Welcome, Enthusiasts!
If you’ve started asking AI about life choices as much as work, you’re in the majority. OpenAI says 73% of ChatGPT use is now personal, while Claude still skews toward code. Google is testing how private training really feels with VaultGemma. Microsoft is planting Copilot Chat inside Word, Excel, and friends. And Google’s quantum team just trained generative models on a 68-qubit chip, a small glimpse of the next platform shift.
📌 In today’s Generative AI Newsletter:
OpenAI + Anthropic 73% personal use; Claude coding
Google VaultGemma privacy model
Microsoft Copilot Chat in 365
Google Quantum 68 qubit generative demo
📊 OpenAI and Anthropic Reveal Global AI Habits
Image Credit: OpenAI
OpenAI and Anthropic have both published new usage data, offering a rare look at how millions of people are actually using ChatGPT and Claude. The findings highlight personal adoption surging, coding dominating Claude’s usage and stark regional divides in where the technology is growing fastest.
Patterns from the reports:
ChatGPT turns personal: Personal use grew from 53% of messages in mid-2024 to 73% in 2025, with advice and decision support now outweighing content creation.
Claude remains technical: Coding and professional problem-solving remain the dominant focus for its users.
Global spread diverges: ChatGPT adoption in low and middle-income countries is expanding at four times the rate of wealthy nations, while Claude’s growth remains centered in developed regions.
Shift in use cases: Users on both platforms are relying more on AI for information-seeking and search functions instead of direct output generation.
The reports make clear that AI is no longer confined to professional tasks but is becoming embedded in daily choices, conversations and problem-solving across very different parts of the world. When millions are as likely to ask an AI for life advice as they are to use it for work, it shows just how quickly the technology is finding its place in the rhythms of everyday life.
🔒 Google Releases VaultGemma, a Privacy-Preserving AI Model
Image Credit: Google
Google has released VaultGemma, a 1B-parameter model trained with differential privacy. It is the largest open model of its kind, developed with DeepMind and serves as both a research benchmark and a public experiment in how far privacy-preserving training can go.
What the research shows:
A new scaling map: Google derived formulas linking compute, data and privacy budgets, offering a structured way to train private models without guesswork.
Measured performance: VaultGemma performs on par with non-private models from five years ago, quantifying the cost of privacy in current methods.
Concrete privacy guarantees: Sequence-level protections prevent memorization, with tests confirming no detectable reproduction of training data.
Trade-offs exposed: Effective DP training demands far larger batch sizes and higher compute, raising the price of building models at scale.
VaultGemma is less a breakthrough in capability than a line in the sand for what privacy-first AI currently costs. By making the weights public, Google shifts the question to the community: how fast can researchers close the gap between privacy and performance and who will pay the compute bill along the way?
📝 Microsoft Expands Copilot Chat Across 365 Apps
Image Credit: Microsoft
Microsoft is rolling out Copilot Chat and AI agents inside Word, Excel, PowerPoint, Outlook and OneNote, bringing a unified sidebar that embeds the assistant directly into the apps millions rely on at work. The move makes Copilot a constant presence across documents, spreadsheets, presentations and email.
Key updates in this release:
In-app chat: Copilot can now answer questions and generate content directly in the sidebar, tailored to the file currently open.
Expanded inputs: Users can upload multiple images in chat, reference files with “/” search and draft longer prompts in an enlarged input box.
Agent support: Licensed Copilot users gain access to specialized Researcher and Analyst agents for data exploration, project ideas and visualization.
GPT-5 integration: The rollout includes Microsoft’s latest model upgrade, offering faster, more consistent responses.
For everyday users, Copilot Chat comes at no additional cost, though premium licenses unlock deeper reasoning across an organization’s files and communications. This expansion makes Copilot harder to ignore. By embedding it across the Microsoft 365 suite, the company is betting that workers will treat AI not as a separate product but as a built-in layer of productivity software.
⚛️ Quantum Processors Take on Generative AI in New Google Study
Credit: Research Paper
Researchers from Google Quantum AI, Caltech and Purdue have published a new study showing how generative models can be trained and sampled efficiently on a quantum processor. Using a 68-qubit superconducting chip, they demonstrated tasks that are very difficult for classical computers.
What they achieved:
Efficient learning: The models are structured to avoid common training failures that affect quantum machine learning.
Practical methods: The team introduced tools like an exact mapping between circuit types and a “sewing technique” that simplifies training.
Quantum edge: Certain tasks, such as finding simplified circuit versions, could only be performed efficiently on a quantum device.
Measured outcome: The results provide a concrete case where generative modeling benefits directly from quantum hardware.
The study is an early but important step toward practical quantum machine learning. It shows how generative models can begin to take advantage of quantum resources, offering researchers a path to explore problems that classical systems cannot handle with the same efficiency.
🚀 Boost your business with us. Advertise where 13M+ AI leaders engage!
🌟 Sign up for the first AI Hub in the world.
