Monday, 19 October 2026
Anthropic engineers become managers of AI agents; Linux 7.3 ships as a record-breaker; Flock license-plate surveillance ruled a warrantless Fourth Amendment search
Engineering
LWN.net
Linux 7.3 released
Linus Torvalds released Linux 7.3 on October 18, 2026, with the merge window for 7.4 opening the next day. LWN calls it 'a record-breaker on all fronts,' with 18,765 non-merge commits from 3,021 developers, 807 of them first-time kernel contributors. Notable additions include a completed sched_ext subscheduler framework that lets engineers replace the scheduler without patching the kernel, global per-CPU data for BPF programs, a new failfs filesystem that fails all operations to help test error-handling code, the fchroot() system call, and working-set tracking for virtual machines via userfaultfd. Group-scheduling improvements target large multi-core systems, ntfs3 gains alternative data-stream support, and there are many SMB server improvements.
arXiv
Hybrid Reasoning Agents: LLM agents that think in latent space, act through text
Zou et al. propose Hybrid Reasoning Agents, a design that lets an LLM agent reason in latent space and act through text within a single generation pass. The system introduces HybridLink, a lightweight component that switches between text and latent modes, and a continuous working memory that preserves hybrid context across turns. Training proceeds in two stages: HybridSFT teaches the backbone to coordinate with HybridLink, and HybridRL learns when to reason latently versus when to act through text. Evaluated at three model scales across 11 benchmarks in mathematics, science, medicine, search, and agentic reasoning, the approach reports a +10.4% performance gain, 35.5%–54.7% fewer tokens, 1.8×–2.6× faster inference, and 2.7×–3.7× cost savings over text-based and pure-latent baselines. Code and models are released at github.com/togethercomputer/HybridAgent.
Citizen Lab
Exposing AdFection: how Discover-Tech turned targeted ads into a pipeline for zero-click spyware
Citizen Lab reports that Discover-Tech, an Israeli ad company, delivered zero-click spyware through ordinary targeted ads — a technique the researchers call AdFection. Customers select targets by mobile advertising ID, browser ID, or IP address; when a targeted ad renders, its script sends the identifiers to a Discover-Tech server without any user click. The server returns exploit code only to selected targets, which can silently install spyware such as Candiru. At least 45 Al Jazeera devices were targeted in 2024–2025, and ad data shows reach into government, military, and NATO networks. The report exposes a structural flaw in the ad ecosystem: any party that can insert an ad into a publisher's inventory can also silently fingerprint and exploit a pre-selected list of devices.
Karri Saarinen
Someone behind the software
Karri Saarinen argues that good products carry the perspective, values, and experience of their makers — and that as software becomes more additive and AI-driven, that perspective can be diluted until a product becomes a collection of features with no unified point of view. Using Linear as a case study, he describes product intuition as judgment built through customer conversations and deliberate choices about what not to build. He warns that a feature can work in isolation yet make the whole product worse, and that agent-driven independent changes produce what he calls a 'Frankenstein's monster.' For engineers, the piece reframes the AI-acceleration era: the bottleneck shifts from building to deciding whether something should exist at all.
404 Media
Footage of the Flock search that a judge ruled unconstitutional
Flock Safety and Motorola's Vigilant operate networks of automated license-plate reader (ALPR) cameras that log timestamped, geolocated photos of vehicles nationwide. In the reported case, a Tulsa County deputy searched Flock and Vigilant records for a car with California plates before he had even finished stopping it; the records showed 79 location hits for the driver in 30 days, and 11 agencies had scanned the rental car more than 50 times in five days. U.S. District Judge Sara Hill ruled the ALPR query a Fourth Amendment search conducted without a warrant or probable cause — the only basis was the out-of-state plate — and excluded both the travel history and the evidence found in the resulting vehicle search. The ruling is the first to treat a nationwide passive ALPR database query as a constitutionally significant search, raising questions about the legal status of mass-collection infrastructure more broadly.
Humanities
Aeon
Horrible, unmanly, indispensable: the Renaissance gun
Historian Catherine Fletcher traces how Renaissance Italian writers first condemned the gun as diabolical and unchivalrous, then within roughly 50 years came to accept and even celebrate it. She attributes the shift to overlapping pressures: militia service, civic shooting contests, court luxury culture, hunting, and the rhetoric of popular liberty. Florence fielded a 500-handgun militia in 1506; Venice and Lucca were holding shooting contests by the 1520s; luxury wheellock guns appeared at court. By the mid-16th century firearms had spread into both rural and urban life and become associated with masculinity. Rising ownership also brought accidents, intimidation, and crime, leading authorities to impose licensing and spatial bans — a pattern of rapid adoption followed by regulatory catch-up that recurs in the history of disruptive technologies.
Terry Tao's blog
Will it be a golden age or a dark age for mathematics (and science)?
In a guest post on Terry Tao's blog, mathematician Francesco Maggi argues that LLMs will tip science toward either a golden age or a dark age, depending on how the field responds. A golden age requires deliberately maintaining education, expertise, and open access to frontier tools. A dark age arrives when output grows fast but independent human research weakens and knowledge concentrates in private hands — the specific fear being that machine-generated proofs, such as the long Lean scripts LLMs already produce, become alien to human understanding. Maggi identifies concrete risks: credit attribution, secrecy, atrophied student training, and private control of the most capable tools. He calls for disclosure standards, public research infrastructure, and mathematicians who continue to write and review work in forms humans can read.