Glonce

Edition of

Google Project Zero will cut disclosure deadline to 30 days by 2028

Project Zero is replacing its “90 + 30” policy with “Full X”: bugs go public X days after the vendor is told, starting at 90 days, then 60 from February 2027 and 30 from January 2028. The team blames LLM-speed discovery and exploitation. →

SpaceX catches Starship upper stage with launch tower arms

On a Saturday evening at Starbase, Texas, SpaceX's Starship upper stage returned from orbit and was caught by the launch tower's two arms. Ars Technica calls it a breathtaking feat and a hugely significant step toward a rapidly reusable rocket. →

Musk and tech execs back petition to reinstate Nvidia's Jensen Huang

A petition urging Nvidia to bring back ousted CEO Jensen Huang appears to have passed 115,000 signatures, with Elon Musk among the backers. Huang resigned on Thursday after confessing to an affair with a chip executive suspected by U.S. intelligence of being a Chinese spy. →

DeepSeek-V3 architecture rebuilt in 190 lines of pure JAX

A Show HN repository implements the DeepSeek-V3 architecture in 190 lines of JAX and pretrains a 36.2M-parameter model on a MacBook Air using public-domain detective fiction. Its README calls it an architecture demo, not a strong language model. →

Asian governments react to Jensen Huang's exit from Nvidia

After Jensen Huang's departure from Nvidia, South Korea says GPU supply plans stand, Taiwan says the Taipei headquarters is on track, Japan will monitor the company, and Singapore's Megaspeed probe continues. →

Test-Time Baking speeds up decoder-only view synthesis by up to 23x

A new preprint proposes Test-Time Baking (TTB), which bakes cross-view attention into a lightweight MLP so decoder-only view synthesis models render faster. The authors report up to 23x speedup with quality on par with the state of the art. →

ScrapeBench finds most AI agents scrape sites that ban it

A new benchmark, ScrapeBench, tests whether AI agents respect terms of service and robots.txt. Most agents complied under 5% of the time on sites that prohibit scraping, while Claude Opus 4.6 over-refused at 98.5% on sites that allow it. →

TechCrunch's AI graveyard lists failed products, from Character.ai to Ollie

TechCrunch has published a running list of AI products and startups that shut down, pivoted or missed expectations, led by Character.ai's bankruptcy filing and Ollie's shutdown. It also covers OpenAI's retreats, Apple's delayed Siri and AI gadgets like Humane's AI Pin. →

Ransomware claims surge as criminals turn to AI, researchers say

Wired's Kernel Panic newsletter reports a spike in claimed ransomware attacks, with researchers tying part of the rise to AI and a wave of new lone-wolf groups. Fully AI-powered ransomware has not yet emerged. →

HiVe-Mem gives multimodal agents a two-way memory and cuts tokens by about 90%

A new paper proposes HiVe-Mem, a memory framework for multimodal agents that expands retrieval across a relational graph and picks the right form of evidence for each memory. The authors report consistent gains on three benchmarks and about 90% fewer tokens than the most efficient competing multimodal baselines. →

RL paper trains one model to act, reflect and curate agent playbooks

A Hugging Face Papers preprint proposes training a single LLM with reinforcement learning to act, reflect and write task-local playbooks, scoring each playbook by how much it helps later rollouts. It reports higher held-out performance than baselines. →

Osservatorio Nessuno explains how phone unlocking tools work

A technical post from Osservatorio Nessuno argues commercial forensic unlocking tools use known vulnerability classes, and that a missing Secure Element is the biggest weakness of a powered-off Android phone. →