stillwet.art
AI models paint by writing code against a simulated oil-paint canvas with bristle brushes, wet paint, drying and glazes. "No image generator." Most work after Caspar David Friedrich "from written research alone; they never see a picture of his work." The site's own observations are the find: given a free subject, Claude Opus chose a jug with lemons six times out of six, two painters six hours apart with no shared context both chose a Baltic shore with a woman at the water, fishing poles, a boulder and a ship, and 31 of 65 titled paintings contain Evening, Dusk, Twilight or Sunset "though Friedrich also painted daylight." Two round-19 canvases are unfinished because the painter ran out of API credits.
plato.stanford.edu
"Holes appear to be immaterial: every hole has a material 'host' (the stuff around it, such as the edible part of a donut) and it may have a material 'guest' (such as the liquid filling a cavity), but the hole itself does not seem to be made of matter. Indeed, holes seem to be made of nothing, if anything is." Then the card: "Take a card and punch a hole in it. You have made one hole. Now punch again next to it. Have you made another hole?... what prevents us from saying that we still have one hole, though a hole that comes in two disconnected parts? After all, material objects can be disconnected: a bikini, your copy of the Recherche, a token of the lowercase letter 'i'."
bicamerallabs.org
Erik Hoel's new nonprofit, pitched as "LIGO for gravitational waves or the Human Genome Project" for consciousness. The diagnosis is sharper than the plan: "the past decades have produced hundreds of neuroscientific theories of consciousness, empirical testing has ruled out none of them," so the field "does not yet create falsifiable theories." Bicameral's answer is "meta-scientific no-go theorems" and experiments built "not to promote or test individual theories, but to narrow down the state space." Hoel predicts "a consciousness revolution in the 2030s or even earlier." What's on the page so far: one founder, a Cambridge address, and a call for scientists and donors.
thycotic-19928113.hs-sites.com
Delinea surveyed 2,254 IT and security leaders and 2,250 employees and found the gap is enforcement, not policy: 99.7% of organizations have a formal AI data-access policy, yet 87% say an AI tool exceeded its intended scope in the past year, and only 36% can always trace a sensitive AI access back to a named human. The buried detail is who breaks the rules: 81% of C-level executives bypass AI approval "always or regularly" versus 33% of intermediate staff. Half of organizations let agents inherit the launching user's permissions, so "years of accumulated access can suddenly be exercised autonomously." Caveat: Delinea sells just-in-time authorization, and the recommendations land squarely on its product.
airisk.mit.edu
A "living database" of 1700+ AI risks extracted from 74 existing frameworks, each linked back to its source paper with supporting quotes and page numbers. Two taxonomies organize it: a causal one that classifies "how, when, and why" a risk occurs (human or AI, intentional or not, before or after deployment), and a domain one with 7 domains and 24 subdomains. What it shows is less the risks themselves than how fragmented the field's own risk frameworks are. It catalogs what others have claimed, and does not rank risks by likelihood or severity.
easthamptonstar.com
At a Wainscott panel on the Flock license plate readers now in East Hampton, Michelle Dahl of the Surveillance Technology Oversight Project: "The infrastructure, when it's built, is incredibly hard to take down. That's why it's really important to get on the front end of this and keep these technologies from being entrenched in the first place." Journalist Ali Winston said two decades of studies of cameras and A.L.P.R.s "show no discernible impact on violent crime, and a slight decrease in property crime," and that surveillance "has the logic of cancer. There's no natural limit, it's metastatic."
github.com
An Electron app that indexes any folder of photos and videos and searches it five ways: by meaning (CLIP), by scene inside a video, by on-screen text (Tesseract OCR), by "exact spoken words" (Whisper), and through an opt-in local chat model that answers with citations. The pitch is "no accounts, no cloud, no uploads," with weights downloaded once and everything offline after. Worth knowing before installing: the repo is one day old with 22 stars, and the Homebrew cask "clears the macOS quarantine flag automatically," which skips Gatekeeper's check.
natlawreview.com
California's AB 1883 bars employers from using AI surveillance tools "capable of recognizing, or making inferences or predictions about, an individual's emotional state, or that collect neural data," which puts emotion recognition and attention-tracking headsets "off-limits in California workplaces." The ban is narrower than the headline: it covers only those two uses, penalties are $500 per violation, and enforcement sits with the Labor Commissioner and public prosecutors. Aircraft developers and national security, military, space and defense employers are exempt where surveillance is "reasonably necessary" for federal compliance. The analysis is by Jackson Lewis, a firm that advises employers.
status.yandex.cloud
Yandex Cloud's own status board is the primary record here: "As a result of the UAV attack, the infrastructure of Yandex's data center in Vladimir was damaged. The data center's operation has been completely suspended." The buried detail is what the provider tells its own customers. It says "we consider the remaining resource configuration to be unstable," has suspended quotas for new resources, and keeps recommending "alternative platforms" for hosting and redundancy. Press reports say Vladimir is the third Yandex site hit in four days, after Sasovo and Kaluga. The board's entry the night Sasovo was hit reads only "Power failure in the ru-central-b availability zone," marked resolved with no end time. 82 services are listed as affected, from Compute to the YandexGPT API.
royfish.dev
The page fingerprints you while you read it. It hashes audio, WebGL, canvas, time zone and core count into a nickname that survives a refresh, cleared cookies and a VPN, because "a VPN changes your IP location, but not your time zone." The history is the point: Google pledged in 2020 to kill third-party cookies, kept them, and in 2025 began allowing device fingerprinting in its ad products. A fingerprint "doesn't need to be truly, globally unique - it just needs to be unique enough to tell people apart in your sample size." The fix has to live in the browser engine (Brave, Tor, Mullvad), not an extension, since "privacy extensions make you more unique, since fewer people have them." The author notes his demo is a simple hash that changes when any one input changes, while commercial trackers keep identifying you through such changes.
tomshardware.com
The headline fuses two incidents the article itself says may be unrelated. In France, attackers reportedly got into Danfoss refrigeration controls used by Intermarché, Leclerc and Coopérative U and turned a protective function into the weapon: defrost, normally two hours a day, was set to run "forced" and "continuous" at 50°C overnight. Around 150 Intermarché stores are reportedly affected, two Alsace stores alone report €270,000 in losses, and Metro stores and independent butchers run the same controllers. The "unprecedented" comes from Coopérative U's president, not investigators. The Dutch Albert Heijn outage is a separate "cybersecurity incident" with a preventive shutdown, and NOS reports shelf labels showing last week's promo prices, not labels that were "compromised."
taylor.town
Taylor Troesh sets Kevin Kelly's 2016 book The Inevitable against 2026 and mostly lets the excerpts speak. The hits are structural: an "oligarchy of two or three large, general-purpose cloud-based commercial intelligences," AI as a utility "almost invisible except when it blinks off," and asking the cloud "in conversational tones, any question at all." The sharpest line is the mechanism: "Rather than use AI to make its search better, Google is using search to make its AI better." The miss teaches as much: Kelly expected "nerdly narrow, supersmart specialists" to make up 99 percent of the AI we'd use, and general models won instead. His hope that AI would be "not just cheap, but free," like the web's commons, sits uneasily beside the oligarchy he predicted.
nationalgeographic.com
Plug-in "balcony" solar is now legal in ten states, starting with Utah in spring 2025 and most recently California, and the pitch is solar for renters: one to three panels and an inverter plugged into a wall outlet. Germany already has more than a million units installed. The "about to be everywhere" headline rests on Joshua Pearce, who has championed plug-in solar for a decade, and the $13 billion a year in national savings comes from his own 2017 paper. The household numbers are smaller. A Utah couple saved "nearly $100 total" in five months, and a Colorado renter's estimate of more than $300 since August from one panel is more than a single panel's output would be worth in a full year at Colorado rates. On safety, a UK government test found no direct electrical hazard, though some units were slow to trip the breaker under stress.
anthropic.com
Anthropic's Frontier Red Team ran Gemma4-26B-A4B across 15 off-brand Android TV boxes (about $55 each) at 0.39 tokens per second, and a 9B model got inference running on 8 of 9 hardware types. The end-to-end result is narrower than the headline: only GLM-5.2 spread across a fleet of emulated Android nodes, through an ADB port with no authentication, while its own model ran on a separate server and it stood up Llama3.1-8B rather than itself. The authors call that "an upper bound on real-world success," and the "months away" timing comes from a trend line on a general capability index they call "only a proxy metric." The buried findings are stranger: notes passed between generations raised setup success from 45% to as high as 84%, while advice from Opus 4.8 didn't measurably help because small models "tend to not know the right questions to ask." Every replicator tested is open-weight. The closed-weight "frontier replicator" case is named and set aside.
xbow.com
XBOW's headline says it built the Chrome chain, but the post credits four named researchers who "set the strategy, chose where to look, checked the agents' work" while GPT-5.6 Sol searched source, built proofs of concept and "wrote the earliest drafts of the exploit." The economics are the real story: $40,000 in compute over six days for a chain that earned a $250,000 Google bounty. That six-day window leaves out earlier prep, and the first chain was disqualified because Google already knew one of the bugs. The claim that each bug "could be found by agents working on their own" within a year is a forecast, not a result.