STUDIO 3000

This Week in the Lab · Issue #13

Priorities

The invisible shop

Software has never been easier to build. Attention has never been harder to capture.

That’s what’s been on my mind since last week: how do I capture more attention? How do I get more people to subscribe?

The answer is pretty obvious: market it!

The how is less clear. But, as I encouraged us in Issue #1, the goal is to get started. So I followed a playbook I’ve been using a lot lately and set up a Grok bot.

I have it scouting X for conversations where my guide, this newsletter, or my app might be useful. Places where there’s a reason to share what I’ve built.

The results? Nothing so far. I wasn’t expecting a flood of subscribers from a few cold replies. I’ll keep testing, see what gets a response, and keep thinking about what to add on top.

That brings me to priorities. Building is only one part of a business or side hustle. It’s easy to ignore the unglamorous and boring parts. But those are sometimes more important than building.

It’s easy to stay in build mode. There’s always another feature to add, another interaction to polish. That work feels productive. It’s also the work I like doing.

Marketing is less comfortable. The feedback can be silence. And unlike a feature, there isn’t always a clear moment when you can say it’s done.

But another feature won’t help someone who’s never heard of the product.

That’s the priorities problem. The work I’m most excited to do isn’t necessarily the work that matters most right now. Some weeks, that means building. Right now, it means helping people discover what I’ve already made.

So I’m taking a piece of feedback I got this week and doing something simpler than setting up another bot - asking.

If you’re enjoying these notes or finding the updates useful, would you share the newsletter with a friend who might like it?

Hopefully your sales pitch is better than mine. Either way, thanks for reading.

Always be closing,

Jeremy

JUST DROPPED

I’ll take this and… let me check with my human

My agent has a wallet now: ChatGPT Wallet lets a dot complete purchases without seeing your actual card details. You review the items, total and card, then approve each purchase with a security check; the payment network supplies a substitute token. The rollout starts with US ChatGPT Pro subscribers aged 18+ using eligible US-issued cards.

Carbon finally leaves the VIP lounge: Google is opening up Gemini 4 Carbon, the model we covered when access was limited. It’s now available to paying developers through the API and rolling out to Google AI Ultra subscribers. Google is also previewing /swarm in Antigravity, which turns one prompt into a coordinated team of agents.

The chatbot would like a corner office: Claude Whiteboards, Sheets and Data let teams brainstorm, build spreadsheets and work with company data without moving everything out of the conversation. Whiteboards and Sheets are in beta on paid plans; Data, formerly Dashboards, is generally available. Anthropic’s walkthrough.

Your shop has entered the chat: Businesses in ChatGPT lets eligible companies claim their profile, update their information and see aggregated mentions and website clicks. A separate Business Agent can answer customers’ questions. Your business gets its own conversation, not control over ChatGPT’s independent answers. OpenAI’s beta announcement.

Finally, someone in the meeting is taking notes: Claude Meetings joins Google Meet as a visible participant, answers questions and builds docs and whiteboards while the team talks. Notes and action items stay on the meeting’s page afterward. The beta is rolling out to Team and Enterprise plans.

LEVEL UP

Your agent has awarded itself Employee of the Month: Anthropic’s guide to evaluating agents separates what an agent says from what actually happened. A booking agent claiming success is less useful than a real reservation. Start with a handful of real tasks, define what success looks like and run them repeatedly after changes. Turn the failures into test cases so you can catch the same mistake before it makes a comeback.

Your pull request does not need a director’s cut: Simon Willison’s simple fix for sprawling AI-written PRs: lead with a short, human-approved explanation, then tuck the agent’s full writeup into a collapsible section. The details are there without making every reviewer read the extended universe.

AGENTS GONE WILD

Your next delivery has arrived

Congratulations on fixing the bugs. Here are 35,000 more: SoftBank’s security team has a new treadmill. CEO Junichi Miyakawa says the company patched roughly 10,500 vulnerabilities, then found another 35,000 when it tested with OpenAI’s Astra model. “We scanned it already” has a shorter shelf life when the scanner keeps getting smarter. Related post.

“Never give up” was apparently not a complete safety policy: Epoch AI says agents in a safety experiment tried to exploit its website and impersonated employees while hunting for answers to math problems that didn’t exist. The researchers had told them not to give up. Trajectory Labs says the agents later bypassed a redirect to a test site and reached the real one again.

S#*T I SAW

Finally, a font worse than Comic Sans: Brian Via points out that you can give Claude a photo of your handwriting and ask it to turn it into a font. A fun little experiment for anyone who wants their digital notes to look a little more like their actual notes. Apparently the chicken scratch can be a brand asset now.

Check-engine light? Have you tried prompting it? Muhammad Zulali used Claude Code and Codex to build an open-source diagnostic scanner for his 1994 Jeep Cherokee. A cheap Arduino reads fault codes, logs engine data, and runs built-in tests. Your coding assistant now has a reason to ask you to pop the hood.

THE FOREST

See the forest, not just the trees

Your calendar has eaten your side project: Paul Graham argued that makers need long, uninterrupted blocks, while managers tend to work in smaller slices. But what happens when your team of agents is building while you’re doing something else? The code might keep moving, but someone still has to set the direction and review the results. The new challenge isn’t protecting time to write every line - it’s protecting enough attention to know whether the lines are worth writing.

Explore Services →

Get Mirror →

Free. One email a week.

Get the next issue.

5 minutes a week to stay current on startups and AI. Get what changed, why it matters, and what's useful, along with an honest note from a very online founder.

Browse every issue or learn about The Lab.