Welcome, humans.

It’s official: the legendary AI researcher (former OpenAI, former Tesla) and arguably the industry’s biggest influencer, Andrej Karpathy (he coined the term “vibe coding”), has chosen a side and joined Anthropic and is now heading back to frontier R&D.

This makes sense if you’ve been paying attention; it’s not like he’s going to join OpenAI (he already left), and he’s too independent for Google or Microsoft. That, and by every measure it appears Anthropic is ahead in the big lab model race. If you’re gonna play… why not play for keeps?

There were so many good memes about this, including translating it for non-AI folks as the equivalent of Ronaldo joining Manchester United, the obligatory “what did Karpathy see?” (nice one Dan) and this one (which transitions perfectly into our main story of the day; thanks Trung!):

Here’s what happened in AI today:

• 😼 Google put Gemini agents across Search, Workspace, and Android.

• 📰 OpenAI added provenance tools for AI-generated images.

• 📰 Anthropic brought Claude to KPMG's 276,000 workers.

• 🍪 Google AI Studio now builds native Android apps.

• 📖 The post-transformer debate asks what AI stores away.

…and a whole lot more that you can read about here

Hey: Want to reach 700,000+ AI-hungry readers? Advertise with us! 

P.S: Love robots? We’re starting a new robotics newsletter! Sign up early here.

😺 Google Put Gemini Agents Everywhere at I/O

DEEP DIVE: Everything Google announced at I/O 2026

Google held its annual I/O conference yesterday, and they used it to make one thing clear: Gemini is becoming the layer underneath all of the apps people already use… whether they like it or not.

Here’s what happened, in brief (read the deep dive for everything):

• Gemini 3.5 Flash became Google’s new fast model for agents, coding, and long tasks.

• Gemini Spark introduced itself as a 24/7 personal agent that can work across Workspace apps.

• Search got information agents that monitor the web, plus mini apps for ongoing tasks.

• Workspace got voice features for Gmail, Docs, and Keep, plus Google Pics and AI Inbox.

• Antigravity 2.0 turned coding agents into a managed desktop, CLI, and SDK workflow, plus the ability to launch a team of parallel coding agents.

• AI Studio now lets you generate full Android apps.

…and a whole lot more than that you can read about in our deep dive.

Why this matters: The real story, though, was pretty simple: Google wants Gemini to stop feeling like a chatbot and start feeling like the operating layer across search, email, docs, shopping, video, glasses, coding, and app building. At some point, the assistant stopped being shoehorned into the apps and will now become the plumbing beneath it.

This matters for Google because the AI race has drifted away from “which chatbot answers best?” and toward “which company can turn answers into action?” Google’s answer is obvious: put Gemini inside everything, from Search, Gmail, Docs, YouTube, shopping, Android, coding, app building, and glasses.

Our take: Google has the distribution advantage, to be sure. ChatGPT and Claude are destination apps. Meanwhile, Google owns the end work surfaces: inbox, docs, browser, phone, YouTube, Search, shopping, and the developer stack. All the endpoints where the work you do with the AI in the middle actually touches the rest of the world.

That means Google can make agents feel less like a new habit and more like a new setting inside the tools you already open all day. If it works well, it will feel like your tools just getting better.

The hard part now is permission. A search engine that answers is useful. A search engine that monitors, books, shops, builds, writes, and edits needs trust.

If Google gets that right, the next AI interface may feel less like opening a chatbot and more like telling your computer what outcome you want. Which is how AI SHOULD work.

FROM OUR PARTNERS

Is your service strategy ready for the AI era?

The old service management playbook — endless tickets, manual triaging, and bolt-on chatbots — can't keep up in the age of AI. Atlassian's new white paper outlines a bold, AI-native approach to service that shifts teams from reactive to proactive, turns rich context into a competitive advantage, and puts human-AI collaboration at the center of every interaction. If you're rethinking how your organization delivers service in the AI era, this is worth the read.

Learn More

🎓 AI Skill of the Day: Use Gemini Like a Task Router to Help Best Use Gemini

Google now has Gemini everywhere. So the skill today: make Gemini pick the right surface before you start. Google’s own prompt guide says Gemini works best with direct, structured instructions, so give it a routing job first.

I want to use Gemini for this task: [task].

Route me to the best Google tool, then give me:
1. The safest first prompt
2. The files or inputs to attach
3. What should stay read-only
4. What success looks like
5. One follow-up prompt

Good rule of thumb: Start with reversible tasks: monitor, draft, compare, storyboard, or prototype before you let an agent send, buy, delete, or publish.

Have a specific skill you want to learn? Request it here. 

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

1. *Incogni - remove your personal data from the web so scammers and identity thieves can’t access it. Use code NEURON to get 55% off.

2. Google AI Studio now builds native Android apps from prompts and lets you pull in Workspace data; this is sick.

3. Stitch turns text, voice, code, or design files into app interfaces you can steer in real time, available globally today.

4. Antigravity 2.0 orchestrates parallel coding agents, background tasks, CLI workflows, and SDK access, included in Google AI plans.

5. Gemini Spark handles proactive tasks across Google products under your direction, rolling to trusted testers first and U.S. AI Ultra users next week.

6. Carbon processes DNA sequences so you can continue sequences, score mutations, and visualize proteins, available as a free demo.

7. Google Pics edits image objects, text, and Workspace visuals with precise controls, launching to Trusted Testers first and Pro / Ultra users this summer.

NEW from The Neuron: Our Recap of the Great Post Transformer Debate

Our most popular podcast interview to date on The Neuron AI Explained was with Zuzanna Stamirowska, CEO of Pathway and whose company invented one of the most credible “post transformer” model architectures to date (called BDH). So when Pathway hosted an entertaining boxing-themed debate between the transformer and so-called post-transformer models, you know we had to tune in and see what’s up.

Check out the video above, or read our recap of the debate here.

📰 Around the Horn

• OpenAI added C2PA conformance, SynthID watermarking, and an early public verifier for OpenAI-generated images.

• Anthropic and KPMG announced a Claude rollout for KPMG's 276,000+ employees and client-facing Digital Gateway platform.

• The European Commission released draft high-risk AI Act guidance and opened consultation through June 23.

• Greg Brockman described Guaranteed Capacity as discounted token availability for 1-3 year enterprise commitments.

• Claude Managed Agents added self-hosted sandboxes and MCP tunnels for safer enterprise API connections.

• Google DeepMind published Co-Scientist in Nature, showing a multi-agent research system that generated and lab-tested biomedical hypotheses.

• Standard Chartered said it would cut 7,000+ jobs over four years while using AI to replace “lower-value human capital.”

• A new Shai-Hulud wave reportedly compromised 600+ npm packages and targeted developer secrets across package pipelines.

FROM OUR PARTNERS

Build durable AI agents and workflows in TypeScript. Trigger.dev provides the orchestration for long-running tasks with zero timeouts, automatic retries, and deep observability. From streaming LLM responses to custom build extensions, get the control you need without the ops overhead.

Focus on your logic and deploy reliable agents that scale on demand.

📖 Midweek Wisdom

• The Anti-AI Trap - Manoel Horta Ribeiro argues that dismissing AI as pure hype can be as lazy as believing every vendor demo.

• François Chollet on agent memory - Chollet’s point is simple and useful: real tasks depend on what happened before, so agents need faithful long-term trajectory memory.

• AI will change the world without taking it over - Boaz Barak and Ben Edelman offer a useful antidote to “AI plays 3D chess” stories.

A Cat’s Commentary

P.P.S: Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.