State Books is a full-service, Black-owned general interest bookstore with a great selection of fiction & non-fiction titles. We stock literary gifts, stationery, greeting cards & puzzles for all ages. We have a carpeted, playful children’s books area in both stores for kids & parents alike to spread out & read together. Come by for weekly children’s story hours, catch monthly book groups, author readings/signings, local interest panels, political conversations & more!

🔖

minisearch

MiniSearch is a tiny but powerful in-memory fulltext search engine written in JavaScript. It is respectful of resources, and it can comfortably run both in Node and in the browser.

🔖

Documents the Department of Justice tried to disappear.

In May 2026, the Justice Department began systematically removing material from its web sites regarding the many indictments and convictions related to the Jan. 6 attack on the U.S. Capitol. This archive reconstructs the vast bulk of those thousands of deleted records.

🔖

The Justice Department Erases History; Lawfare Restores It

Last week, the Justice Department began systematically removing material from its web sites regarding the many indictments and convictions related to the Jan. 6 attack on the U.S. Capitol.

The operation started without fanfare or formal announcement and proceeded largely unnoticed. Until, that is, journalists such as the Washington Post’s Meryl Kornfield took notice of certain press releases and other materials that had conspicuously disappeared from www.justice.gov.

“The Trump admin is quietly deleting info about the Capitol attack from the DOJ website as it prepares to give funds to J6ers,” Kornfield posted. “This week, DOJ deleted a press release about one man with an ongoing child solicitation case who came to the Capitol with bear spray.”

Then, with typical bombast, the Justice Department responded by taking issue with one particular aspect of Kornfield’s characterization. “Nothing ‘quiet’ about it,” the DOJ Rapid Response account replied. “We are proud to reverse the DOJ’s weaponization under the Biden administration. We will do everything in our power to make whole those who were persecuted for political purposes. This includes stripping DOJ’s website of partisan propaganda.”

We are not erasing history quietly, the Justice Department seemed to suggest. We are erasing history loudly and proudly.

At Lawfare, we have restored the vast bulk of what was deleted. We have also started to preemptively archive a raft of material that has not yet been deleted but probably will be, given its thematic relationship to the material that was 86ed.

🔖 Data

Center Policy Database

Data centers are the physical facilities that power cloud services, AI systems, streaming, and nearly every digital platform people use each day. As demand for artificial intelligence accelerates, data centers are becoming major sources of electricity demand and local infrastructure pressure, which means their growth affects energy systems, communities, and long-term public planning.

🔖

The State of Data Center Policy in the United States

The regulatory landscape for data centers in the United States has shifted dramatically in recent years from a period of aggressive economic incentives to a phase of intense scrutiny, restriction, and community-led resistance. To track these legislative changes, the DIGS Lab at the University of Virginia reviewed more than 700 federal, state, and local policies related to data centers. The data center policy database aims to bring transparency around zoning, permitting, and regulating data centers and their impacts on communities. This is what we found.

🔖

R.E.M. Live - 1981-11-07 Viceroy Park, Charlotte, North Carolina

Another early R.E.M. set, from the same state but a different city as the previous show. Pretty much the same library of songs, but this one’s the superior show to get - it sounds slightly nicer and doesn’t have the equipment failures of the previous show. There’s already a source on here, but that’s a different master of the same recording.

🔖

itsjunetime / tdf

A terminal-based PDF viewer.

Designed to be performant, very responsive, and work well with even very large PDFs. Built with ratatui.

🔖 ratatui

Ratatui (ˌræ.təˈtu.i) is a Rust crate for cooking up terminal user interfaces (TUIs). It provides a simple and flexible way to create text-based user interfaces in the terminal, which can be used for command-line applications, dashboards, and other interactive console programs.

🔖 Sonny Rollins

with Paul Jeffrey

Emission “Who is” de la BBC 2 Producer / Director : Dick Fontaine Produced by Allan King Associates Canada LTD. 1968