arstechnica.com●Covered by 4 sources
SCOTUS rules 6-3 that your phone's location history generally needs a warrant
The Supreme Court held that cellphone location data held by companies like Google is protected by the Fourth Amendment, rejecting the argument that handing over slices of it or sharing with apps kills your privacy expectation. The ruling guts the geofence warrant, the police trick of vacuuming up everyone in an area, though the court stopped short of calling those warrants flatly unconstitutional.
So what? ↓Until now, cops could ask Google for a list of every phone near a crime scene and sweep up total strangers in the dragnet; this says they generally need a warrant first, which is the rare privacy win that protects you even when you did nothing wrong.
Source: arstechnica.com
theregister.com●Covered by 3 sources
South Korea is throwing roughly $1T at memory chips and humanoid robots
Seoul's "3S+1F" plan runs to about a quadrillion won (figures range from $550B in private commitments to nearly $1T total), with the world's two biggest memory makers vowing new fabs to ease the supply crunch dubbed "RAMageddon" and a target of commercial humanoid robots by 2028.
So what? ↓Memory chips sit inside your phone, your laptop, and every AI data center, so a trillion-dollar bet to ease that supply crunch is the rare government plan that could actually show up in what you pay for hardware.
Source: theregister.com
techcrunch.com●Covered by 3 sources
Blue Origin still can't say why New Glenn blew up, but swears it flies again this year
A month after the explosion that deleted its launchpad, Blue Origin still has no root cause for the failure, yet the CEO says reconstruction has already begun and the rocket will return to flight before year's end. The Register flagged that timetable as ambitious, which is generous phrasing for fixing a problem you haven't diagnosed.
So what? ↓"We'll be flying again this year" means a lot less when it follows "we don't know what went wrong," and every customer with a satellite waiting on New Glenn is now staring at a schedule built on a mystery.
Source: techcrunch.com
Meituan●Covered by 4 sources
Meituan trained a 1.6-trillion-parameter model on Chinese chips, with zero Nvidia
Meituan's LongCat 2.0 (also tagged Owl Alpha) clocks in at 1.6T total parameters with roughly 48B active, a 1M-token context, and 35T training tokens, and it's reportedly the first near-frontier model trained at scale entirely on domestic Chinese accelerators. The whole pitch is the supply chain, not the benchmark: no foreign silicon touched it.
So what? ↓Nvidia export bans were supposed to keep China a generation behind, and a working trillion-parameter model trained without a single Nvidia card is the loudest sign yet that the chokehold is loosening, which means the AI price war that keeps your tool bills down isn't going anywhere.
Source: Meituan
Meta●Covered by 4 sources
Meta's brain-to-text decoder hits 61% word accuracy, no surgery required
Meta announced Brain2Qwerty v2, a real-time sentence decoder that reads raw, non-invasive brain signals at roughly 61% word accuracy overall (78% for its best participant), and it released the v1/v2 training code on Hugging Face with BCBL putting out the v1 dataset. That beats other non-invasive methods, though 61% means it still misses about two words in five.
So what? ↓A typing interface that works without drilling into your skull is the version regular people might actually use someday, but "reads your brain signals" plus "Meta" is a sentence worth reading twice before you volunteer your thoughts.
Source: Meta
Cursor●Covered by 4 sources
Cursor's coding agents now live in your pocket, public beta and all
Cursor shipped an iOS app, now in public beta, that lets you launch and review coding agents from your phone using always-on cloud agents or remote control of the ones already running on your computer, with Live Activities and diff review on the small screen. It's pitched for handling incidents, customer issues, and feedback while you're away from the desk.
So what? ↓The whole point of an AI agent is that it works while you're not watching, so putting the launch button and the diff review on your phone means the code keeps shipping during your commute, your dinner, and the weekend you swore you'd take off.
Source: Cursor
techcrunch.com●Covered by 3 sources
Gemini will now draw you pictures based on what it knows about you, for free
Google opened up Gemini's Nano Banana-powered image generation to eligible free users in the US, letting the app make images from your interests and data pulled from connected Google apps, no explicit prompt required. It reads your preferences and generates accordingly, which is a lot of trust to hand a chatbot for the price of nothing.
So what? ↓The part where it skips the prompt entirely tells you exactly how much of your Google footprint it's already quietly reading to decide what to draw.
Source: techcrunch.com
DeepSeek●Covered by 3 sources
DeepSeek open-sources DSpark, claiming up to 85% faster responses per user
DSpark is a speculative-decoding system where a small model drafts token candidates and the big model verifies them in batches, boosting per-user response speed 60 to 85% without changing outputs. DeepSeek claims 30.9% longer accepted lengths than Eagle3 and 16.3% over DFlash, and it already plugs into vLLM.
So what? ↓Squeezing more out of fewer chips could chip away at China's reliance on US high-end hardware, and the cheaper a lab runs, the harder it can lean on your AI bill in your favor.
Source: DeepSeek
The Verge●Covered by 2 sources
Quantum computers still can't do a single useful thing, but the deadlines keep coming
To date, no quantum computer has conclusively performed one commercially useful task; existing machines are too small and error-prone. That didn't stop a June 22nd executive order to speed up US quantum efforts and Trump's science adviser promising a machine "powerful enough for scientific discovery by 2028," while Microsoft touts its Majorana 2 chip and one startup says it'll leapfrog everyone (with hardware it doesn't yet have).
So what? ↓When the political deadlines and vendor promises all race ahead of a technology that has yet to clear its first real-world task, the safe bet is on the asterisk, not the press release.
Source: The Verge
The Verge●Covered by 2 sources
Microsoft is reportedly weighing five Xbox studio closures, with Marvel's Blade on the chopping block
The Verge reports Microsoft will announce another wave of Xbox layoffs next week, with at least five studios facing possible closure, spinoff, or merger, and the Marvel's Blade game (from Dishonored maker Arkane) said to be in line for cancellation. Xbox union members are already pushing for transparency and good-faith bargaining ahead of the cuts.
So what? ↓The studio that gave you Dishonored is reportedly on the brink, a reminder that when the spreadsheet says cut, it's the people who actually make the games who go first.
Source: The Verge
techcrunch.com●Covered by 2 sources
Proton's privacy-first chatbot Lumo 2.0 learns to draw
Proton shipped Lumo 2.0 this week, adding image generation and editing to its privacy-focused AI chatbot alongside a broader set of new capabilities.
So what? ↓A maker whose entire brand is not snooping on you is a rare pick for anyone who wants to generate pictures without making their data the price of admission.
Source: techcrunch.com
the-decoder.com●1 source · panel-picked
US political campaigns now run on AI at basically every step
A New York Times report says both Republican and Democratic campaigns now lean on AI across the pipeline, from vetting opponents to micro-targeting voters, while Europe is drawing tighter limits on the same tools.
So what? ↓The persuasion machine deciding which ad, text, and talking point lands in your feed is increasingly automated, so the version of a candidate you see is tuned to you specifically, and Europe is betting that's a line worth regulating.
Source: the-decoder.com
thereallo.dev●1 source · panel-picked
A blog post says Claude Code is steganographically marking your requests
A post on thereallo.dev claims Anthropic's Claude Code hides invisible markers inside the requests it sends, a write-up that climbed to 314 points and 100 comments on Hacker News. We've got the claim and the receipts trail, not Anthropic's side, so file the why under unconfirmed for now.
So what? ↓If a coding tool is quietly tagging the text you feed it, that's worth knowing about even before anyone agrees on what the tags are for.
Source: thereallo.dev