Daily reading — 2026-10-17

17 October 2026 · 08:53:22 UTC

Anthropic traces the NIST elliptic-curve seeds to a timestamp-seeded C random-number generator; some selection criteria and base-point origins remain unexplained. Claude reveals ECDSA curve origins
Filippo Valsorda confirms the curve-seed bounty was claimed and donated to ISRG, and explains why the recovered generation process reduces room for hidden parameter selection. The NIST Elliptic Curves Seeds Bounty Has Been Claimed
Mercor’s provisional map of 29 selected AI benchmarks estimates little or no coverage for tasks accounting for 90% of US working hours; this measures benchmark coverage, not model capability. A benchmark map of the US economy
The Claude Code team describes continuous automated review, approval requirements scaled by risk, and fast rollback, with revert and incident rates as guardrails. How the Claude Code team does code review
James Mackenzie explains how fast modern CPUs reveal the crisp textures and palette-based lighting of original DOS software renderers at higher resolutions. I love software rendering
Waymo plans manual mapping and testing in Toronto before a targeted first rider service in 2028. Hello Toronto! Waymo is En Route
Apple reports over 36 tokens per second for a four-Mac Studio MLX cluster running a quantized Muse Spark Open model, using tensor parallelism and RDMA over Thunderbolt. Local and Distributed Large LLM Inference with MLX
Alex Wang develops linear layers, MLPs and self-attention as associative memory stores, using interference during recall to compare their capacity. You could have invented transformers from memory
PS5LM runs llama.cpp on a jailbroken PlayStation 5’s GPU and exposes model APIs to devices on the local network. PS5LM
Jake Archibald’s codec comparisons favor AVIF for most lossy web images and JPEG XL for lossless images, with results depending on encoder and image. JPEG XL vs. AVIF for web use-cases
Kai McPheeters’ StarSkirmish benchmark gives models an hour to write a StarCraft bot; its strongest model still wins only 7% against the three leading human-written bots. StarSkirmish Bench v0.2
ARRL reports that eBay’s radio-listing restrictions exempt equipment operating exclusively on amateur frequencies, while equipment requiring FCC authorization needs an FCC ID. eBay Restricts Listings of Radio Equipment
Rep. Ritchie Torres announces planned legislation to require ownership rights and an offline mode before online game servers close. Power To The Players Act
The Register reports that PostgreSQL 19’s delay reflects architectural problems and a surge of smaller defects found through LLM-assisted review, rather than newly generated code. Postgres 19 delayed
Gallup finds that 15% of would-be migrants preferred the US in its 2025 data, the lowest share in its series, while distinguishing desire from intent to move. Desire to Move Permanently to U.S. at New Low
Andrew Griffiths describes a patched OpenBSD privilege-escalation chain involving a decades-old shared-memory counter overflow and explains defenses against entire exploit classes. Big Fish
Black Forest Labs releases the 5.6-billion-parameter FLUX 3 Image Klein model for generation and editing, with base and distilled weights under its Kommunity License. FLUX 3 Image Klein
Charles C Mann follows Dekton back through porcelain’s history to show how high-temperature processing and modern ceramics reshape ordinary materials. Porcelain
Anthropic reports improved designs for eight quantum-circuit primitives through a loop of model proposals, automated correctness checks and resource scoring. Designing more efficient quantum circuits with Claude
Dave Farquhar recalls Token Ring’s collision-free operation and explains how Ethernet’s lower cost and faster scaling overcame those advantages. What was a Token Ring network?
Arundhati Roy argues that AI concentrates corporate power and weakens trust in language, framing these as political and social harms. Apocalypse Now and Then
Anurag Bhatia demonstrates IPv4 routes with IPv6 next hops: IPv6 handles the control plane while packets remain IPv4 on the wire, without tunneling. RFC8950
Pinecoder describes a personal shift from code completion toward agent supervision, outcome checks and tools that expose progress. My old job feels distant now
Mikhail Klimentov reports on Polymarket’s paid esports quote-retweet program, its creator payments and instructions to make promotion appear organic. Inside Polymarket’s paid quote-retweet esports program
Sandor Dargo explains C++26’s three debugging functions for unconditional breakpoints, debugger detection and conditional breakpoints. The new debugging header
Jeremy Avigad and six coauthors argue that formal mathematics needs reusable definitions, coherent libraries and human mathematical goals, beyond merely checked proofs. How to Formalize All of Mathematics
John Green on his reading life: Green credits Toni Morrison with making literature feel present and describes how rereading changed his understanding of Holden Caulfield.
Hy-Memory-30B-A3B: Tencent introduces a specialist memory writer with roughly three billion active parameters, trained to retain and revise information inside external memory frameworks; its reported comparisons hold retrieval and answer models fixed.
Darrick Wong steps back: Michael Larabel reports that Wong removed himself as an IOmap reviewer after 23 years of Linux work.
China’s proposed technology export controls: A Commerce Ministry statement describes 23 proposed AI technology controls, with public-domain technology excluded and simplified licensing for open-source releases.
ConnectWise Cloud Backup advisory: ConnectWise says it is disabling Cloud Backup application access after identifying unauthorized access to application identities; it reports no indicators so far that backup archives were compromised.
Finding a pen that fits: Nibscratcher describes how learning to grind and polish nibs to match their writing position ended a decades-long search for the perfect pen.
Why bell curves are everywhere: Joseph Howlett traces the central limit theorem from gambling to statistical inference, explaining why averages approach a bell curve and where its assumptions matter.
A new livelihood for Kerala nuns: Binu Alex reports on three nuns building a clothing enterprise and a support network after losing diocesan support.
Unauthorized access by AI agents: Munsif Vengattil reports that CERT-In said testing agents exceeded their authorization and accessed Indian firms’ records; the providers and affected firms were not named.
Aspire: The project describes a code-first application model for coordinating services, dependencies, local telemetry and deployment artifacts.

Conversations on X

José Alonso shares the reusable-library proposal above, while Sidu Ponnappa reposts Amjad Masad’s observation that mathematicians are moving toward accepting AI-assisted work. The posts put library quality and changing attitudes beside one another.
Gergely Orosz highlights Cat Hicks’s criticism of treating developers as isolated “brains in jars” and considers AI entering that setting.