Today’s Highlights#

Start with the investigation that dominated the day: Bloomberg’s reconstruction of a U.S. strike on an Iranian elementary school. A Pentagon internal review says a former military compound that had become a school a decade ago, visibly separated by walls and by 2017 already showing playgrounds and painted surfaces, was retargeted as a day-one aimpoint after an intelligence database was never updated, a 2019 analyst note landed in a disconnected system, and heavy reliance on Palantir’s AI-enabled Maven Smart System — which fuses more than 150 data feeds to accelerate targeting — compressed hours of vetting into minutes while civilian-harm reviewers were largely sidelined. The result was more than 150 deaths, including at least 123 children, and a UN finding of failure to verify. Around that hard question of speed versus care, today branches in two directions: on one side, a 25-line Jev replica and a deep dive on tokens getting too cheap to meter explore how splitting judgment from generation can make AI cheaper and more local; on the other, a first-in-the-nation city ban on data-driven grocery pricing and a 600,000-posting audit of ghost jobs push for transparency in everyday markets. In the lab, Claude surfaces a CRISPR-like enzyme system from DNA databases, and in a Scottish neighborhood, volunteers reset a clock that has kept a community on time since 1877.

Policy and Governance#

Pentagon Says Overreliance on AI Contributed to Missile Strike on Iran School#

Bloomberg’s investigation traces the February 28 strikes on the Shajarah Tayyebeh school in Minab through interviews with officials involved in the Pentagon’s internal probe. According to the account, the site was catalogued years earlier as a naval facility but was rebuilt as a school on the same parcel, with walls and entrances separating it from the base by 2017 and bright playground markings visible in 2018. A satellite check from 2019 that noted those changes was filed outside the primary targeting database, and the main library was never reclassified, the officials said. As the administration sought to approve more than a thousand aimpoints in days, the review says, work that once took hours was compressed to minutes around Maven, and no civilian-harm specialist rechecked the site. A UN fact-finding mission said the United States failed to do everything feasible to verify the target and that the lapse went beyond negligence; the Pentagon says the full report is still under review. The piece reads less as a single-model failure than as a chain of stale data, disconnected systems, and time pressure. On Hacker News, the thread’s main split was whether AI was the cause or the cover story. Top comments argued any database lookup on stale data — AI or SQL — would have failed, while others countered that placing an unaccountable system in a lethal decision chain is itself a governance failure. A former joint terminal attack controller added that multi-source eyes-on verification was skipped.

Discussion: Hacker News thread

Seattle City Council Votes to Ban Surveillance Pricing in Sale of Groceries#

The Seattle City Council passed the Fair Pricing and Transparency Act, which would make Seattle the first U.S. city to prohibit surveillance pricing — charging different people different prices for the same groceries — based on personal data such as browsing history, real-time location, or inferences about income and household size. The bill keeps ordinary discounts but requires clearer explanations and limits how shoppers can be profiled. Consumer Reports, which advised on the draft, cited past investigations where profiles ran to 62 pages and pricing experiments produced spreads as high as about 23% for the same basket at the same time, with Maryland, Connecticut, and New Jersey having adopted narrower statewide limits. The digest’s reading is that the city is testing whether product-level fairness rules can anchor broader data-governance debates. HN reactions divided on scope: many pointed to the EU’s General Data Protection Regulation as proof that stricter limits can work, while several pushed back that a sweeping ban could spill into necessary record-keeping for schools or clinics and urged HIPAA-like tiered protections instead.

Discussion: Hacker News thread

Tech and Products#

Jev in 25 Lines of Python#

A parody-slash-demo from NobodyWho punctures the hype around TypeSafe’s Jev, a so-called System One model that does not generate text but returns calibrated probabilities over a preset set of choices. The post loads a small local model, Qwen3-0.6B as a GGUF file, feeds an email-classification prompt with labels A/B/C for Legitimate/Spam/Phishing, and computes the next-token logits — the model’s raw scores for what comes next — for those three markers into a softmax to get probabilities. The output is local, fast, and private, and the authors note this is intentionally minimal, pointing to OpenJev for fuller implementations. The thread’s most useful critique was that reading the first label token’s log-probability (the model’s internal confidence signal) is brittle with chat-tuned models, which may want to continue in prose. Top comments argued for constrained or structured generation that forces a full label string and for grammar-guided decoding, with a few readers noting Jev’s real edge is calibration and cost, not novelty.

Discussion: Hacker News thread

Claude Code Reads AGENTS.md Only When Telemetry Is On — Fixed#

A close look at Claude Code reading AGENTS.md shows a local file that required a remote permission slip. The author found that version 2.1.277 was supposed to pick up AGENTS.md when CLAUDE.md was absent, but with telemetry off the file never loaded. The write-up traces the feature to a built-in plug-in called agents-md that shipped disabled by default and gated behind a remote flag, tengu_agents_md_mod; if the flag could not be fetched, the file was ignored even though it lived on disk. A canary-word test in an otherwise empty directory reproduced the block across both DISABLE_TELEMETRY and CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC, and per-project env overrides did not help. The behavior was later corrected in 2.1.281. Several readers on HN questioned why a filename alias needed a full plug-in and a server-side kill switch, and the post author argued that coupling local instruction loading to network reachability harms predictability. An Anthropic engineer replied in-thread to apologize and link the source.

Discussion: Hacker News thread

Tokens Too Cheap to Meter#

Jyn’s essay gathers hardware, model, and inference-engine benchmarks to argue that what is getting cheap is not tokens per se but tasks: for a given quality bar, 2026 is about two orders of magnitude cheaper than 2025. The piece walks through GPU energy efficiency roughly doubling every two years, vLLM gaining about 40% in a little over a year, and vendor stacks adding tens of percent, then overlays pareto frontiers — curves that show the best quality you can buy at each price — for cost versus intelligence. Specialized classifiers such as Jev and Laya, which trade generation for calibrated choices, push that curve another one to two orders for narrow judgments. The author’s forward-looking question is what product shapes make sense when judgment calls become infrastructurally cheap. HN’s top pushback invoked Stein’s Law and the lower bound of compiled tools: several readers argued calling a language model will asymptote toward, not blast past, the per-call cost of grep, with different parallelization and determinism profiles, and that the more useful framing is embedding cheap classifiers inside tools rather than replacing them.

Discussion: Hacker News thread

Business and Platforms#

Transit Rewards#

Waymo’s transit rewards program is launching in the San Francisco Bay Area, first for employees and then more widely. Riders who link a Visa card in the Waymo app and tap the same card on any of the region’s 27 transit agencies that accept contactless Visa within two hours of a Waymo ride automatically receive $2.85 in Waymo Cash, the price of a San Francisco bus fare. The company frames it as a complement to the Bay Area’s tap-to-pay rollout and regional transfer discounts from last December, alongside a deal to lease 40 dedicated spaces at Caltrain stations for staged vehicles. Earlier pilots in San Francisco and Los Angeles informed the design, and Waymo says more cities will follow. HN commenters generally liked the multimodal intent but questioned whether a rebate can patch a fragmented network, citing the long-suspended 47 Muni link and the unfunded Caltrain downtown extension, and several did back-of-the-envelope math on BART and Muni operating costs per trip versus fares collected.

Discussion: Hacker News thread

28% of Job Postings on Company Career Sites Have Been Open Over 90 Days#

The Ghost Jobs Report for September 2026 from Unlisted Careers measures age directly from employers’ own career sites, covering 607,050 open postings across 15 applicant tracking systems. Of postings with a publish date, 28.3% had been open more than 90 days and 94,106 more than 180 days, with a median open age of 36 days; among recently closed postings, 14.6% disappeared within a week and 4.2% reappeared within 30 days at the same employer and title. The report notes that a single generic requisition often collects resumes for many hires, so age alone does not prove a posting is fake, but it cautions that prolonged listings and reposts can waste applicant time. The main split on HN was between hiring insiders who said evergreen requisitions and hard-to-fill specialties explain long windows and other readers who warned that stale or performative listings still obscure true demand, with practical advice to weigh posting age against company size and how generic the description reads.

Discussion: Hacker News thread

Science and Research#

Claude Discovers a Novel Enzyme System with CRISPR-like Repeats#

An update from Anthropic’s life-sciences lab describes a small human team that gave Claude a high-level prompt — search DNA databases for uncharacterized reverse transcriptases — and let agent swarms do the rest. Over about 21 hours, roughly 950 agents using about 210 million tokens surveyed families of reverse transcriptases (enzymes that copy RNA into DNA), reproduced known results as a check, and flagged outliers. One candidate stood out: a reverse transcriptase from a jumbo phage (a virus that infects bacteria) associated with an array of non-coding repeats and a mystery accessory protein, a combination the team calls array-associated reverse transcriptases, or ART. The underlying enzyme had been seen before, but the array-plus-protein system had not. The team has put up a preprint and begun bench validation, noting the function is still unknown. Commenters on HN debated what kind of reasoning this represents, with the more grounded take that pattern recognition plus careful literature replay and anomaly triage explains most of the behavior, while others cautioned that programmability akin to CRISPR remains unproven.

Discussion: Hacker News thread

Society and Culture#

Fixing the Portobello Police Station Clock#

A gentle field note, Fixing the Portobello Police Station Clock, follows an invitation from the community owners of a 19th-century police station in Portobello, near Edinburgh. The building, started in 1877, now belongs to the local group Action Porty, which asked for help resetting the tower clock after acquiring the site. The author and a friend climbed a steep ladder through a dusty attic into the clock chamber, which still has its three dials and original mechanism, now electrically driven with a nighttime chime cutoff. The piece lingers on the small, tangible work of keeping public time accurate and the quiet way a community asset stays cared for after a formal handover. HN responses were warm and anecdotal, from dust-triggered airport swab stories (ion mobility spectrometry is sensitive and prone to false alarms) to reflections that these modest maintenance rituals do more for belonging than grand plans.

Discussion: Hacker News thread

I Don’t Want the Details#

In I don’t want the details, Michael Heap revisits being cut off in a post-incident review by an SVP who said, in effect, I trust your competence — tell me what we will change. At first it felt dismissive, Heap writes, but the line reframes the usual why-did-this-happen narrative, which often produces a plausible timeline everyone can nod along to and then leaves urgency behind. Asking what we will change, the post argues, forces a systemic fix rather than a comforting story; understanding is not the same as remediating. The essay recommends foregrounding verifiable changes and the incentives around them. The thread’s main split was sharp: some readers welcomed the strategic focus and said good leadership should manage outputs, not walk through every technical weed, while others warned that without details you cannot judge whether a big, morale-taxing change is well targeted, urging that details be kept in the written record even if the live conversation centers on next steps.

Discussion: Hacker News thread

Closing#

Today moved between accountability and everyday life. The same impulse that makes systems faster — compressing checks, personalizing prices, stretching listings — is what makes careful human verification, transparent rules, and ordinary upkeep matter more. Whether the tool is a targeting aid, a classifier you can run on a laptop, or a clock tower that just needs the right hands on the right ladder, the through-line is not replacing judgment but placing it better. Keep a little room for the curiosity that does not come recommended. See you next time.