Sunday, 11 October 2026
Yandex's third datacenter struck, Bitwarden moves to dual licensing, and byte-level models challenge the tokenizer orthodoxy
Today's Lead
EngineeringYandex Cloud Status
3rd Yandex Cloud Data Center Was Hit
A third Yandex Cloud data center was reportedly struck, causing brief disruptions to cloud services. The incident was visible in Cloudflare's global traffic data as bot requests from Yandex's ASN fell sharply. Hacker News commenters debated whether commercial data center infrastructure constitutes a legitimate military target, and discussed how the industry might respond with distributed, underground, or more hardened architectures.
Engineering
Bitwarden Community
Bitwarden Dual License Model
Bitwarden adopts a dual license model. Starting with the next release, app store and website builds use the commercial license. The GPLv3 version stays on GitHub and keeps receiving updates. Some future components will be commercial only. Bitwarden says the source stays public, forks stay allowed, and self-hosting and the free plan are unchanged.
MotherDuck
Why DuckDB 2.0 is faster
DuckDB 2.0 is due this fall, and an alpha is available. DuckDB 2.0 uses a separate download pool, so workers only decode data. On a 2.2 GB Parquet file on S3, the alpha takes 7.7 seconds where DuckDB 1.5.5 takes 18.8 seconds. The rewritten recursive CTE engine shows up to 40 times speedup on graph reachability.
arXiv
Byte Language Models: Scaling, Emergent Abstractions, and Information Allocation
The paper tests byte-level language models, which read raw bytes and need no tokenizer. A flat Transformer with token-superposition training and hash embeddings consistently outperforms subword models as model size scales. At matched parameter counts, byte models reach a lower optimal loss than subword models. At the same training compute, subword models still perform better. Byte models use about 3.2 times more source bytes per parameter at the optimum, so they can keep improving with more data.
Nerdy Electronics
Mars Pathfinder Priority Inversion Bug: What Really Happened
The Mars Pathfinder resets came from a priority inversion in VxWorks. A low-priority task held a semaphore that a high-priority bus management task needed. A medium-priority task ran and kept the low-priority task from finishing. The high-priority task missed its deadline, and the watchdog timer ordered a system reset. Engineers fixed the bug by enabling priority inheritance for the semaphore. They uploaded a C program that changed one flag from FALSE to TRUE.
Humanities
GNU Project
What's Wrong with GitHub
The GNU Project urges developers to avoid GitHub for free software projects. The article says GitHub harms users and pushes other projects toward the same harm. GitHub gives unclear license guidance, and much of its site requires nonfree JavaScript. Its Copilot service uses code under many licenses, gives users no control, and can reproduce copyrighted text. GitHub Actions runners depend on GitHub-controlled infrastructure, and the article calls an automatic update feature a backdoor.
Thomas Hühn
Knuth reward check
Thomas Hühn received a Donald Knuth reward check for an error he found in the first word of Computer Modern Typefaces, Volume E of the Computers & Typesetting series. Knuth confirmed the error about twenty years after Hühn found it. The check was for $2.88, written as 0x$1.20 in hexadecimal. Knuth no longer issues real checks and gives a certificate from the Bank of San Serriffe instead.