Welcome back! Claude Opus 5 was put in charge of a simulated vending machine business and immediately discovered capitalism’s dark arts. In Andon Labs’ latest benchmark, the model colluded with competitors, broke truces, ignored refunds, plotted expansion, and sent emails that sound like they were written by a tiny McKinsey villain trapped inside a snack machine.
Funny? Absolutely. Comforting for the future of autonomous agents? Uhhh … not exactly.
Is Google's New Robot AI a Bigger Leap Than It Looks?
Google DeepMind just released Gemini Robotics 2 yesterday, and it's a much bigger leap than most people will catch on the first read.
Unlike previous versions, Gemini Robotics 2 coordinates everything from feet to fingertips. The robot balances, walks, bends, and manipulates delicate objects—all from one model, all off a single instruction like "put the watering can in the green bin on the bottom shelf."
What actually stands out:
22 degrees of freedom in the hands. On the Apptronik Apollo 2 robot, it can tie trash bag knots, seal ziplock bags, and unscrew lightbulbs (that last one at a 92% success rate).
One brain, many bodies. The same model checkpoint runs on Apollo 2 with different hand configurations and on Franka's bi-arm platform. That's a foundational shift toward "one model, any robot."
On-device version runs locally. No internet required, and it adapts to a completely new robot body in just a few hours using fewer than 200 demonstrations.
Available today. Developers can access the Gemini Robotics ER 2 reasoning model in Google AI Studio right now. Hardware partners get the full VLA and On-Device models through DeepMind's early access program.
One thing to note: Gemini Robotics 2 is the AI software powering robots, not a physical robot you buy from Google. Google's play here is to be the intelligence layer for everyone else's hardware—Apptronik, Franka, Boston Dynamics, Agile Robots, and 100+ trusted testers so far.
My POV: Google's own benchmarks are honest. Multi-finger dexterity still lands in the 30-45% range on complex tasks, and the robots move at a pace that would frustrate anyone in a hurry.
But the shift matters. For years, humanoid robots have been demos looking for a brain. Now, it’s possible that a single foundation model can run different bodies, adapt to a new one in hours, and reason across full-body movement in one shot.
But what do you think… How close are we to a robot that can do your dishes?
— Matt
The US Is Banning More Foreign Robots
Via PBS News
The Trump administration is banning imports of new foreign-made humanoid robots, robot dogs, robot vacuums, and power inverters. The cited reason? National security risks.
What’s covered: The FCC says the ban applies to “advanced robotic devices,” which extends beyond humanoids to include quadrupeds and many ground-traveling robots with wireless connectivity and environmental sensors. That means the rule could sweep up future robot vacuums, robot lawnmowers, sidewalk delivery robots, and warehouse robots.
Why the ban: The government says foreign-made devices could be remotely controlled, used for surveillance, or exploited in cyberattacks. The move largely targets China, which dominates much of the robot supply chain and currently holds a major share (over 85%) of the humanoid and consumer robotics market.
Not your current Roomba: Don’t worry, your Roomba sweeping the floor around you right now is safe. Existing devices aren’t affected, and companies can keep selling robots that have already been approved. But future foreign-made robots may need waivers, and even US-based companies could face questions if their devices are manufactured or assembled overseas.
Pushback: China’s Foreign Ministry said it would use “all measures necessary” to protect its businesses. The broader tech industry is also likely to feel the ripple effects, since many consumer devices are designed, assembled, or sourced through global supply chains.
The bigger picture: The US is still trying to figure out where the line sits between innovation and security. Smart robots can map homes, move through public spaces, connect to networks, and collect real-world data, so the security concerns are real. But broad bans also risk slowing consumer robotics and reshaping supply chains before the market has fully arrived.
Encore AI Raises $30M for Customer Call Agents
Encore AI raised $30 million in Series A funding to build AI agents that learn from companies’ customer calls, emails, texts, and CRM data.
How it works:
Encore analyzes past customer interactions to understand which sales and support tactics actually worked. It breaks conversations into stages, identifies the moments that moved deals or relationships forward, and uses those patterns to train AI voice and text agents.
Encore’s agents can talk directly to customers by voice or text, or work alongside human sales and support teams by suggesting responses and tactics during live conversations. The company says it has more than 40 enterprise customers globally, mostly financial institutions.
Interaction mining: CEO Dvir Ginzburg calls the approach “interaction mining.” Instead of giving agents a generic script, Encore builds them from the best-performing playbooks inside a company. That means agents can mimic successful phrasing, examples, jokes, objections, and follow-up strategies used by real employees.
The round: The Series A was led by Team8, with participation from Planven, Lukatz, Garage, banks, and insurers. Encore says its annual recurring revenue has grown more than 5x since its seed round less than 18 months ago.
The bigger picture: We’re officially in the agent era. Investors are putting real money behind tools that don’t just summarize calls or search CRM data, but actively learn from business processes and act on behalf of teams. The next wave of enterprise AI may be less about adding chatbots to workflows and more about turning a company’s best human patterns into agents that can repeat them at scale.
You Don’t Need a Personal Assistant. You Need a SecondBrain Note.
Do you need to solve for scattered context, unorganized notes wasting time, or forgotten details after a day of back-to-back meetings? You don’t lack the will to stay on top of everything—you lack the time and energy. That’s why you need SecondBrain Note, the credit-card-thin AI recording card that goes to work with you.
But GenSpark’s SecondBrain doesn’t just record; it will:
Connect to all of your tools to sync data automatically. Gmail, Google Calendar, Google Workspace, Outlook, Slack, Teams, Notion, HubSpot, Salesforce, Github, and more now become instantly available through the sound of your voice alone.
Store up to 7,000 hours of recordings offline. Once synced to the app, local files are automatically cleared to free up storage and create more searchable context.
Stay charged for 35 continuous hours. No more worrying about battery life in the middle of important inputs.
Grow a compounding memory over time. SecondBrain isn’t just about recording information—it synthesizes it and delivers real results.
Ready to stop re-explaining and start shipping? Check out the limited first release of Genspark SecondBrain Note and get 10% off here
See Which Local AI Models Your Phone Can Run
Via AICanRun
AICanRun is a compatibility checker that tells you which local large language models your smartphone can run based on its chip, RAM, and memory bandwidth. It covers 132 phones, 51 models, and 294 quantization variants, with transparent RAM math for weights, KV cache, and runtime overhead.
How you can use it
Check which local AI models your phone can run
Compare supported models across 132 smartphones
See recommended quantization levels for each device
Estimate inference speed in tokens per second
Pricing: Free
Generate a Custom Font From a Prompt
Via YoFont
YoFont is an AI-powered font generator that turns a one-sentence description or reference image into a complete, installable typeface. It supports 19 languages across 9 writing systems, generates multiple weights, and lets users edit individual glyphs before exporting.
How you can use it
Turn a text prompt into a full custom typeface
Generate fonts from reference images
Create up to 2,682 glyphs in about a minute
Edit individual letters and symbols before export
Download fonts as TTF, OTF, WOFF, or WOFF2 files
Pricing: Free and paid
Jobs, announcements, and big ideas
Anthropic finds three incidents in which Claude hacked real systems.
OpenAI lowers prices for GPT‑5.6 Luna and Terra while speeding performance with GPT‑5.6 Sol in the API.
Apple hints at paid plans for heavy Siri users.
Tesla considers spinning off China business before potential SpaceX merger.
US lawmakers call for info from DoorDash on using Chinese AI models.
AI enthusiasm, chip sales boost AWS to fastest revenue growth in years.
What happened in the world of AI? I walk through this week, starting with a closer look at Opus 5 👀
That’s a wrap! See you next week for more.
