Welcome back! Pocket raised $11M for a $129 AI note-taking gadget that sticks to the back of your phone and records real-world conversations.
I know what you’re thinking. “Matt, another wearable?” Hear me out: It’s sold more than 130,000 units since launching last year, which makes me think the “AI gadget” category may not be dead. Maybe it just needed to stop trying to replace your phone and start capturing the parts of life your phone misses.
P.S. The team will be off on Friday to celebrate July 4th. 🎇 We’ll see you back here next week for more.
What If the Next Big AI Wave Has Nothing to Do With Words?
Here's a story that flew a little under the radar this week, but I think it's a much bigger signal than most people realize.
Google just announced it's bringing SandboxAQ's Large Quantitative Models (LQMs) onto Google Cloud Marketplace, giving researchers access to a category of AI built for something completely different from chat: chemistry, biology, materials science, and drug discovery.
The core distinction:
Large Language Models (the ones powering ChatGPT, Claude, Gemini) are trained on text. They predict the next word in a sentence—great for writing, summarizing, and reasoning over language.
Large Quantitative Models are trained on numerical data and scientific equations from real-world lab experiments. They predict things like how strongly a molecule binds to a catalyst surface, or which drug candidates are worth testing in the lab, making them useful for non-text use cases.
LQMs are already integrated with Claude, and now they’re coming to Gemini. The first two models hitting Google Cloud Marketplace are AQCat (Q3 2026, for materials and catalyst discovery: think batteries, semiconductors, sustainable fuels) and AQPotency (later this year, for drug discovery). Researchers can prompt in plain English and have Gemini hand off the actual scientific calculations to a physics-grounded model underneath. The pattern is basically: LLMs as the interface, LQMs as the engine.
A couple of things stand out about why this matters more than it looks on the surface:
SandboxAQ was originally spun out of Alphabet, is chaired by former Google CEO Eric Schmidt, and just landed a $500 million CHIPS Act award.
SandboxAQ says LQMs are aimed at the $50+ trillion global quantitative economy—biopharma, energy, advanced materials, and financial services.
My POV: We've spent the last three years teaching AI how to communicate like humans. The next phase is teaching AI how to understand the physical world. If that lands, the impact doesn't show up as better emails or spreadsheets. It shows up as faster drug discovery, better batteries, new materials, more efficient energy systems, and breakthroughs across industries that collectively represent trillions of dollars in economic activity.
Domain-specific AI is quietly where some of the biggest business opportunities are starting to emerge. If you're thinking about where AI goes after the current chatbot moment, this is the lane to watch.
— Matt
Congress Wants to Ban AI Companies From Selling Your Health Data
Via New York Times
Lawmakers are preparing a new version of the Health and Location Data Protection Act that would ban the sale of Americans’ health and location data to brokers, including information people share with AI chatbots.
What’s changing:
The original version of the bill was introduced in 2022 and focused on stopping data brokers from collecting and selling health and location data.
Over the past several years, Sen. Warren and Rep. Scanlon continuously reintroduced the bill into subsequent congressional sessions to tighten data privacy loopholes.
The updated version expands the ban to companies selling data to brokers and specifically covers health information entered into AI systems like ChatGPT or Claude.
AI gets personal: The timing matters because AI companies are pushing further into health. OpenAI introduced ChatGPT Health and ChatGPT for Healthcare earlier this year, Anthropic followed with Claude for Healthcare, and Elon Musk publicly encouraged people to upload medical records like MRI scans to Grok.
Enforcement muscle: The bill would require the Federal Trade Commission to create rules within 180 days of the bill passing; allow the FTC, state attorneys general, and affected individuals to sue to enforce it; and earmark $1 billion for FTC enforcement over the next 10 years.
Why it matters: AI health tools are moving faster than the privacy rules around them. The question isn’t whether people should be able to use AI for sensitive medical questions—the reality is, they already are. The harder question is whether the government can protect that data without freezing the kinds of AI health products people may actually want.
Robot Hand Startup Raises $11M After Settling Tesla Lawsuit
Proception, a startup building high-dexterity robotic hands, raised an $11M seed round led by First Round Capital, with participation from Y Combinator and BoxGroup.
Backstory first: Proception founder Jay Li was a technical lead on Tesla’s Optimus humanoid robot program. Tesla sued him last year, accusing him of taking trade secrets to start Proception. The two sides reached a settlement last month.
What he’s building: Proception is working on robotic hands that can move and sense more like human hands.
Its hand has 22 degrees of freedom and multiple joints per finger, and the company is now shipping its first batch to researchers and robotics companies.
The company is also building a sensor-packed glove that can capture human hand interaction data without needing a robot in the loop. That same glove can act as the robot hand’s “skin,” giving Proception a way to collect more scalable, task-specific data for dexterous manipulation.
The bigger picture: Humanoid robots get a lot of attention, but hands are one of the hardest parts to get right. If robots can’t reliably pick up, feel, grip, and manipulate objects, they’re limited in what they can actually do. Proception’s story is also a reminder that frontier tech has a very active talent revolving door—today’s former Tesla lead can become tomorrow’s supplier to the same robotics ecosystem.
Email Infrastructure for AI Agents
Via Sendmux
Sendmux is an email API built for AI agents. It lets developers create dedicated agent inboxes, send email through multiple providers with weighted routing and automatic failover, and receive inbound mail as compact structured JSON to reduce LLM token waste.
How you can use it
Create dedicated inboxes for AI agents
Send email through your own providers with routing and quotas
Receive inbound mail as structured JSON
Trigger live workflows with webhooks, SSE, and MCP
Pricing: Paid
Autonomous Software Delivery Agents
Via Revolte
Revolte is an AI-driven software delivery platform that uses autonomous agents to plan, generate, test, deploy, and operate code while keeping engineers in the loop for review and governance. It integrates with tools like Jira, Git, and Figma, and includes platform-as-code YAML, a CLI, and delivery intelligence dashboards.
How you can use it
Turn product intent into planned engineering work
Keep engineers involved for review and governance
Connect delivery workflows across Jira, Git, and Figma
Track reliability, observability, and release progress
Pricing: Paid and free plans
Jobs, announcements, and big ideas
Netflix uses an AI-generated Gene Wilder voice in its new Wonka golden-ticket reality show.
The Commerce Department lifts restrictions on Anthropic’s AI models.
Google unveils Gemini Omni Flash and Nano Banana 2 Lite, faster lightweight models for everyday tasks.
Anthropic debuts Claude Sonnet 5, a model tuned for coding and demanding professional work.
OpenAI introduces GeneBench-Pro, a new benchmark measuring AI performance across genomics tasks.
Google clears Gemini 3.5 Pro for a July launch.
Don’t fall for this trap! Every AI tool makes you think you should automate everything. But what people miss is the how. Watch along as I explain how you can get the most out of your tools:
That’s a wrap! See you next week for more. Enjoy the long weekend!
