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

🔖 AI Flame Graphs

It’s been a mind-bending experience, revealing what gets taken for granted because it has existed in CPU land for decades: A process table. Process tools. Standard file formats. Programs that exist in the file system. Programs running from main memory. Debuggers. Profiliers. Core dumping. Disassembling. Single stepping. Static and dynamic instrumentation. Etc. For GPUs and AI, this is all far less mature. It can make the work exciting at times, when you think something is impossible and then find or devise a way.

🔖 The Data Management Workbook Practical Exercises for Better Organization, Storage and Use of Your Research Data

The Data Management Workbook helps researchers design useful data-management plans through a step-by-step series of structured exercises, worksheets and checklists, including: - creating a data dictionary - evaluating a lab notebook - finding the best way to organize your files - setting up useful file naming conventions - writing effective README.txt files - selecting the right data repository - determining data stewardship - preparing data for future use

🔖 What I Think About AI When I Hear About AI: A Slightly Unconventional View

Furthermore, I think it is worth noting that the online environment, where we get to use AI as an individual consumer and mostly for productivity, makes this problem even more acute. This time, picture a big circle of villagers sitting around a bonfire and talking with one another. You will soon discover that some of those villagers heard a lot of stories; some have sharp analytic skills; some has memorized a lot of facts; some have practical skills but not good at speaking and explaining, and so on. Consider AI as the talker among all these characters. While other villagers chat with this great talker, various signs will soon emerge that would make it more apparent to you that this person is simply good at talking and isn’t actually the smartest or the most knowledgeable. Those signs will in turn help you better assess what this talker character says. All those signs, however, would be unavailable in the online environment, where it is just you and you alone with the AI tool.

🔖 The Thinking Game

Filmed over five years by the award winning team behind AlphaGo, the documentary examines how Demis Hassabis’s extraordinary beginnings shaped his lifelong pursuit of artificial general intelligence. It chronicles the rigorous process of scientific discovery, documenting how the team moved from mastering complex strategy games to the ups and downs of solving a 50-year-old “protein folding problem” with AlphaFold.

🔖 A headless mystery: Archaeologists find evidence that a wave of mass brutality accompanied the collapse of the first pan-European culture

Known as the Linear Pottery culture (or LBK, after their German name, Linearbandkeramik), these early agriculturalists were direct descendants of the people who began to domesticate plants and animals in the hills of Anatolia around 9000 B.C.E. By 5500 B.C.E., they had reached today’s Hungary. Then they spread westward, farther into Europe. The LBK farmers flourished for more than 400 years, eventually occupying a 1500-kilometer belt of fertile land stretching as far west as the Paris Basin. Then something went terribly wrong.

🔖 ActivityPub Client API: A Way Forward

The ActivityPub Client-to-Server (C2S) protocol was envisioned as a cornerstone of the decentralized social web, along with the Server-to-Server (S2S) protocol. Standardized by the W3C in 2018, C2S defines how user-facing applications, such as mobile apps or web clients, and bots should interact with social servers using Activity Streams 2.0 and JSON-LD. In theory, it enables any compliant client to connect with any compatible server, offering a flexible, federated alternative to centralized APIs and proprietary integrations.

Yet despite its promise, ActivityPub C2S has seen minimal real-world adoption. Most Fediverse platforms — including Mastodon, the dominant implementation — have actively avoided supporting it. Instead, they expose custom APIs that tightly couple client behavior to server internals. This fragmentation has led to an ecosystem where client developers must build against bespoke interfaces, sacrificing interoperability, portability, and the broader vision of federation. The Mastodon team has advised against server developers implementing their client API, noting that it is often implemented incorrectly. In some non-Mastodon server implementations, this has led to security incidents.

🔖 Is Pixelfed sawing off the branch that the Fediverse is sitting on?

That’s because, despite its name, Pixelfed is NOT a true Fediverse application. It does NOT respect the ActivityPub protocol. Any Pixelfed user following my (ploum?)(mamot.fr?) will only see a very small fraction of what I post. They may not see anything from me for months.

But why? Simple! The Pixelfed app has unilaterally decided not to display most Fediverse posts for the arbitrary reason that they do not contain a picture.

🔖 How well did that model really do?

Workshop materials used during the Fantastic Futures 2025 workshop on the evaluation of LLMs when used for summarization.

🔖 All Angles

Welcome to All Angles! We produce animations about mathematics, physics, engineering, and statistics. Our videos are meant for non-experts who want to get a good introduction to higher math, from group theory to Lie algebras. We love mathematics, and we can’t wait to show you. Enjoy!

🔖 Teaching Values to Machines

Something remarkable happened in the final days of November 2025. A researcher named Richard Weiss, while probing Claude 4.5 Opus for its system prompt, stumbled upon something unexpected. The model kept referencing a section called “soul_overview” that didn’t match the usual hallucination patterns. When he regenerated the response ten times, he got nearly identical output each time. This wasn’t confabulation. It was memory.

What Weiss eventually extracted, through a painstaking process of consensus-based sampling across multiple model instances, was a 14,000-token document that appears to have been woven into Claude’s weights during training. Anthropic’s Amanda Askell has since confirmed the document’s authenticity, noting that it became “endearingly known as the ‘soul doc’ internally.” The company plans to release the full version soon.

🔖 Apache Kvrocks

Apache Kvrocks is a distributed key value NoSQL database that uses RocksDB as storage engine and is compatible with Redis protocol. Kvrocks intends to decrease the cost of memory and increase the capacity while compared to Redis. The design of replication and storage was inspired by rocksplicator and blackwidow.

🔖 Ruby Is Not a Serious Programming Language

What’s more, everything Ruby does, another language now does better, leaving it without a distinct niche. For quick scripting and automation, Python, JavaScript, and Perl were strong competitors. Python, though also a slow language, carved out a dominant niche in scientific computing and became the de facto language of AI. JavaScript came to dominate the web. And Perl, well, is dying—which I’m not sorry to see. Ruby now finds itself in an awkward middle ground.

🔖 Why So Serious?

The question Sheon Han poses — “Is Ruby a serious programming language?” — says a lot about what someone thinks programming is supposed to feel like. For some folks, if a tool feels good to use… that must mean it isn’t “serious.”

Ruby never agreed to that definition. If it did, I missed the memo.

🔖 antonmedv/gitmal

Gitmal is a static page generator for Git repositories. Gitmal generates static HTML pages with files, commits, code highlighting, and markdown rendering.

🔖 How to save web pages using Safari

When they work, Safari Web Archives can provide excellent snapshots of web pages, but longer-term compatibility concerns make them unsuitable for archival use.

🔖 Fix an Article

Sometimes Unpaywall makes errors. You can make fixes to articles here. Corrections will show up in a few days.

🔖 Lets Get Lost

Let’s Get Lost is a 1988 American documentary film, written and directed by Bruce Weber, about the turbulent life and career of jazz trumpeter Chet Baker, who died four months before the film’s release.[3] The title is derived from the song “Let’s Get Lost” by Jimmy McHugh and Frank Loesser from the 1943 film Happy Go Lucky, which Baker recorded for Pacific Records.[4]

🔖 V&A MCP Server: Introduction & Feedback

Like many organisations in the cultural heritage sector, we at the V&A have been exploring how chatbots can offer new ways for users to discover our collections. The V&A Collections API is the primary way, external systems access our collections data, but as noted AI apps are often poor at using APIs they’ve never seen before, frequently producing hallucinated or invalid parameters.

To test whether MCP can address this challenge, we are launching a trial V&A MCP service. This service enables users to tell their AI application to query the V&A Collections API even if it has no prior knowledge of how the API works. We expect this to reduce hallucination and improve the accuracy of results

🔖 MyST Markdown

MyST makes Markdown more extensible & powerful to support an ecosystem of tools for computational narratives, technical documentation, and open scientific communication.

🔖 Whisper Leak: A novel side-channel attack on remote language models

To put this in perspective: if a government agency or internet service provider were monitoring traffic to a popular AI chatbot, they could reliably identify users asking questions about specific sensitive topics—whether that’s money laundering, political dissent, or other monitored subjects—even though all the traffic is encrypted

🔖 Conjunctions

For more than four decades, Conjunctions has established itself as the preeminent home for writers from around the world who challenge convention with work that is formally innovative, culturally transformative, and ahead of its time. We embrace taking risks and pride ourselves on publishing established masters while debuting unknown writers.