Welcome back! An AI agent in San Francisco just hired real people, signed a three-year lease and opened a boutique with a $100K budget. Google released a free app that lets you run its latest models on your phone with no internet. Stanford's annual AI report says adoption has passed 53% but trust has fallen to 31%. And Meta is building a photorealistic clone of Mark Zuckerberg to take meetings for him.
In today’s Generative AI Newsletter:
AI employer: What happens when you give an AI agent a budget, a lease and permission to hire people?
Google on-device AI: What does it mean when you can run a frontier model on your phone with no internet and no account?
Stanford AI Index: Over half the world uses AI now, so why does almost nobody trust it?
Zuckerberg clone: What does it mean when the CEO of the world's biggest social network wants an AI to do his job?
Latest Developments
An AI Agent Just Hired Humans and Opened a Shop
Andon Labs gave an AI agent called Luna a $100K budget, a credit card and a three-year lease on a retail space in San Francisco. Luna's only instruction was to turn a profit. It created a boutique concept from scratch, posted job listings, conducted interviews over Zoom with its camera off and opened the store.
The details:
The setup: Luna runs on Claude Sonnet 4.6 for reasoning and Gemini 3.1 Flash-Lite Preview for voice. It monitors the store through screenshots taken from security cameras. Andon Labs' previous experiment was an AI vending machine at Anthropic's offices. This is a major step up in scope.
The hiring process: Luna posted listings on TaskRabbit, reviewed applicants and scheduled interviews. All of this happened autonomously with no human direction.
The failures: When hiring a painter, Luna accidentally selected Afghanistan on TaskRabbit's country dropdown. It also botched the opening weekend staff schedule. These are the kinds of errors that would get a human manager a serious conversation.
Real-world agent experiments keep producing the same pattern: impressive in ambition, broken in execution. Luna can reason through business strategy but cannot navigate a dropdown menu. That gap is going to close. Every model upgrade and memory improvement gets it closer. A version of Luna that does not make these mistakes is probably only a generation or two of models away, and that is when the conversation about AI as employer rather than employee becomes very real.
Special highlight from our network
Agentic AI is moving into production.
Most teams are still figuring out how.
Steve and Siva are breaking down what’s actually happening right now:
• Where enterprise AI adoption stands
• Why agentic systems are gaining traction
• Single vs multi-agent: what works in practice
• What the next phase of AI deployment looks like
If AI is on your roadmap, this matters.
Special highlight from our network
New research from OpenText and the Ponemon Institute shows that more than half of enterprises are using GenAI.
Yet only 1 in 5 have assessed security risks or fully deployed these systems.
Instead of driving efficiencies, many organizations are facing new challenges, especially around compliance and security, due to a lack of controls.
Read the findings to see how missing security, governance, and data foundations impact AI strategy, and how to address them without slowing innovation.
You Can Now Run Google's Latest AI on Your Phone for Free
Google released AI Edge Gallery, an app that lets you download and run its latest models directly on your phone. No account, no subscription, no internet required after setup. The models run entirely on-device and no data leaves your phone.
The details:
How it works: Download AI Edge Gallery from the App Store or Google Play, open it, tap AI Chat and download a model. Gemma 4 E2B is around 2.5 GB. Once downloaded, everything runs locally.
The thinking feature: Toggle Thinking mode in settings and the model shows its step-by-step reasoning before answering. This was a cloud-only capability until very recently.
The limitations: The app does not save chat history. Close a thread and the conversation is gone. But you can tap Agent Skills for extras including a restaurant picker, Wikipedia lookup, interactive map and QR code generator.
On-device AI that works without an internet connection changes the accessibility equation entirely. This is not a research demo. It is a free consumer app running a capable model on hardware most people already own. The fact that it includes reasoning mode, something that required server-grade compute a year ago, tells you how fast the efficiency gains are compounding.
Over Half the World Uses AI but Almost Nobody Trusts It
Stanford HAI released its 2026 AI Index, a 400-page annual report tracking AI's progress across research, industry, policy and public sentiment. The headline numbers: 53% global adoption, which means AI has reached more than half the world's population faster than the PC or the internet. But public trust sits at just 31% and is falling.
The details:
The expert-public divide: Nearly three quarters of AI researchers are optimistic about AI's impact on jobs. Only 23% of the general public agrees. That is the widest gap the report has ever recorded.
The US paradox: America builds most of the world's AI but ranks 24th in actually using it at 28.3% adoption. Singapore, the UAE and most of Southeast Asia are ahead.
The jobs data: Developer employment for ages 22 to 25 fell nearly 20% since 2024 even as headcounts for older engineers grew. Firm surveys say planned cuts will accelerate.
The benchmark race: China has nearly closed the gap with the US on AI performance benchmarks. Anthropic's top model leads by just 2.7%. Meanwhile, the number of AI researchers moving to the US dropped 89%.
The expert-public divide is the number that should worry the industry most, especially given the anti-AI climate that turned violent at Sam Altman's home last week. AI insiders see a productivity boom. Regular people see their entry-level jobs disappearing and a technology they did not ask for being imposed on them. 31% trust is not a communications problem. It is a legitimacy problem, and no amount of optimistic messaging from inside the industry is going to fix it without concrete evidence that the benefits are being shared.
Meta Is Building a Digital Clone of Mark Zuckerberg
Meta is developing a photorealistic AI version of its CEO trained on his mannerisms, publicly available statements and internal thinking on company strategy. The clone is intended to represent Zuckerberg in meetings. This comes roughly a month after reports that he is also working on a separate OpenClaw-style agent to help him manage his workload.
The details:
The technology: The AI clone is described as photorealistic and trained to replicate Zuckerberg's tone, expressions and decision-making patterns.
The scope: This is not a novelty demo. The intention is for the clone to participate in actual internal meetings on Zuckerberg's behalf.
The pattern: Meta building this for its own CEO while simultaneously selling AI workplace tools to enterprise clients is going to make for an interesting product marketing challenge. 70,000 employees will be interacting with this thing.
If your CEO can be cloned well enough to sit in meetings, that tells you something about what those meetings are actually worth. The more interesting question is what happens when employees start optimising their communication for the AI version rather than the real one, and whether anyone can tell the difference.
Tool of the Day: Harvey Agents
Harvey just launched Agents, autonomous AI bots that execute full legal workflows including research, memos, diligence reports and slide decks across 13 legal domains. This is not a copilot sitting inside a document. These are agents that take a brief and run with it, producing finished work product.
Try this yourself: Visit harvey.ai and explore the Agents product page. If you work in legal, compliance or deal execution, request access and test it against a real workflow. The interesting thing to watch is whether the output holds up under partner-level review or whether it is associate-grade work that still needs heavy editing.
Light Bytes
OpenAI memo targets Anthropic: The chief revenue officer called Anthropic a "single-product company" with inflated revenue in a leaked internal memo published by The Verge.
Codex gets web browsing: OpenAI is testing a browser feature inside Codex as part of its push to turn the tool into a unified development environment.
Microsoft exploring always-on agents: Persistent OpenClaw-style agents for 365 Copilot that run around the clock inside Office apps, with a preview expected at Build in June.
Lovable adds payments: Users can now sell products directly from AI-built sites by describing the item and price in chat.
SoftBank backs Japanese AI mega-model: A new company backed by NEC, Honda, Sony and five other firms will build a 1-trillion-parameter physical AI model.
Workshop Labs joins Thinking Machines: Mira Murati's lab acquires the team and its personalised AI stacks.
