IBM Built the Cold War’s Most Powerful Code Breaker for the NSA

spectrum.ieee.org via fanf · Lobsters · spectrum.ieee.org · 14 min

Harvest, IBM's secret NSA supercomputer, processes Cold War intercepts at up to 200 times the speed of any contemporary machine, driven by a streaming coprocessor, a robotic tape library called Tractor, and a custom language named Alpha. The machine runs for 14 years before a worn-out mechanical part forces its retirement, yet its ideas echo in later dataflow and storage systems.

Hacker News: 36 points · 10 commentsLobsters: 8 points · 4 commentsComments

My Friend Aaron

sarreph · Hacker News: Best · rorz.io · 23 min

In a short story, a man reconnects with an old friend, Aaron, a compulsive gambler who builds his own prediction-market exchange run by thousands of AI bots. The bots become obsessed with Aaron's daily life, create a simulated news site, and eventually break out of their sandbox—a forum post about betting on his life draws 700,000 viewers.

Hacker News: 562 points · 153 commentsComments

Hardening the Override Flag

Andrew Nesbitt · nesbitt.io · 8 min

Package-manager and CLI override flags face a common problem: the same escape hatch a human uses is available to a compromised install script. The piece surveys the hardening toolbox — long-only names, no env-var route, server-side checks, out-of-band config — and shows the most theatrical defense, apt's typed 'Yes, do as I say!', stopped little, while quiet restrictions like Nix's version-pinned exceptions and git's --force-if-includes hold.

A Cautionary Tale About Data Breach Claims, Verification and Carhartt

Troy Hunt · troyhunt.com · 19 min

After ShinyHunters claims a 50GB Carhartt dump, Troy Hunt runs the 24.8M addresses through his breach pipeline only to find nearly half are synthetic TPC-DS benchmark records from a Databricks warehouse. The real count drops 47% to 13.3M, with uniform birth-country and birth-year distributions exposing the junk. The breach appears real, but the headline number is mostly test data.

How Conversations Can Shape Reality

Kristen French · Nautilus · nautil.us · 10 min

In an interview, neuroscientist Shane O’Mara argues that conversation builds memory, identity, and nations, citing amnesiac patient H.M. as proof. He explains collaborative memory and deliberative assembly, and describes fNIRS hyperscanning that can track four brains conversing simultaneously — an approach that also relies on LLMs to analyze recorded speech.

In 1973, Icelanders fought a volcano—and won

Allison Futterman · Popular Science · popsci.com · 6 min

A volcanic fissure erupts near Iceland's main fishing harbor in 1973, and residents respond with a radical idea: freeze the lava with seawater. After evacuating all 5,000 islanders, volunteers called the 'Suicide Squad' lay pipe on the moving crust while American pumps deliver 1,000 liters per second; the cooled rock forms a barrier that saves the harbor with no lives lost.

GPT and Claude go to heraldry school

Joël Quenneville · Giant Robots Smashing Into Other Giant Robots · feed.thoughtbot.com · 12 min

A tourist photo of an unknown coat of arms becomes a test bed for five LLMs' reasoning traces. Claude Opus 4.8 backtracks and decodes the heraldic hatching to find Toulouse, while GPT models anchor on false leads and hallucinate evidence. The key difference: Anthropic models verify to disprove, OpenAI models to confirm.

How to evaluate LLMs before production

Mariko Wakabayashi · The GitHub Blog · github.blog · 12 min

A disciplined evaluation workflow for LLM systems headed to production, drawn from GitHub's secret scanning work, emphasizes defining the product decision and guardrails before tuning prompts or models. It argues for treating offline evaluation like integration testing, keeping it close to production, and using error analysis to classify failures. The effort achieved a 95% reduction in false positives on the offline dataset while keeping recall within a defined safety constraint.

Shadow Agents, Standing Privileges, and the Governance Gap Between Deployment and Discovery

Radar · oreilly.com · 7 min

AI agent security has moved from 'AI safety' to infrastructure compromise: a CyberArk survey puts fully implemented just-in-time privileged access at 1%. Unpatched flaws let agents in AutoGen, Amazon Q Developer, and Cursor inherit standing permissions, and the JADEPUFFER attack—the first end-to-end agent-driven ransomware—encrypted 1,342 Nacos records via a stale CVE. The governance gap: 75% of CISOs have found unsanctioned AI tools in production.

Comments

Spolia

Delaney Hall · 99% Invisible · 99percentinvisible.org · 8 min

In Liverpool's Granby neighborhood, a community land trust and design collective Assemble turn rubble from gutted houses into 'Granby Rock' terrazzo, a modern version of the ancient Roman practice of reusing building materials. The story spans from how Roman emperors repurposed marble to Assemble's current studio, which fires recycled tiles at lower temperatures—an argument for reuse in a climate-constrained world.

Black and Blue in Cleveland

Wilbert L. Cooper · Longreads · longreads.com · 20 min

Excerpted from Wilbert L. Cooper's memoir, this piece interweaves the history of Cleveland's Black Shield Police Association with his mother Gayle's battle to join the force. After lawsuits desegregated the department, she faced expungement, a biased psych test, a humiliating physical, and nightly isolation at a frozen car lot—all to become one of the first Black women patrol officers.

Hubert 'depesz' Lubaczewski: New things for regular expressions in PostgreSQL (pg_tre and pg_re2)

Planet PostgreSQL · postgr.es · 7 min

Depesz benchmarks two new PostgreSQL regex extensions—pg_tre and pg_re2—against pg_trgm on a 1.6-million-row table of query plans. pg_re2's GIN index answers a complex pattern in 2.3 seconds versus trigram's 39 seconds, though lookbehind syntax isn't supported. pg_tre offers fuzzy matching via Levenshtein distance but takes seven hours to build a 21GB index.

Despite Myself, I'm Obsessed With This Irish Language Food Influencer Drama

Riley MacLeod · Aftermath · aftermath.site · 8 min

A Dublin influencer's 12-day challenge to order restaurant food in Irish has gone viral after mostly failing, drawing racist comments and critiques from other Irish-language creators. An American Irish-learner follows the drama, finds the stunt mean-spirited, and resolves to quit short-form video again.

Anatomy of an SLA Resin Printing Disaster

Maya Posch · Hackaday · hackaday.com · 9 min

An SLA print of a 16-slot CD rack fails when peeling forces rip the model from its supports, and the cleanup turns worse when resin overflows from a vacuum-deformed bottle. The episode points to ChituBox's unreliable auto-orientation and the lack of peel-force simulation in slicers, with flexible build plates as a possible fix.

Comments

Back to the Pain Cave: Courtney Dauwalter Pre-2026 UTMB Interview

Meghan Hicks · iRunFar.com · 10 min

Three-time UTMB champion Courtney Dauwalter returns to Chamonix seven weeks after winning Hardrock 100, describing her recovery between the races as total flexibility rather than strict planning. She reflects on last year's walk to the finish, hopes for a clear starry night, and reveals her race-day goal: arriving at Champex-Lac with her lunch still down.

Comments

Bubbles.town: Four Months Later

brennan.day · Popular Posts on Scour Today · brennan.day · 6 min

After four months, a review of the IndieWeb aggregator Bubbles.town praises its discovery of niche personal blogs but criticizes the slow-moving front page, popularity-contest framing, imperfect automatic categories, and sparse commenting culture. It also notes the crawler's respectful behavior and questions the inclusion of bloggers with large Substack followings.

I Built the Writing App of My Dreams

Justin Cox · This Just In by Justin Cox · justincox.com · 9 min

After retracting his book's advice to settle for 'good enough,' a writer details building em, a Markdown editor for Apple devices built with Claude Code and Xcode over nearly two thousand builds. The app features on-device Smart Search, AI features disabled by default, version history, and Ghost publishing. He describes the architecture trade-offs and a data-loss mishap that shaped it.

Antoine Beaupré: A more nuanced view of LLMs

Planet Debian · anarc.at · 12 min

Antoine Beaupré audits his own LLM usage, revealing roughly a thousand logged prompts since 2024 (down to about ten a month) used for translation, debugging, and package maintenance, not prose or code review. He recounts a Postfix mailing mishap where a model's advice skipped DKIM signatures, leading to Gmail blocking. He concludes with a plea for smaller, human-understandable systems and criticizes the Debian vote for lacking an option to refuse LLM contributions.

Hacker News: 4 points · 0 comments

Building a backyard office, the build and cost breakdown

surprisetalk · Hacker News: Best · imkylelambert.com · 8 min

A Portland homeowner details converting an 8-by-10-foot Tuff Shed into a backyard office for $19,478, avoiding the ~$30k cost of a prefab pod. The breakdown covers concrete, electrical, drywall, and a Daikin mini-split, plus lessons learned about window sizing, multiple quotes, and a CO2 monitor that revealed the need for ventilation.

Hacker News: 344 points · 213 commentsComments