Topical
This chart from Our World In Data tracks not just the different median age by world region, but also how it changed between 2000 and 2023.
USAFacts illustrate how the last two midterm elections (2018 and 2022) had higher turnout than most elections in the previous four decades.
This is an analysis of what the data shows about the flydubai flght 1073 jet's plunge.
According to analysis by The European Correspondent, young men are shifting rightward while young women move left.
Tools & Tutorials
"523 lessons. 20 phases. Every algorithm built from raw math before a single framework gets imported."
This article presents the experience of training a GPT model on a million 15-puzzle solutions, realising this solves 99.9% of random scrambled boards but fails on the easiest puzzles, even an already-solved one. A fairly simple, strong example of the fallacy of thinking AI as intelligence rather than as advanced statistics, at least for this class of problems.
"Upstate NY and New England often use the foliage as a tourist attraction so every so often you see reports of whether the foliage is at "peak" or not along with the weather reports. There's regional dashboards that show the foliage status for large areas of the country. So there must be data collection going on here – let's dig into that this week!"
"My machine learning models are written in Python, but most of the work around them happens in R: preparing and editing the input files, previewing the outputs, post-processing the results, generating figures…"
This is an interactive course teaching signal-processing concepts (invertibility, causality, stability, Fourier series, sampling, aliasing, Laplace/z-transforms, etc.) through real Wikipedia traffic data labs.
Dataviz, Data Analysis, & Interactive
Based on Global Data Tracker, this chart tracks net forest-area change (in million hectares) from 2000-2023 for the ten countries with the largest forests as of 2000, using FAO/World Bank data.
![[OC] Forest-area change from 2000 to 2023 in the ten countries with the largest forests in 2000](https://assets.buttondown.email/images/22ead789-581d-481a-991a-b2ffdef424ac.png)
"We inspected 2,061 files from one public n8n workflow collection and grouped them by normalized nodes and connections. This analysis counts one representative from each of the resulting 2,030 groups. Of those representatives, 1,721 (84.8%) lacked all five signals listed below. That measures declarations in these files, not production failure rates or every possible way to handle errors.
The five signals are: at least one node with retryOnFail=true, onError=continueErrorOutput, onError=continueRegularOutput, or continueOnFail=true; or a node type ending in errorTrigger. An absent signal means the tested condition was not satisfied, not that every related field is missing."
A pretty cool project, which draws its data from an online game.
This is a great bit of geospatial curiosity that I wasn't aware of: West and East Germany used different coordinate systems: DHDN (Bessel 1841 ellipsoid) vs. Pulkovo 1942/83 (Krassovsky ellipsoid). After reunification, surveyors and GIS teams spent years merging the two geodetic systems into today's ETRS89/UTM standard. An interesting exercise: plotting the Brandenburg Gate under each system yields coordinates almost 600 meters apart despite using the same projection method.
"Data-center spending is greater than that for the canals, railroads and grid combined—creating jobs and wealth but also boosting inflation."
(via Peter Wood)
"We read 2.2 million recipes from American recipe websites to find the dishes people write down most, the handful of ingredients nearly all of them share, and what you could cook tonight with what's already in your kitchen."
"Aggregated Quantitative Results of European Municipalities", part of the Digilog research project on digital transformation in local government.
(via Ettore Ferranti)
AI
A great visualization of the AIAAI (AI, Algorithmic and Automation Incidents and Controversies) database.
OpenAI Astra, its first GPT-6 model, seems to be particularly good at controlling computers and therefore, the authors claim, at robotics.
ACADEMIC PAPER KLAXON 1
"We ask whether AI-generated text can be identified one level deeper, from structural signatures". It looks like it can.
(via Link Molto Belli)
ACADEMIC PAPER KLAXON 2
"The performance of LLM-based agents is jointly shaped by their base models and the harnesses that mediate their interaction with the environment. Because different models exhibit distinct behaviors, effective harness design is inherently model-specific. Yet agent harnesses are still largely engineered by human experts, a paradigm that scales poorly as modern LLMs become increasingly diverse and rapidly evolving. In this paper, we introduce Self-Harness, a new paradigm in which an LLM-based agent improves its own operating harness, without relying on human engineers or stronger external agents. We operationalize Self-Harness as an iterative loop with three stages: Weakness Mining, which identifies model-specific failure patterns from execution traces; Harness Proposal, which generates diverse yet minimal harness modifications tied to these failures; and Proposal Validation, which accepts candidate edits only after regression testing. We instantiate Self-Harness across Terminal-Bench-2.0, SWE-bench Verified, and AppWorld using a minimal initial harness and three base models from diverse families: MiniMax M2.5, Qwen3.5-35B-A3B, and GLM-5. Across all nine model--benchmark combinations, every final harness improves both held-in and held-out pass rates, with overall relative gains of up to 132%. Qualitative analyses further show that the retained mechanisms address benchmark-specific bottlenecks in artifact handling and runtime control, software-patch verification, and application-state retrieval. These results suggest a path toward LLM-based agents that are not merely shaped by their harnesses, but can also participate in reshaping them."
"Jev went live on September 15, selling a new category: a model that doesn’t write a response, it picks one, a choice or a score or a yes-or-no with a confidence number stapled on. TypeSafe, the startup behind it, raised a $40 million seed at a $200 million valuation the same week and cleared a 140,000-person waitlist in 36 hours. Within a day, 13% of Vercel’s paid AI Gateway customers had already wired it into a workflow. Within three days, Cloudflare, LangChain, and Langfuse had shipped integrations for it."
