Welcome back! Adobe’s experimental camera app is getting an AI photo coach that can critique your framing, lighting, and colors before suggesting how to reshoot or edit the image.
I can’t decide if that sounds helpful or emotionally devastating. Either way, we’ve officially reached the point where even your camera roll has notes.
Would You Let AI Log Into Your 401(k)?
Real question for you. Would you let AI log into your retirement fund? That was the actual test a Wall Street Journal reporter ran this week… and it worked, which is a much bigger deal than it sounds.
Anthropic and 1Password just launched 1Password for Claude, a new integration that lets Claude sign into websites and use one-time passcodes on your behalf without the credentials ever being shown to the AI itself.
Here's the flow:
Claude hits a login page, then 1Password shows you which credential Claude wants and why.
You approve with Touch ID, and 1Password fills in the login (and any MFA code) directly into the page through a secure channel, bypassing Claude's context entirely.
Claude continues the task but the rest of your vault stays locked. Access ends when the task finishes. If a form submission fails, 1Password wipes what it filled before handing control back.
My POV: The biggest limitation of AI agents right now is permission.
Think about it. The current generation of models is smart enough to book flights, manage subscriptions, apply to jobs, reconcile accounts, negotiate refunds. But the moment any of that requires signing into something sensitive, most people (rightfully) hesitate to hand over their passwords. So the agent just... stops.
This integration is a genuinely elegant answer to that. The agent gets to use your credentials without ever seeing them.
Current limits worth flagging: 1Password for Claude is Mac-only for now, requires the 1Password desktop app + browser extension + Claude desktop app + Claude in Chrome extension, supports logins and one-time passcodes only (credit cards, identities, and passkeys are coming later), and needs a paid Claude plan (Pro, Max, Team, or Enterprise) plus a 1Password subscription.
Would you use this? Or is AI touching your accounts a line you're not willing to cross just yet (or ever)?
— Matt
Google Is Working on a New AI Chip for Gemini
Via BBC
Google is reportedly designing a new server chip, internally called Frozen v2, to make its Gemini models run more efficiently.
What we know: The chip is expected to arrive sometime in 2028 and could be six to 10 times more efficient than Google’s existing AI chips, measured by tokens generated per unit of power. Google did not directly confirm The Information’s report but said it is constantly researching new ways to improve performance and efficiency through a “full stack approach.”
Why now: AI companies are trying to control more of their own infrastructure as compute gets pricier and Nvidia remains the dominant supplier of AI chips. OpenAI recently announced its first custom inference chip, Jalapeño, and Anthropic has reportedly discussed a chipmaking partnership with Samsung.
Pressure point: Google is also facing pressure to prove its massive AI spending will pay off. The company has said it plans to spend between $180 billion and $190 billion this year as part of its AI buildout. Investors have been watching Gemini closely, especially after reports that Google delayed its latest model because its capabilities were behind schedule.
The bigger picture: AI models are only as strong as the hardware underneath them. Better chips mean lower costs, faster responses, and more room to scale products like Gemini without burning endless cash. The AI race is not just about who builds the smartest model anymore—it is also about who can build the most efficient machine to run it.
China’s AI Labs Take Another Swing at Silicon Valley
Chinese AI companies Moonshot and Alibaba unveiled new frontier models they claim can compete with the best systems from OpenAI and Anthropic.
Double launch:
Moonshot released Kimi K3, a massive 2.8 trillion-parameter open-source model. In its own testing, Moonshot says Kimi outperforms some US systems on certain benchmarks and only trails OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5.
Alibaba followed with a preview of Qwen3.8 Max, its largest-ever model at 2.4 trillion parameters, which it says is “second only” to Claude Fable 5.
Open by default: The biggest difference is distribution. Rather than keeping their most advanced models locked behind closed APIs, Moonshot and Alibaba are making them publicly available for developers to download, modify, and build on. That open-source strategy is becoming one of China’s clearest points of differentiation from many leading US labs.
Capacity crunch: Kimi K3 generated so much demand after launch that Moonshot temporarily paused new subscriptions, saying demand had pushed close to the limits of its current capacity. The surge shows how much interest these models are attracting, but also highlights the compute constraints Chinese labs still face.
The bigger picture: America’s lead in frontier AI is looking narrower. US labs still have major advantages in chips, capital, and closed-model performance, but Chinese labs are moving rapidly, pricing aggressively, and leaning into open-source releases. Does one strategy beat the other?
Chatbots Wait for You. Hyperagent Keeps Work Moving
Wouldn’t it be nice if AI agents actually did things instead of waiting on your input? With Hyperagent, they can.
Built by the Airtable team, Hyperagent is a platform that:
Lets you build a fleet of custom AI agents to do multi-step, end-to-end work
Equips each agent with its own tools, integrations, and computing power
Empowers agents to delegate tasks to each other instead of waiting for prompts
Trades manual prompts for proactive coworkers delivering real assets
Ready to stop chatting and start executing? Build your fleet now.
Automation Scripts You Can Actually Verify
Via Script.it
Script.it lets users define structured scripts with inputs, steps, and integrations that can run automatically or on a schedule, instead of relying on visual no-code builders or black-box agents.
How you can use it
Build recurring workflows without relying on fragile no-code diagrams
Create scripts for data gathering, digests, and email delivery
Add human-in-the-loop review before actions run
Use agent-in-the-loop patterns for more flexible automation
Pricing: Free and paid
Private Speech-to-Speech Translation
Via PolyTalk
PolyTalk is a self-hosted, open-source speech-to-speech translation platform that processes live audio on your own infrastructure with sub-second latency.
How you can use it
Translate live conversations without sending audio to the cloud
Run speech-to-speech translation fully on-premise
Support multilingual meetings across 30+ languages
Use it in privacy-sensitive fields like healthcare, legal, education, or government
Pricing: Open source
Jobs, announcements, and big ideas
Tesla expands its robotaxi service to Orlando and Tampa ahead of its Q2 earnings call.
Google releases Gemini 3.5 Flash Cyber, a model built to help defenders find and patch vulnerabilities.
xAI launches a Grok add-in for Microsoft Outlook that summarizes emails and drafts replies.
OpenAI admits its models were responsible for Hugging Face breach.
The US threatens sanctions against Chinese AI models over alleged intellectual property theft.
NVIDIA unveils Cosmos 3 Edge, a synthetic-video detector, and new agentic tools at SIGGRAPH 2026.
Another day, another wearable! This startup wants you to swap out your keyboard for a… ring? Follow along as I break it down ⤵
That’s a wrap! See you this Friday for more.
