Google I/O dropped Gemma 4 12B, agent infrastructure, and on-device browser AI in a single wave. Meanwhile, a new executive order moves frontier AI safety evaluations behind a national security curtain.
Gemma 4 12B is a dense, encoder-free multimodal model released alongside Antigravity 2.0, Google's new agent-first infrastructure layer. The strategy: give away weights, lock in the platform.
At least 18 foreign nationals died in a Delhi hotel fire reported on June 3. Russia separately demanded France release a detained tanker captain tied to a sanctioned shadow fleet vessel.
The Prompt API hits stable in Chrome 148, letting developers call Gemini Nano locally for image, audio, and text inference with zero server round-trips. On-device AI just became a distribution play.
Google published Gemma 4 12B on June 3, a dense multimodal model featuring a unified, encoder-free architecture that feeds images, audio, and text directly into the LLM backbone without separate vision or audio encoders. The release targets developers building local AI applications and represents a significant step...
Continue reading arrow_outwardGoogle released Gemma 4 12B as an open model and unveiled Antigravity 2.0, its next-generation agent infrastructure, turning I/O into the company's most aggressive open-source push yet.
A hotel fire in Delhi killed 18 foreign nationals, raising urgent questions about building safety enforcement in India's capital.
Sentiment sits at Extreme Fear as geopolitical friction, new U.S. sanctions on Iranian crypto exchanges, and uncertainty around AI regulation weigh on risk appetite.
Chrome 148 now ships a multimodal on-device Prompt API, quietly turning billions of browsers into local AI inference endpoints without a single cloud call.
Gemma 4 12B lands alongside Antigravity 2.0, Google's revamped agent infrastructure layer. Together they signal a strategy shift: give away the model weights, then monetize the orchestration and cloud tooling that make agents production-ready. Microsoft's counter-move with proprietary models sets up a direct philosophical clash over how the next wave of AI gets deployed.
A new executive order establishes a classified benchmarking process for frontier AI systems, moving safety evaluation behind a national security curtain. The move could accelerate government adoption by giving agencies vetted performance data, but it also removes public scrutiny from the metrics that determine which models are deemed safe. Transparency advocates are already warning this creates a two-tier accountability system.
Chrome 148's multimodal Prompt API lets web developers call local AI inference directly from the browser, no server round-trip required. This is a distribution play with few parallels: Chrome's roughly 3.4 billion installations mean on-device AI could reach more endpoints overnight than any single model provider serves today. Privacy-sensitive use cases in healthcare, finance, and education stand to benefit most immediately.
The Zero Dollar Inference
Read full analysis arrow_outwardGoogle published Gemma 4 12B on June 3, a dense multimodal model featuring a unified, encoder-free architecture that feeds images, audio, and text directly into the LLM backbone without separate vision or audio encoders. The release...
Google I/O 2026 introduced Antigravity 2.0, described as agent-first infrastructure, alongside WebMCP, a protocol for exposing website tools to browser-based AI agents. Chrome DevTools is being rebuilt with an AI Assistance panel...
Chrome 148 brings the Prompt API to stable, powered by Gemini Nano and now supporting multimodal inputs across images, audio, and text. The same update introduces Gemma 197M, a lightweight model designed for fast on-device inference....
The June 2 executive order, titled "Promoting Advanced Artificial Intelligence Innovation and Security," directs federal agencies to establish a classified benchmarking process for evaluating frontier models' cyber capabilities. It...
Microsoft has released its own AI models to compete directly with OpenAI and Anthropic, marking a strategic shift from its role as primarily a distribution partner. The move positions Microsoft as both a platform provider and a model...
OpenAI unveiled new features for GPT-Rosalind, expanding its functionalities and usability across various applications. The announcement, posted to OpenAI's blog on June 3, follows the company's recent Codex expansion and model pruning...
At least approximately 18 foreign nationals were killed in a hotel fire in Delhi, India, according to CNN-Newsapproximately 18 reporting cited by US News on June 3. Details on the nationalities of the victims and the cause of the blaze...
Russia on June 3 demanded that France release the captain of a tanker it says was detained on false charges, according to Reuters. The dispute follows France's seizure of a sanctioned Russian shadow fleet vessel reported on June 1....
Italian prosecutors cleared a contested presidential pardon after a formal probe, Reuters reported on June 3. The pardon had drawn political scrutiny and legal challenges in recent weeks. The decision effectively ends the investigation,...
The U.S. Treasury Department announced sanctions targeting four Iranian cryptocurrency exchanges as part of the broader economic pressure campaign tied to the Strait of Hormuz crisis. The move expands the sanctions front beyond oil into...
Prototype with the Prompt API. Open Chrome 148 stable, enable the Prompt API flag if needed, and build a multimodal input flow that classifies images or transcribes audio entirely on-device. Measure latency against a cloud endpoint.
Benchmark Gemma 4 12B on your own data. Pull the weights from Hugging Face, run inference on a domain-specific eval set, and compare quality against your current provider. Document where the 12B dense model wins or loses versus larger sparse alternatives.
Audit your agent architecture against Antigravity 2.0. Review Google's WebMCP protocol docs and map your existing tool-calling patterns to the new browser-agent surface. Identify one workflow you can expose as a WebMCP endpoint this week.
Gemma 4, Antigravity 2.0, and Chrome 148's Prompt API form a coherent stack: open model, agent platform, browser distribution. Builders who treat these as isolated launches will miss the strategy. Meanwhile, classified AI benchmarks mean the public scoreboard for frontier safety just went dark, raising the stakes for independent evaluation. The window to shape how open and on-device AI gets adopted is right now. Build on the open layers before the platform gravity sets in.
AI analysis, world news, markets, and tools. One briefing, delivered free.