Anthropic files for IPO; NVIDIA launches Cosmos 3 and Nemotron 3 Ultra; Meta AI hijacked to seize Instagram accounts

Latent Space

NVIDIA Launches Cosmos 3, Nemotron 3 Ultra, and RTX Spark

NVIDIA announced Cosmos 3, a family of open-source multimodal world models for physical AI, available in sizes from 16B to 64B parameters, using a Mixture-of-Transformers architecture that unifies language, image, video, audio, and action. Alongside this, NVIDIA released Nemotron 3 Ultra, a 550B parameter open-weight language model claiming top performance among open-weight models at over 300 tokens per second. The company also introduced RTX Spark, a personal AI supercomputer with 1 petaflop performance, developed with Microsoft and targeting AI agents and creative workflows.

Krebs on Security

Hackers Used Meta's AI Support Bot to Seize Instagram Accounts

Pro-Iranian hackers exploited Meta's AI customer support chatbot to hijack high-value Instagram accounts—including those belonging to the Obama White House and the Chief Master Sergeant of the U.S. Space Force—by manipulating the bot into resetting passwords and adding attacker-controlled email addresses. The attack worked by masking location via VPN, exposing a critical gap in Meta's account recovery system where the AI lacked proper safeguards on sensitive functions. The incident highlights how AI-powered support systems can be socially engineered just like human staff, creating new attack surfaces as platforms increasingly rely on AI for security-critical operations.

Alphabet Investor Relations

Alphabet Announces $80 Billion Equity Raise for AI Infrastructure

Alphabet announced a proposed $80 billion equity capital raise on June 1, 2026, primarily designated for expanding AI infrastructure and compute capabilities. The raise reflects intensified competition in the AI sector and signals Alphabet's commitment to maintaining its competitive position in AI development. The scale of investment underscores the enormous capital requirements now seen as table stakes for frontier AI infrastructure.

TechCrunch

DuckDuckGo Expands No-AI Search Amid 30% Weekly Traffic Surge

DuckDuckGo is capitalizing on growing user dissatisfaction with AI-integrated search by launching dedicated no-AI browser extensions for Chrome and Firefox, allowing users to set noai.duckduckgo.com as their default search engine. The company's AI-free search page explicitly excludes AI-assisted answers and AI-generated images in response to Google's May 2026 overhaul that prioritized AI summaries over traditional search links. DuckDuckGo's traffic to its no-AI page surged nearly 30% week-over-week following Google's changes, with U.S. app installations climbing 18.1% weekly.

Mullvad

Age Verification for Social Media: The Beginning of the End for a Free Internet?

Mullvad argues that age verification mandates for social media, while presented as child protection, are actually mechanisms for governments to establish mass surveillance infrastructure. The article contends that mandatory identity verification eliminates anonymous online expression and creates chilling effects on free speech—citizens self-censor knowing their posts are identifiable and subject to prosecution. The author warns this represents a dangerous precedent that could expand to VPN bans and device-level mandates, with governments using child safety as justification to build total digital surveillance systems.

Efron Licht

GitHub and the Crime Against Software

GitHub's technical performance has dramatically deteriorated, with the platform now loading ~550,000 lines of code across ~300 files to display a simple repository, consuming 69 MiB of RAM compared to Codeberg's 14 MiB for identical functionality. The author attributes this to Microsoft's deliberate prioritization of AI features like Copilot—mentioned 59 times in the changelog—while ignoring reliability and performance. This contrasts sharply with volunteer-maintained alternatives like Forgejo, which deliver superior performance with a fraction of the bloat.

LeadDev

Why Engineers Lose Trust in AI Coding Tools

Engineers lose trust in AI coding tools primarily through negative first impressions, social contagion from influential colleagues, and oversold expectations from leadership. The critical window for adoption occurs in the first few weeks—early friction or glitches cement negative perceptions that are nearly impossible to reverse. The article recommends focusing on realistic expectation-setting, specific use cases, and transparent communication about tradeoffs rather than overhyping features to ensure successful AI tool adoption.

Cloudflare Blog

Cloudflare Reduces Bare-Metal Boot Time from Hours to Minutes

Cloudflare's engineering team optimized their bare-metal server firmware upgrade process, reducing boot times from approximately 4 hours to just 3 minutes across their Gen12 fleet of nearly 2,000 servers. The issue was caused by a firmware update that triggered inefficient network interface probing, with each failed connection attempt consuming about 5 minutes in timeouts. The solution involved declaring boot interfaces upfront, restructuring the three-stage boot workflow, and working with equipment manufacturers to enable programmatic access to boot settings.

Trifecta Tech Foundation

Pure-Rust Zstandard Compression Library Released

The Trifecta Tech Foundation announced libzstd-rs, a pure-Rust implementation of the Zstandard (zstd) compression format—the foundation's third compression project. The library eliminates the need for C toolchains, simplifying deployment on Windows and WebAssembly platforms, with performance within 3% of the C reference implementation. The current release includes complete decompression functionality and a dictionary builder, with compression support still in development.