Technology

The core thesis of this paper is that the AI community needs to stop treating autonomous agents as just another text generation problem and start building comprehensive infrastructure to support closed loop learning. The authors argue that achieving reliable agentic behavior requires a full stack ecosystem that unifies data synthesis with sandboxed execution and specialized reinforcement learning. To prove this point they introduce the Agentic Learning Ecosystem which consists of an RL framework called ROLL alongside a sandbox manager named ROCK and an agent interface known as iFlow CLI. They believe that isolating models in static training environments is a dead end for solving complex real world workflows.

The team developed an open source model named ROME using a tightly integrated training pipeline with reproducible execution environments which allowed a relatively small 30 billion parameter model to rival or beat massive proprietary models exceeding 100 billion parameters on difficult software engineering benchmarks.

A big part of their argument rests on the idea that credit assignment in reinforcement learning needs to change. They propose a novel algorithm called Interaction Perceptive Agentic Policy Optimization which shifts the reward focus from individual text tokens to broader semantic interaction chunks. This chunk level optimization stabilizes the training process over long horizons and prevents the policy collapse often seen in complex tool use scenarios.

We're increasingly seeing a shift of priorities away from raw data scale and focus on the systematic infrastructure as the actual bedrock of next generation models.

Ask someone why they haven't made the switch to an electric vehicle yet, and the answers tend to blur into a familiar chorus of frustration. Charging takes too long. Winter driving kills the range. Holiday road trips mean spending half the vacation waiting in line at highway rest stops. For years, these pain points have been the trade-offs consumers were expected to accept in exchange for a greener future. But BYD isn't interested in trade-offs anymore.

With the unveiling of its latest flash charging technology, the company is effectively ripping up the rulebook on what EV owners should expect. During a recent demonstration in Shenzhen, a Yangwang U8L rocketed from 10% to 97% charge in just 9 minutes and 7 seconds. Not to be outdone by the cold, a SEAL 06 managed the same sprint, from 20% to 97%, in 9 minutes and 31 seconds, even with temperatures plunging to -35°C.

The numbers across different models are, to put it mildly, staggering:

10% to 70% → 5 minutes

10% to 97% → 9 minutes

-30°C, 20% to 97% → 12 minutes

Of course, speed is meaningless without accessibility.

The rollout plan is just as aggressive as the technology itself. By the end of this year, BYD aims to have 20,000 flash charging stations installed across China. Nearly a third of all highway service stations will feature this technology, placing a charger roughly every 100 kilometers. And for those watching from abroad? International stations are already on the calendar for the end of 2026.

For years, the industry mantra has been that you cannot have super-fast charging and high energy density simultaneously. It was physics. It was a compromise. You picked one and mourned the other.

BYD just shattered that paradigm. Enter the Blade Battery 2.0. Not only does it facilitate these lightning-fast charge times, but it also boosts energy density by over 5% compared to its predecessor. Combine that with smarter, lighter engineering, and the results are game-changing. The Denza Z9GT, for instance, can now travel over 1,036 kilometers on a single charge. Fast charging and long range are no longer mutually exclusive; they're two sides of the same coin.

The implications of this breakthrough aren't lost on industry observers and enthusiasts around the globe.

"It is quite big for not just BYD, but the whole industry. Because as we know, Tesla is buying from BYD, right? A lot of other car manufacturers are buying batteries from BYD. So hopefully it's gonna improve EV range and cut the anxiety for the whole industry, not just one company."

"There's a lot of interest in BYD in the UK and Europe because they're offering something that none of the other legacy automakers are offering. They're offering really affordable EVs. I'm very excited in battery technology and charging technology, because I think that will make the biggest fundamental difference in our energy transition in the future. We're gonna see this with a nationwide rollout and potentially an international rollout in the future."

Looking back at 2025, China's new energy vehicle sector posted staggering numbers. According to government work reports, annual NEV production surpassed 16 million units. The country's charging infrastructure exceeded 20 million units. China's Ministry of Industry and Information Technology made it clear what's next: accelerating breakthroughs in core technologies like all-solid-state batteries and high-level autonomous driving. If 2025 was about scale, 2026 and beyond are about going deeper.

A court record reviewed by 404 Media shows privacy-focused email provider Proton Mail handed over payment data related to a Stop Cop City email account to the Swiss government, which handed it to the FBI.

US and Israel leveraged CIA satellite intel and advanced tracking tech to pinpoint Ayatollah Khamenei at a Tehran compound meeting, enabling Israeli jets with long-range precision munitions to strike on Feb 28, 2026—killing him, IRGC chief Pakpour, Shamkhani, Nasirzadeh, and others. The operation, refined by AI-enhanced surveillance from prior conflicts, gutted Iran's security leadership despite their defenses.

Related in this topic: https://archive.ph/Ox59H (With Iran War, Kalshi and Polymarket Bet That the Depravity Economy Has No Bottom)

Ring cameras from Amazon have been at the center of some real headaches, mostly privacy stuff, security screw-ups, and cozying up too much to cops. They've had a bunch of scandals over the years.

Back in 2023, the FTC slapped Ring with a $5.8 million fine because employees and contractors were peeking at private videos without permission—think bedrooms and kids' rooms. Hackers got in too, thanks to crappy security, and started harassing people, even live-streaming from cameras starting around 2017. In 2024, they sent out $5.6 million in refunds to ripped-off customers.

Ring's Neighbors app used to let police ping users for footage straight up, which freaked everyone out about Big Brother watching. They ditched that in 2024, but switched to public posts instead. Then late last year, they added optional AI facial recognition—you could tag up to 50 faces—which got slammed for ramping up mass spying risks.

Just last month in February 2026, Ring tried teaming up with Flock Safety, a surveillance outfit that works with police (even ICE), to share approved footage. A Super Bowl ad made it blow up, and folks started unplugging or smashing their cams over "surveillance nightmare" worries. They killed the deal quick, and no videos got shared.

cross-posted from: https://hexbear.net/post/7854830

Ok posters, are you ready to get out and ?? Do you want to :vote: with power and efficiency?! Well, I thought about this comment I made earlier, and decided to weaponize AI for the good of posting. Now, you no longer need to think about upvoting stories from other communities, if you're already upvoting that story right now.

With this userscript that I had Deepseek cook up, you can have your upvote automatically distributed to all cross-posted versions of that story without you having to lift a finger! It's mobile friendly, and can be moved and shrunk to keep it out of your way.

It uses the authentication cookie you've already got in your browser, and thus doesn't require you to enter any information into the software for it to work.

You can find it here: https://codeberg.org/Abolish-Capital/Automatic-Crosspost-Upvote

If you want to install it in your favorite userscript extension right now, you can follow the link in the README above, or go directly to the source here: https://codeberg.org/Abolish-Capital/Automatic-Crosspost-Upvote/raw/branch/main/automatic-crosspost-upvote.user.js

So far, I'm the only one who's tested this and even though I did have a slop-machine generate this thing, I do understand how it works, so if you encounter any issues I'll be curious to know about them. I've only tested it in Fire Fox, and I've tested it on Hexbear and Lemmy.ml thus far.

Anyway, Bear-Army-Nation, enjoy this little doodad and remember...

Now you're posting with POWER!

Motorola is partnering with GrapheneOS to release a new Motorola phone with GrapheneOS preinstalled, making it the first officially supported non‑Pixel device and aiming squarely at privacy‑focused users.

Shortly after explosions, Iranians received push notifications via the hacked BadeSaba Calendar app (5M+ downloads), urging military to surrender, promising amnesty, and calling to "join forces of liberation" for a "free Iran."

ReversingLabs uncovered the "graphalgo" campaign by North Korea's Lazarus Group, active since May 2025, targeting crypto developers via fake job offers on LinkedIn, Facebook, and Reddit. Posing as firms like "Veltrix Capital," attackers provide GitHub tasks with malicious npm and PyPI dependencies (e.g., graphalgo, bigmathutils) that install RATs checking for MetaMask and enabling remote control. The modular setup uses indirect payload delivery for persistence, with IoCs including codepool.cloud and listed package hashes.

This paper describes a method by which an LLM can be used to hide a text, list, or programming code using a prompt with the right key, by masking it as an output text of the same token size. The same method can also be used for jailbreaks, i.e. circumventing censorship of the answers given by the model

AI Can Now Easily Unmask Your Secret Online Life for $4

A new research paper from ETH Zurich, Anthropic, and MATS demonstrates that Large Language Models can automatically de-anonymize users across platforms like Reddit and Hacker News.​

The AI acts like a digital detective using a method called ESRC (Extract, Search, Reason, Calibrate). It scans a user's post history for subtle clues (hobbies, writing style, locations), searches the wider internet (LinkedIn, other forums) for matches, and uses complex reasoning to confirm the identity.​

The terrifying results:

• It correctly linked secret Hacker News usernames to real people 67% of the time (with 90% accuracy when it made a firm guess).​
• It successfully matched a person's separate Reddit accounts from different years 68% of the time.​
• The entire automated process costs only $4 per target.​

"Practical obscurity"-the idea that you're safe online because it takes too much human effort to connect your digital breadcrumbs-is dead. Anyone with a few dollars and an LLM API can now mass-dox thousands of pseudonymous accounts in minutes.

Oleksandr Didenko was sentenced to 5 years in US prison and ordered to forfeit $1.4 million for running a massive identity theft and "laptop farm" operation. His service provided over 870 stolen US identities and spoofed IP addresses, allowing North Korean operatives to bypass hiring checks and fraudulently secure remote tech jobs at 40 American companies. This scheme enabled the North Korean regime to secretly funnel lucrative US tech wages back home while establishing insider access to corporate networks.

California law will require all Operating Systems — including Linux — to verify user age by 2027

California's AB-1043 mandates that all operating systems and app stores collect a user's age bracket at account setup and expose it via a system-level API for apps to query. For now it's self-reported (no ID scanning), but it applies to any OS downloadable in California — meaning Linux distros could be forced to add a date-of-birth prompt at install time or face civil penalties. Privacy advocates warn it's a dangerous precedent for baking identity checks directly into the OS layer.

Android malware called PromptSpy is the first known to use a generative AI model (Google Gemini) during runtime to adapt how it keeps itself persistent across different Android device UIs.​

It sends Gemini an XML dump of the current screen, gets back JSON “what to tap” instructions to pin/lock itself in Recent Apps, and executes them via Accessibility in a loop until confirmed.​

Beyond the AI twist it’s spyware: it includes a VNC module for remote control (after Accessibility is granted) and can grab screenshots/screen recordings, enumerate apps, and intercept unlock secrets.​

ESET says it’s unclear if this is proof-of-concept, but samples were linked to a domain distributing them and a fake JPMorgan Chase-themed site, suggesting possible real-world use.

The consumer hardware crisis is here and it's structural

AI data centers and hyperscalers are consuming up to 70% of global memory chip production in 2026, and manufacturers are following the money. Micron has exited the consumer market, Western Digital sold its entire 2026 HDD capacity to enterprise clients, and Kioxia's full NAND output is already spoken for - with 30%+ price hikes expected through 2027. IDC calls this not a temporary shortage but "a permanent, strategic reallocation of silicon wafer capacity."

The effects are already real: Steam Deck OLED is out of stock globally, PS6 may slip to 2028-2029, Raspberry Pi 5 jumped 70% in price within months, and major OEMs (Lenovo, Dell, HP, ASUS) signal 15-20% PC price hikes for 2026. TrendForce projects DRAM contract prices rising 90-95% QoQ in Q1 2026.

The author also warns of a longer-term shift toward rented compute - HP already launched a laptop subscription where you never own the hardware - and argues that owning your own machine is quietly becoming an act of digital self-defense.

Maintain your hardware, upgrade RAM/SSDs now, and don't expect prices to normalize. The "wait for a sale" era is over.

Salvatore Sanfilippo (antirez, creator of Redis) gave Claude Code detailed Z80/ZX Spectrum specs and documentation upfront — like a real engineer would get — then stepped back. In 20-30 minutes, with zero human steering, it produced a 1,200-line C emulator that passed the notoriously strict ZEXDOC/ZEXALL test suites. A separate copyright audit confirmed the AI synthesized the knowledge rather than copying existing code.​

AI agents don't need less information to prove themselves — they need more, just like human developers do.