Welcome back! Instinct, a viral AI agent startup, is giving an email address to its AI assistant. Sounds like a very normal product update until you dip deeper … the bot can now sign up for services, contact businesses, join group threads, and handle support requests on your behalf.
The dream is less inbox clutter. The nightmare is discovering your AI assistant has better follow-up etiquette than you.
If We've Reached AGI, Why Can't AI Even Spot Other AI?
For everyone claiming we've reached AGI, answer me this: how come AI models are so bad at spotting AI?
I've been testing a bunch of platforms that claim they can detect AI-generated content, and none of them are consistently reliable. AI images and videos that most humans would immediately recognize as generated regularly get flagged as "probably human-made" or just "inconclusive." And when you upload the same content directly to the latest multimodal models (the ones supposedly with advanced image and video understanding), they'll still confidently get it wrong.
I'm not the only one seeing this:
NewsGuard tested 5 leading AI image detection tools in May 2026. The worst performer misidentified 40% of real photos as AI-generated. Only two of the five got them all right.
Unless the content carries an embedded watermark or C2PA metadata, AI seems genuinely bad at recognizing its own output.
My POV: To be fair, humans aren't perfect at this either. But plenty of AI-generated images and videos feel obvious to most people the second they look at them. If AGI means intelligence that matches or exceeds the smartest humans across almost every meaningful task, how can we claim we're already there when AI still struggles with something a human can often do instantly?
Hit reply and let me know: do you think we're at AGI?
— Matt
Anthropic Researcher Leaves With Dire AI Warning
A senior Anthropic safety researcher quit the company with a stark warning: AI labs are racing toward systems they may not be able to control.
Alarm bells: Jacob Coxon, who worked on training AI systems at Anthropic and previously at OpenAI, said he left over what he described as a lax approach to safety. In his departure post on X, he accused the companies of “racing straight to self-improving superintelligence and gambling with our lives,” arguing that even many of the people building these systems believe AI could pose catastrophic risks this decade.
Not a lone voice: Evan Hubinger, who leads one of Anthropic’s AI safety teams, responded directly and agreed with Coxon’s characterization. Hubinger said he personally estimates the chance that AI could “kill all humans” within the next decade is greater than 1 in 10. He also said Anthropic does “not yet have a plan” for ensuring advanced AI stays safe and aligned.
Race dynamics: The warning lands as frontier labs are pushing toward more autonomous systems, preparing for IPOs, and dealing with fallout from rogue agent incidents and other high-profile safety concerns. The worry is not just that models will become more capable, but that competitive pressure will keep labs moving faster than their ability to understand, monitor, or contain them.
The bigger picture: The people sounding the alarm are not outside critics throwing rocks at AI from a distance. They are researchers inside the labs building these systems. That makes the warning harder to dismiss, but it also raises the uncomfortable question: can the same institutions building increasingly powerful models also develop the safety systems needed to control them? Right now, even some of their own researchers do not seem convinced.
Delivery Apps Are Turning Grocery Lists Into AI Orders
Instacart and Shipt both launched AI shopping assistants as delivery apps race to turn natural language into ready-to-buy carts. These features follow other AI shopping moves from Uber Eats and DoorDash this year.
Cart from a prompt: Instacart’s new assistant, Clementine, can turn a conversation, grocery list, recipe, or photo into a cart. Users can ask for things like budget-friendly kids’ lunches, high-protein dinners for two, or their usual grocery order. It can also personalize recommendations around dietary needs, surface deals, suggest cheaper alternatives, and convert handwritten lists or screenshots into shoppable carts.
Same race, same day: Shipt launched its own assistant, Ask Shipt, which can build carts from prompts like a Saturday tailgate for 25 people or a week of school lunches and snacks. Users can also upload a photo of a dish they saw online or at a restaurant, and the assistant will identify the ingredients and add them to a cart.
The bigger picture: Grocery delivery platforms are now competing on more than speed, selection, and fees. They are competing to become the default household planning tool—being there when you decide what to cook before checkout even happens. The features also help keep users inside Instacart, Shipt, DoorDash, or Uber Eats instead of asking ChatGPT or Claude what to make for dinner and then shopping somewhere else.
Hire a Team of Agents Your Whole Team Can Use
Hyperagent does what most local AI setups still struggle to do: lets your teammates share the specialized agents everyone spins up.
Stop reinventing the wheel to collaborate when Hyperagent lets you:
Build specific agents for distinct jobs that will work together and hand off tasks.
Message agents in Slack—stop rebuilding workflows on everyone’s local computer.
Transform your process from individual prompting to multi-agent execution.
Want $100 in free credits? Sign up for a paid subscription today.
Turn Customer Context Into CRM Actions
Via Lightfield
Lightfield is an AI-native CRM that captures, analyzes, and acts on customer interactions without manual data entry. It synthesizes emails, calls, and meetings into a living model of each customer, then uses that context to support pipeline, outreach, meeting prep, and deal revival.
How you can use it
Generate pipeline without manual CRM entry
Prepare meeting briefs from past interactions
Send personalized outreach
Revive stale deals with AI agents
Pricing: Paid
Create Games From Prompts
Via Crayon AI
Crayon AI lets anyone create playable web games by describing an idea in plain language. Users can type a genre, world, character, or concept, and Crayon generates a playable game without requiring coding or engine knowledge.
How you can use it
Create playable web games from text prompts
Generate 2D or 3D game concepts
Publish finished games to Crayon Arcade
Share games through a direct link
Pricing: Paid
Jobs, announcements, and big ideas
Meta's AI agent Muse climbs to No. 2 on the US App Store with 83,000 downloads.
OpenAI opens its Agents API in public beta, bringing Codex-powered cloud agents to developers.
Google expands Dreambeans, its personalized daily story app, to all US users on Android and iOS.
OpenAI debuts ChatGPT for Financial Services, pairing GPT-6 Astra with built-in financial data.
Anthropic publishes red-team research showing Claude models near superhuman performance on intelligence-targeting tasks.
ElevenLabs teams up with Universal Music Group on a licensed AI music platform for fan remixes.
Some really important—and scary—AI news came out this week 🚨
That’s a wrap! See you next week for more.
