AI Pulse: Daily Digest — April 3, 2026
Summaries are AI-generated. Click through to read the original reporting.
Google has released Gemma 4, its first major update to the open model family in a year, introducing small, fast, and multimodal models optimized for local on-device execution. Notably, Google is switching to the more permissive Apache 2.0 license, making the models more accessible for commercial and research use. NVIDIA has already announced acceleration support for Gemma 4 on RTX hardware for local agentic AI workflows.
Read more →Microsoft's MAI group, formed just six months ago, has released three new foundational models covering voice transcription, audio generation, and image generation — a direct challenge to rivals like OpenAI and Google. The launch coincides with CEO of AI Mustafa Suleyman shifting his focus toward a longer-term "superintelligence" strategy following a major company restructuring in March. Suleyman described the new models as the first fruits of Microsoft's ambition to build frontier AI capabilities in-house.
Read more →OpenAI has purchased TBPN, a cult-favorite Silicon Valley talk show that broadcasts live daily and has hosted Sam Altman and executives from Meta, Microsoft, Palantir, and Andreessen Horowitz. The show will continue to operate independently under the oversight of OpenAI's chief political operative Chris Lehane. The acquisition signals OpenAI's growing interest in media and narrative influence beyond its core AI products.
Read more →Anthropic's attempt to contain the spread of leaked Claude Code client source code via DMCA notices inadvertently took down thousands of legitimate GitHub forks and repositories. The company acknowledged the error and retracted the bulk of the takedown notices, but the incident highlights the difficulty of controlling leaked proprietary code once it spreads across open platforms. The leaked code itself revealed details about Anthropic's roadmap, including a persistent agent mode, a stealth "Undercover" feature, and an unreleased virtual assistant called Buddy.
Read more →Google is bringing its most capable generative AI models — Veo 3.1 for video and Lyria 3 for music — into Google Vids, with high-quality video generation now available at no cost. A new "directable avatar" feature lets users customize and instruct AI presenters through natural language prompts, enabling more personalized video creation. The update positions Google Vids as a serious AI-native competitor in the workplace video creation space.
Read more →Security researchers have demonstrated two new Rowhammer-style attacks — dubbed GDDRHammer and GeForceHammer — that exploit GDDR memory in Nvidia GPUs to compromise not just the GPU but the host CPU as well. The attacks can grant an attacker complete control over affected machines, raising serious concerns for AI data centers and gaming systems alike. This is a significant hardware-level vulnerability with no simple software patch available.
Read more →Popular AI meeting note-taker Granola has a significant privacy issue: despite claiming notes are "private by default," any note is viewable by anyone who obtains its link. The app also uses notes for internal AI model training unless users explicitly opt out — a setting that is not prominently surfaced. Users are advised to review their privacy settings immediately, especially given the sensitive nature of meeting content.
Read more →Multiple Baidu Apollo Go robotaxis simultaneously froze in the middle of streets in Wuhan, China, reportedly trapping passengers inside vehicles and contributing to at least one accident in the resulting traffic gridlock. Local police confirmed receiving numerous reports of the autonomous vehicles becoming immobile on highways and city roads. The incident is a high-profile setback for autonomous vehicle deployment at scale and will likely intensify regulatory scrutiny.
Read more →Meta's upcoming Hyperion AI data center will draw power from 10 newly built natural gas plants, consuming enough energy to power the entire state of South Dakota. The revelation intensifies criticism of Big Tech's carbon footprint at a time when AI infrastructure demand is accelerating energy consumption globally. The scale of the commitment underscores the enormous and growing power requirements of frontier AI model training and inference.
Read more →Startup Cognichip has secured $60 million in funding to develop AI systems capable of autonomously designing semiconductor chips, claiming it can cut chip development costs by over 75% and halve the time to market. The company is targeting the expensive and slow traditional chip design pipeline, which has become a bottleneck as demand for AI-optimized silicon surges. If successful, AI-designed chips could dramatically accelerate the hardware side of the AI arms race.
Read more →Google is rolling out two new inference tiers for the Gemini API — Flex for cost-sensitive workloads and Priority for low-latency, mission-critical applications. The tiered pricing model gives developers more granular control over the cost-versus-speed tradeoff when building on Gemini. This move mirrors similar strategies from OpenAI and Anthropic and signals growing maturity in the enterprise API market.
Read more →Switzerland's finance minister has filed a criminal complaint against xAI's Grok chatbot, alleging that its AI-generated "roast" feature produced vulgar and defamatory content targeting women, including public officials. The lawsuit is one of the first government-level legal actions against a generative AI system's content output in Europe, and could set a precedent for how AI humor features are regulated. Elon Musk has previously praised Grok's roasting capabilities publicly.
Read more →A new wave of gig workers around the world — including medical students in Nigeria — are earning income by recording themselves performing physical tasks at home, generating training data for humanoid robot AI systems. The work involves strapping cameras to their heads and mimicking robot-like movements to capture motion data that robotics companies use to teach machines dexterous skills. The trend raises questions about labor conditions, data ownership, and the hidden human cost behind the humanoid robotics boom.
Read more →After seven years of development, California startup Kintsugi is shutting down following its failure to obtain FDA clearance for an AI system designed to detect depression and anxiety from a person's voice. The company will release much of its technology as open-source, and some components may be acquired or repurposed by other organizations. The closure illustrates the steep regulatory hurdles facing clinical AI startups and the high-stakes nature of FDA approval timelines for health-focused AI products.
Read more →Elgato's Stream Deck 7.4 software update introduces Model Context Protocol (MCP) support, enabling AI assistants like Claude, ChatGPT, and Nvidia G-Assist to discover and trigger Stream Deck actions autonomously. The update is a practical example of MCP's growing role as a standard interface for connecting AI agents to real-world tools and devices. It reflects a broader trend of AI assistants moving beyond conversation into direct control of physical and software environments.
Read more →Summaries are AI-generated. Click through to read the original reporting.