Welcome back to Daily Zaps, your regularly-scheduled dose of AI news ⚡️ 

Here’s what we got for ya today:

• 👂 Google adds audio search results

• 🗽 New York RAISE Act to prevent AI harm

• 🧐 Meta + Scale AI update

• 🧠 How Claude Research was built

Let’s get right into it!

Sign Up | Advertise | Tools

BIG TECH

Google adds audio search results

Google is experimenting with AI-generated Audio Overviews in its Labs, powered by the latest Gemini models. For complex search queries, users can click a "Generate Audio Overview" button to receive a roughly 40-second voiced summary drawn from top search results, featuring two engaging voices in a conversational style.

The feature allows playback speed adjustments and cites source webpages. While this is the first time Audio Overviews appear in Search, a more advanced version already exists in Google’s NotebookLM, enabling follow-up questions. Currently, the Audio Overviews in Search are available only as an opt-in experiment through Google Labs.

GOVERNMENT

New York RAISE Act to prevent AI harm

New York state lawmakers passed the RAISE Act, a bill aimed at preventing frontier AI models from companies like OpenAI, Google, and Anthropic from causing catastrophic harm, including deaths or billion-dollar damages, by establishing the nation’s first legally mandated transparency and safety reporting standards for large AI labs. The bill, supported by prominent AI safety advocates, targets only the biggest AI models trained with over $100 million in computing resources and requires reporting of safety incidents, with penalties up to $30 million for noncompliance.

Designed to avoid stifling innovation, especially for startups and academics, it differs from California’s vetoed SB 1047 by excluding kill-switch mandates and broader liabilities. While Silicon Valley pushes back, critics’ fears that companies might withdraw from New York are tempered by the state's large market, and sponsors maintain the law strikes a balance between safety and innovation as it heads to Governor Hochul’s desk.

BIG TECH

Meta + Scale AI update

Meta’s $14.3 billion investment in Scale AI, acquiring 49% and bringing CEO Alexandr Wang to lead a new “Superintelligence” unit, has disrupted the AI data industry. While boosting Meta’s AI leadership, the deal caused major Scale clients like OpenAI and Google—Meta’s rivals—to reduce their business with Scale, fearing data exposure. Competitors of Scale have seen a surge in demand as AI labs seek neutral partners for sensitive training data, highlighting a major reshuffling of AI data supply chains.

The industry is shifting from basic data labeling toward expert-generated data for advanced “reasoning” AI models, increasing the value of secretive, high-quality human training input. Meta’s stake in Scale raises concerns that it could gain access to rivals’ proprietary data, potentially accelerating its AI progress and reshaping the competitive landscape.

STARTUPS

How Claude Research was built

Claude’s new Research feature uses a multi-agent system where a lead agent coordinates specialized subagents to search and analyze information across the web, Google Workspace, and integrated tools in parallel, enabling dynamic, flexible exploration of complex, open-ended research tasks. This architecture excels at breadth-first queries and complex investigations by decomposing problems and running simultaneous searches, significantly outperforming single-agent approaches.

Key lessons from development include the importance of prompt engineering, clear delegation, efficient tool use, and managing token consumption—since multi-agent systems use many more tokens but provide much higher performance on demanding tasks. Evaluations combine automated LLM-based grading and human review to ensure quality, while production challenges like state management, debugging, and deployment require robust engineering practices. Despite complexity, multi-agent research systems help users solve challenging problems faster and more thoroughly, demonstrating scalable, reliable AI-driven research in real-world use.