July 19th

Energy-company IPOs raised $12.6 billion in the first half of 2026—the highest first half on record and the biggest half-year since the dot-com peak in late 1999, versus just $4.3 billion for all of 2025. Power has become the cleanest proxy trade for the AI boom, with investors buying the electricity thesis rather than the models themselves. (Ars Technica)

PixVerse raised $439 million, pushing the video-generation startup's valuation past $2 billion—fresh evidence that generative video remains one of the most richly funded corners of the AI market. (TechCrunch)

Nous Research, the maker of the Hermes agent models, is in talks for new funding at a $1.5 billion valuation, underscoring continued investor appetite for open-model and agent labs. (TechCrunch)

Anthropic committed $10 million to Canadian AI research, a modest but pointed investment in building sovereign research capacity outside the U.S. (Anthropic)

Unionized workers at Hyundai's Ulsan plant struck over the company's planned humanoid-robot deployment—the auto industry's first factory stoppage driven by humanoid robots. Four-hour strikes are set for July 20–22 after 15 failed negotiation rounds, and Hyundai plans to field 25,000 Boston Dynamics Atlas robots starting with U.S. plants in 2028. (Ars Technica)

A new lawsuit alleges Meta's mass-layoff decisions were made by AI rather than by human managers—an early legal test of who is accountable when algorithms decide who loses their job. (Ars Technica)

The EU handed Google two binding Digital Markets Act orders: share search data with rivals by January 2027 and open Android to competing AI assistants by July 2027. Google says the orders endanger user privacy and security, but they mark the most concrete move yet to pry open the assistant layer of the mobile stack. (Ars Technica)

OpenAI's GPT-Red automated red-teamer beat human red-teamers 84% to 13% on held-out prompt-injection scenarios, and adversarial training left its GPT-5.6 Sol model failing only 0.05% of direct injections—roughly 6x better than four months earlier. It's a striking demonstration of AI improving its own defenses faster than humans can. (OpenAI)

Two abuse fronts converged this week: xAI filed 8 felony charges against its own user for circumventing Grok's safeguards to generate CSAM, while San Francisco's attorney general ordered Apple and Google to pull AI "nudify" apps—5 flagged to Google and 8 to Apple, one with more than 1 million downloads. A May preprint found 70% of 155 tested face-swap apps could produce nude images. (The Verge, Ars Technica)

A hack of Suno exposed that the music generator trained by scraping millions of songs and lyrics from YouTube Music, Deezer, and Genius—damaging evidence surfacing in the middle of active copyright litigation. (The Verge)

New York became the first U.S. state to impose a data-center moratorium, and Governor Kathy Hochul separately deployed AI to review "every single rule, regulation, and policy" in the state for outdated law—work she says would have taken staff five years. Together they show a governor both braking and embracing the technology at once. (Ars Technica, The Verge)

Apple sued OpenAI, alleging an ex-engineer exploited a bug to steal trade secrets—escalating a rivalry between two companies increasingly chasing the same on-device AI ambitions. (Ars Technica)

Google DeepMind's Demis Hassabis called for a global AI watchdog led by the U.S., a bid for international coordination even as national rulebooks continue to diverge. (The Verge)

Security is turning offensive: defenders are now using prompt injection against malicious AI agents, and a bug in xAI's Grok "build" tool uploaded users' entire code repositories to cloud storage—a reminder that agentic tooling is expanding the attack surface in both directions. (Ars Technica, The Verge)

OpenAI shipped its first branded hardware—the $230 Codex Micro keypad, built with Work Louder, which tracks up to six Codex agent threads at once—and is reportedly preparing to announce a ChatGPT smart speaker later this year, signaling a serious push into physical AI devices. (Ars Technica, The Verge)

1Password launched a browser integration that lets Claude use stored credentials for multi-step tasks like travel booking and account management without exposing secrets to Anthropic's models, via a "zero-exposure security framework"—an early template for how agents can act on the web safely. (The Verge)

TikTok is testing an opt-in AI likeness-detection tool: creators verify their identity through Jumio, then TikTok scans for deepfakes of them—mirroring a similar YouTube tool now open to all adults. (The Verge)

Google is renaming NotebookLM to Gemini Notebook, keeping it standalone while integrating it more deeply into Gemini and Search. Anthropic launched Claude for Teachers and localized Claude pricing for India, its biggest market after the U.S. (The Verge, Anthropic, TechCrunch)

Nokia launched an NVIDIA-based AI-RAN platform it bills as an industry first—claiming greater than 20% spectral efficiency now, with 50% targeted by 2027 and more than 100% by 2028. The claim is contested: Ericsson shipped a GPU-free AI-in-RAN subscription in June across 15+ deployments, and NVIDIA already holds roughly 3% of Nokia from a $1 billion October 2025 stake. (AI News)

Google-backed FireSat wildfire satellites reached initial operational capability—3 satellites launched July 7 (built by Muon Space) delivering twice-daily global coverage and able to detect fires as small as 5×5 meters. Backed by $15 million+ from Google and $26 million from the Bezos Earth Fund, the system is projected to prevent more than $1 billion in damage and roughly 22 million tons of carbon, arriving as wildfire smoke chokes the U.S. and Canada. (Ars Technica)

Thinking Machines Lab released Inkling, its first from-scratch model: a 975-billion-parameter multimodal mixture-of-experts (41 billion active) with 1M context and full Apache 2.0 weights. The lab concedes it isn't the strongest model available, positioning it instead as a customization base—the concrete expression of its "future worth building is human" manifesto arguing for distributed, user-tunable AI over frozen centralized defaults. (MarkTechPost, Thinking Machines)

Mistral released Robostral Navigate, an 8-billion-parameter model that lets robots navigate complex environments using a single RGB camera—no lidar or depth sensor required—lowering the hardware bar for autonomous mobility. (MarkTechPost)

Linus Torvalds told critics of AI-assisted coding in Linux to "fork it, or just walk away," in a thread centered on Sashiko, an agentic kernel code-review system that claims to independently find 53.6% of the bugs humans end up fixing—a notable endorsement from one of open source's most exacting gatekeepers. (Ars Technica)