Unlock AI-powered productivity

HoneyBook is how independent businesses attract leads, manage clients, book meetings, sign contracts, and get paid.

Plus, HoneyBook’s AI tools summarize project details, generate email drafts, take meeting notes, predict high-value leads, and more.

Think of HoneyBook as your behind-the-scenes business partner—here to handle the admin work you need to do, so you can focus on the creative work you want to do.

Try HoneyBook for free

Today:

• Google Unveils Sweeping I/O Upgrades

• Microsoft Showcases New AI Agent Tools

• Manus Launches Paid Team Plan

• Aronofsky, DeepMind Team On AI Films

• Apple To Open On-Device AI

Google Unveils Sweeping I/O Upgrades

Google’s I/O 2025 unleashes sweeping upgrades: Veo 3 makes videos with voices; Imagen 4 improves pictures; Flow turns plain text into movie clips; Lyria 2 composes music. Gemini app now crafts infographics and shares screens; new Pro and Ultra plans open more features. Gemini in Chrome sums up pages; AI Mode revamps Search. Faster Gemini 2.5 models and freely-shared Gemma 3n widen access. Beam creates 3-D calls for remote teams.

Why this matters

1. All-in-one creativity boost – Video, image, music and filmmaking tools arrive in a single ecosystem, pushing multimodal creation into everyday use.
   

2. Rise of “do-it-for-me” agents – Chrome help, Agent Mode and AI Search preview assistants that handle multi-step tasks, raising expectations across the industry.
   

3. Wider access and competition – Open models, student deals and clear pricing broaden who can build with cutting-edge AI, likely spurring fresh research and startup innovation.

Microsoft Showcases New AI Agent Tools

At Build 2025, Microsoft unveiled new tools to speed research and business work with AI agents. The Discovery platform links agents to a knowledge graph to draft hypotheses, run simulations, and trace sources. Copilot Studio now hosts cooperating agents that can control software and even WhatsApp. Windows and Azure AI Foundry bridge on-device and cloud models while the open Model Context Protocol aims to let any agent exchange data safely everywhere.

Why this matters

1. End-to-end agent workflows – Discovery and Copilot Studio show agents handling full research or business tasks, proving AI can move beyond chat into real, connected work.
   

2. Open standards beat silos – MCP and NLWeb aim to let agents and websites talk freely, reducing lock-in and encouraging a shared “agentic” web.
   

3. Local-plus-cloud flexibility – Windows and Azure AI Foundry give developers one path from laptop tests to cloud scale, widening access to advanced models.

Manus Launches Paid Team Plan

Manus launched Manus Team, a $39-per-seat plan for small groups, requiring five seats. Teams share 19,500 task credits, access beta features, dedicated servers, and peak-hour priority. Two tasks run at once, with “high-effort” mode boosting reliability. Viral in March, the startup recently raised $75 million at a $500 million valuation, launched mobile apps, and plans global expansion to Japan and the Middle East. Manus now uses Claude 3.7 Sonnet and Qwen models.

Why this matters

1. Commercial viability of agents – A clear pricing model shows how agentic platforms can turn buzz into repeat revenue for real teams.
   

2. Growing model mix – Reliance on both Anthropic and Alibaba models signals a shift toward flexible, multi-provider AI stacks.
   

3. Global small-business focus – Affordable team plans and region expansion push advanced AI automation beyond big enterprises, widening the adoption base.

🧠RESEARCH

Chain-of-Model Learning for Language Model

This paper introduces Chain-of-Model (CoM), a new way to build language models that makes them easier to scale and more efficient. By linking layers in a chain-like structure, models can grow over time and run at different sizes depending on the task. Their version, CoLM, matches Transformer performance while offering better training and inference flexibility.

AdaptThink: Reasoning Models Can Learn When to Think

AdaptThink is a new method that teaches reasoning models when to "think deeply" and when to skip straight to the answer. By switching modes based on task difficulty, it cuts response time by over 50% and even boosts accuracy. It uses reinforcement learning to train models for smarter, faster reasoning.

AdaCoT: Pareto-Optimal Adaptive Chain-of-Thought Triggering via Reinforcement Learning

AdaCoT is a smart system that teaches language models when to use step-by-step reasoning and when to skip it. Using reinforcement learning, it balances accuracy with speed, cutting unnecessary reasoning on simple tasks. It reduced response length by 69% and still performed well on harder questions—saving time without sacrificing quality.

🛠️TOP TOOLS

Rose AI - Cloud-based data platform that revolutionizes financial data analysis and visualization through the power of generative AI

DupDub - AI-powered content creation platform designed to streamline the production of multimedia content

Gliglish - AI-powered language learning platform designed to help users improve their speaking and listening skills through interactive conversations

ClassPoint AI - Instantly generate engaging quiz questions from slide content

MyVocal AI - AI-powered platform that offers rapid voice cloning and text-to-speech capabilities

📲SOCIAL MEDIA

🗞️MORE NEWS

• Director Darren Aronofsky’s Primordial Soup is partnering with Google DeepMind to explore how AI can enhance storytelling. They’re producing three short films using cutting-edge video models, with the first premiering at Tribeca Festival 2025.

• Apple plans to let developers use its on-device AI models through a new software kit, starting with smaller local models. This move, expected to be announced at WWDC on June 9, mirrors Google’s Gemini Nano approach and aims to expand AI features in third-party apps without relying on cloud access.

• The Chicago Sun-Times published a summer guide filled with fake, AI-generated books and quotes from made-up experts. The paper says it didn’t approve the content, which came from a third-party provider, and is now revising its policies.