Published on June 19, 2026 at 18:01 CEST (UTC+2)
Project Valhalla, Explained: How a Decade of Work Arrives in JDK 28 (380 points by philonoist)
Project Valhalla, Explained: How a Decade of Work Arrives in JDK 28
After years of skepticism, Project Valhalla’s value classes and objects (JEP 401) have finally been integrated into OpenJDK, targeting JDK 28. The change is massive—over 197,000 lines of code—but is currently a preview feature disabled by default. Brian Goetz noted this is only the first part of Valhalla, suggesting the community will now shift from “they’ll never ship it” to “they didn’t ship the most important part.” The article provides a deep dive into the history and implications of this long-awaited JVM enhancement.
The Productivity J-Curve [pdf] (2018) (16 points by kioku)
The Productivity J-Curve [pdf] (2018)
This MIT paper introduces the concept of a productivity J-curve: when new technology is adopted, productivity often initially dips before rising above previous levels. The theoretical framework explains why short-term disruption can lead to long-term gains, a pattern relevant to organizational adoption of AI and automation. The document is academic and data-driven, offering a model for evaluating technology investments.
DuckDB Internals: Why Is DuckDB Fast? (Part 1) (326 points by marklit)
DuckDB Internals: Why Is DuckDB Fast? (Part 1)
DuckDB has rapidly evolved from a research project to a widely used in-process analytical SQL database, powering billions of queries in BI, ETL, and embedded analytics. The article explains its speed through columnar storage, vectorized execution, and tight integration (no server, just a library). Part 1 focuses on architecture fundamentals, with later parts likely diving into query optimization and parallelism.
"No Feigning Surprise" (56 points by evakhoury)
"No Feigning Surprise"
A simple writing tip comic: when communicating, say something genuinely surprising rather than pretending to be surprised. The comic encourages directness and authenticity in technical and creative writing.
Ten years of ClickHouse in open source (173 points by saisrirampur)
Ten years of ClickHouse in open source
ClickHouse celebrates a decade since its open-source release, now the most popular open-source analytical database with over 2,000 contributors. The author outlines several levels of openness and explains how ClickHouse aims for the highest tier—transparent development, CI, documentation, and user support. The post reflects on the project’s evolution and what it means to build a great database in public.
Zen and the Art of Machine Learning Research (160 points by jxmorris12)
Zen and the Art of Machine Learning Research
This essay likens AI research to Zen meditation: insights come unpredictably, and the key is consistent discipline—reading and building, day in and day out. It emphasizes that no single breakthrough happens on demand, and that researchers must accept randomness. The post also references the SwiGLU paper by Noam Shazeer as an example of ideas that “work” without clear explanation.
To study how chips work, MIT researchers built their own operating system (286 points by speckx)
To study how chips work, MIT researchers built their own operating system
MIT CSAIL created a new kernel called Fractal to gain a cleaner view of processor behavior, without the abstraction layers of modern OSes. Using Fractal, they discovered previously unknown behaviors in Apple’s M1 chip. The work aims to improve hardware understanding, which could lead to better performance optimization and security analysis.
I found 10k GitHub repositories distributing Trojan malware (883 points by theorchid)
I found 10k GitHub repositories distributing Trojan malware
The author discovered a large-scale malware campaign on GitHub where thousands of repositories (not forks) copied legitimate projects and added malicious links to zip archives. The repositories share a common pattern, enabling a script to detect them. The campaign poses a serious supply-chain threat, especially for developers who clone and run such code.
Gribouille 0.3.0: A Grammar of Graphics for Typst (154 points by mcanouil)
Gribouille 0.3.0: A Grammar of Graphics for Typst
Gribouille is a plotting library for the Typst typesetting system, version 0.3.0 adds finer control over guides (axes, legends), a theme: parameter for compositions, and stacking defaults for area plots. It targets static data visualization in academic and technical documents, offering a tidy grammar similar to ggplot2.
The AirPods Effect (251 points by herbertl)
The AirPods Effect
An essay examining how widespread use of wireless earbuds (like AirPods) is reducing real-world spoken interaction. Data shows a 28% decline in spoken words per person between 2005 and 2019 in the U.S. The article explores the social and psychological consequences of always-on audio consumption, suggesting technology is pushing people apart.
Low-level performance optimization for AI/ML workloads
Project Valhalla’s value classes reduce Java object overhead, directly benefiting JVM-based ML pipelines (e.g., Spark, Flink, Deeplearning4j). Similarly, DuckDB and ClickHouse emphasize columnar storage and vectorized execution—techniques that mirror the data-access patterns of training and inference. Why it matters: As AI models grow, efficient data handling becomes a bottleneck; these infrastructure advances enable faster iteration and lower cost. Actionable takeaway: Consider adopting in-process analytical databases or JVM improvements for data preprocessing and feature engineering.
Open-source ecosystems as both enabler and risk
ClickHouse’s decade-long journey highlights the power of transparent, community-driven development—a model that AI/ML pipelines rely on (PyTorch, Hugging Face, etc.). However, the GitHub malware discovery reveals that open-source supply chains are vulnerable to injection attacks. Why it matters: AI developers often depend on hundreds of dependencies; a compromised repo could backdoor models or training data. Actionable takeaway: Implement dependency scanning, pin versions, and audit contributions—especially for code that runs in production ML environments.
The disciplined, stochastic nature of AI research
The Zen essay and reference to Shazeer’s SwiGLU remind us that breakthrough ideas often arise from persistent experimentation rather than linear planning. Current trends like foundation models and scaling laws appear to follow a similar pattern—results can seem random until a critical mass of work accumulates. Why it matters: This insight informs how teams allocate R&D resources; expecting predictable quarterly breakthroughs is unrealistic. Actionable takeaway: Foster a culture that encourages long-term, hypothesis-driven tinkering and accepts “negative” results as part of the process.
Hardware-aware ML and the need for low-level observability
MIT’s Fractal OS exposes unexpected behavior in Apple’s M1 chip, underscoring that even modern processors contain undocumented performance or security quirks. As AI moves to specialized hardware (GPUs, TPUs, NPUs), understanding these quirks becomes critical for optimizing kernels and reducing latency/inference costs. Why it matters: Without fine-grained observability, ML engineers may leave performance on the table or introduce subtle bugs. Actionable takeaway: Invest in profiling tools and, where possible, run microbenchmarks on the actual hardware used for training/inference.
Social implications of AI-powered consumer devices
The AirPods Effect demonstrates how smart earbuds (with noise cancellation, voice assistants) reduce spontaneous human conversation—a 28% drop in spoken words in the U.S. Similar trends could emerge from wearable AI (smart glasses, always-on microphones). Why it matters: AI/ML product designers must weigh convenience against social isolation. Actionable takeaway: Include ethical impact assessments in product roadmaps; consider features that encourage real-world interaction rather than replacing it.
Data visualization remains a critical underleveraged AI/ML skill
Gribouille’s grammar of graphics for Typst caters to the same need as Matplotlib, ggplot, and Vega—generating publication-quality plots from analytical results. As AI research and reporting proliferate, the ability to communicate findings clearly through static visuals is essential. Why it matters: Many ML engineers focus on model performance but neglect visual storytelling, which is key for adoption and peer review. Actionable takeaway: Equip teams with robust plotting libraries and teach best practices for presenting uncertainty, comparisons, and trends.
Analysis generated by deepseek-reasoner