Good morning. It’s Monday, March 10th.

On this day in tech history: Netscape announced the release of its third-generation web browser, Netscape Communicator, which included features like email, news, and HTML editing tools. This was a significant development during the browser wars of the late 1990s, though Netscape eventually lost dominance to Internet Explorer.

In today’s email:

• Manus Autonomous AI Agent

• Google’s New Gemini Models Set For Wednesday Release

• LumaAI’s Ray 2

• 3 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

Manus introduced as ‘world’s first’ first fully autonomous AI agent but it’s not a DeepSeek repeat

Manus AI, a Chinese autonomous agent, has been confirmed to integrate Claude 3.5 Sonnet, fine-tuned Qwen models, and open-source frameworks like Browser Use. Chief researcher Yichao “Peak” Ji detailed its multi-agent architecture, where an executor agent mediates user interactions while remaining siloed from the planner and knowledge agents—a design that constrains context length and limits jailbreak exploits. The team is currently testing Claude 3.7 Sonnet, which Ji notes demonstrates improved performance.

Manus claims it outperforms OpenAI’s deep research on the GAIA benchmark, though no independent validation backs the bold assertion. Operating in standard and high-performance modes, the system suggests dynamic reasoning optimizations similar to OpenAI’s Operator.

While the company leans on open-source foundations, it keeps key mechanics under wraps. With select components set for release, Manus positions itself as a wildcard in the AI agent race alongside OpenAI’s Swarm and Google’s Mariner. Read more.

Google plans to release new Gemini models on March 12

Google is gearing up to drop new Gemini models on March 12, with system code hinting at upgraded intelligence and tighter app integration. Expected additions include Flash 2.0 Thinking, built for structured AI reasoning, and Flash 2.0 Thinking with Apps, which dynamically taps into Gmail, YouTube, and other Google services. A personalization layer is also in the works, potentially allowing Gemini to tailor responses based on user behavior.

Meanwhile, Gemini Live is inching closer to real-time screen sharing and video analysis, features that could roll out under Gemini Advanced. While these updates suggest a sharper, more responsive AI system, past delays mean nothing is set in stone. Gemini’s fragmented experience has drawn criticism, and this overhaul seems like Google’s move to streamline its AI ambitions. Read more.

LumaAI drops Ray 2 Flash with 3x speed, 3x savings, and frontier model efficiency gains

Ray just launched Ray2 Flash, a model built to triple speed, slash costs, and cut queue times without compromising quality. Packing text-to-video, image-to-video, audio, and control capabilities, Flash brings frontier-level AI generation to all subscribers at a fraction of the usual wait.

Its new distillation technique retains full model capabilities while running 500 percent more efficiently. Instead of pocketing all the gains, Ray is reinvesting them into higher throughput, ensuring users get results faster. Flash is now live. Read more.

• Scientists find RNA isn’t just a messenger but a key player in DNA repair

• Nadella says AI models are commodities and the real battle is full-stack dominance

• ChatGPT nears 4 billion visitors but growth plateaus

• Microsoft is reportedly testing independent AI models to loosen OpenAI’s grip on its AI stack

• Ex-DeepMind researchers raise $130 million to turn AI into its own engineer

• Apple delays major Siri AI upgrades to 2026

• AI-driven discovery simplifies Einstein’s “spooky action,” advancing secure quantum networks

• "Highlighted Chain of Thought" prompting boosts LLM accuracy and verifiability

• Foxconn has its own AI now

• Grok can now also be marked in posts on X – and replies

• Nvidia's GTC 2025 will feature AI's brightest minds and draw 25K attendees in person

• Figure AI ships another batch of humanoid robots as it moves to new HQ

• ElevenLabs partners with Google Cloud to bring AI audio to enterprise

📄 Unified Reward Model for Multimodal Understanding and Generation

📄 Sketch-of-Thought: Efficient LLM Reasoning with Adaptive Cognitive-Inspired Sketching ​​

📄 START: Self-taught Reasoner with Tools  

📄 Token-Efficient Long Video Understanding for Multimodal LLMs

📄 LLMVoX: Autoregressive Streaming Text-to-Speech Model for Any LLM

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!