Every Inc.

Four major AI launches in three days. If you didn’t have time to keep up, here’s the rundown: 🪔 Genie 3 – Google’s tool that turns text into interactive 3D worlds you can walk around in. 💻 Claude Opus 4.1 – Anthropic’s coding assistant that got quietly but meaningfully better. 🔒 gpt-oss – OpenAI’s first open-source models in five years, now runnable on your own hardware. 🧠 GPT-5 – OpenAI’s new flagship model, designed to handle everything from quick chats to deep reasoning. Each suggests different priorities: immersive experiences, precision coding, open experimentation, and one-model-for-everything. We’re entering a phase where no single AI model or company sets the tone. Instead, multiple races are unfolding at once—open vs. closed, specialized vs. general, headline-grabbing demos vs. everyday tools. We read all the benchmarks, watched the demos, and tested what we could—so you don’t have to. Full breakdown of the week from Katie Parrott in the comments.

OpenAI just dropped GPT-5—and it’s a big deal. Yes, it’s faster. Yes, it’s cheaper. But here’s what really matters: GPT-5 is the first time most people will feel what it’s like to talk to a reasoning model. We’ve spent the last few weeks testing it across writing, coding, research, and AI-powered workflows. Some moments felt like magic. Others felt like the future hasn’t quite arrived yet. But for the average knowledge worker? This changes the baseline. You don’t need to pick a model anymore in ChatGPT. You just ask—and GPT-5 figures out whether to respond instantly or “think” for a while before returning a deeper answer. It routes your query to the right tool under the hood. That alone makes the whole experience feel more fluid—and more powerful. Some quick takeaways from our testing: ✍️ It’s a strong first-draft collaborator—less formulaic, more expressive. 🧠 It’s great at technical research. Think systems architect, not search engine. 🤖 It’s still not built for autonomous agents. But for focused tasks, it’s sharp. 💸 It’s 12x cheaper than Claude Opus in the API. Expect a ripple effect. This won’t replace your job. But it will change how you start it—and what kind of support you now have by default. Get our full Vibe Check in the comments 👇

What happens when your AI agent tries to create another AI agent? We tested Claude Code’s new agent feature the night it dropped—trying to chain them, orchestrate workflows, and debug the weird failure modes. Some of it worked. A lot of it didn’t. 1. Agents don’t share memory 2. They can’t call slash commands 3. You’re still the glue between them But with the right setup, you can manage multiple agents in parallel—each with its own persona, each running independently. It’s a glimpse of what agentic AI might become—and what’s still missing to get there. Read the full Vibe Check, brought to you "live" from Every's Think Week ↓

When should you trust the algorithm—and when should you go with your gut? Mike Taylor has spent the past year building and testing AI research tools. In In the latest edition of Also True for Humans, he shares three patterns he’s seen over and over—ways AI gets it wrong, even when it sounds right: - It operates on outdated context - It mimics social approval over real behavior - It parrots the loudest advice, not the best If you rely on AI to think better, this is a must-read on how to think with it—critically, clearly, and with context that only you can provide. Full post in the comments 👀

🎙️ We’re hiring a freelance podcast producer! You’ll help shape AI & I—our weekly show on how the smartest people use AI. If you’re excited about AI, fluent in Descript, and love turning big ideas into great episodes, we want to talk to you! Details in the JD ↓

For 20 years, Google's toothbrush test filtered every major tech bet: Will people use this twice a day? No daily usage meant no billion-dollar business. But AI agents are rewriting this rule—and expanding what's possible for builders and entrepreneurs. Dan Shipper calls this the "magic minimum": Can your product periodically deliver enough unexpected value to be irreplaceable, even if users only engage once or twice a month? Here's what's changing: User memory isn't the constraint anymore. Agents remember themselves. They work proactively, popping up when they've created value—not waiting for you to remember they exist. The result: Thousands of specialized tools that were "too niche" suddenly become viable businesses when they can proactively deliver value. For builders: Stop forcing daily engagement. Focus on periodic magic. For everyone else: The tools that transform your work might not be habits. Instead, they might be partners working in the background.

Ever wonder why switching between work tools feels like starting from scratch every time? Why your project data in Notion knows nothing about your tasks in Asana? Why every new AI assistant needs you to explain your entire context again? Former Stripe and Google executive Alex Komoroske traces this frustration back to a single architectural decision made by Netscape engineers in the 1990s. It's not a bug. It's not laziness. It's the hidden physics of how software works. But for the first time in decades, we might have the tools to change these physics entirely. Imagine if your data could carry its own rules about how it can be used. If AI assistants could see across all your tools without compromising privacy. If software could truly adapt to your workflow instead of forcing you into theirs. Link to today's Thesis in the comments 👇

The most valuable skill in AI might be knowing when NOT to use it. This week's Context Window explores why the friction we're racing to eliminate might be exactly what we need to preserve. → Willem Van Lancker argues that "productive friction" is essential for developing expertise—even as AI makes everything instant → Former executive assistant Shreeda Segan evolved from ghostwriting tweets to building AI systems that write better than humans → Alex Komoroske is building AI designed to help you become who you want to be, not just check off today's tasks → Dan Shipper offers a new definition of AGI: when it makes economic sense to never turn your agent off If you're wondering how to maintain your edge as AI handles more of the "doing," this issue maps out 10 strategies for intentional friction—plus a glimpse at the new hybrid roles emerging between writing and engineering. Full issue in the comments.

You tap a button, and your photo looks professional. GPS guides every turn. AI writes your emails. But what if the friction we're eliminating is exactly what makes us irreplaceable? Willem Van Lancker helped design Google Maps for iPhone. Now, as a partner at Terrain, he's swimming against the current of frictionless AI. His contrarian take: The struggles that feel inefficient today build the judgment that sets you apart tomorrow. His friction playbook reads like productivity heresy: - No auto-scheduling tools—manually coordinating meetings forces intentionality - Sketches before screens—analog creation naturally restricts ideas to their essence - Cooldown periods before shipping—deliberate pauses reveal better solutions - MidJourney over ChatGPT for visuals—embracing chaos preserves creative control - Strategic uncertainty in projects—committing beyond comfort builds adaptability Yes, AI systems amplify our capabilities. But Van Lancker's approach suggests that preserving some friction isn't about being a Luddite—it's about maintaining the struggles that develop our unique perspective. In an age where AI can do the work, the real differentiator is knowing which work to keep doing yourself.

AI can (and should) strip away drudgery, but our edge comes from the work we still have to wrestle with. That “wrestling” is what investor-designer Willem Van Lancker calls productive friction—the deliberate, feedback-rich effort that sharpens judgment and taste. His framework for recognizing productive friction: - Immediate feedback: You understand when and why you've failed - Cumulative learning: Each attempt builds your reference library - Transferable principles: The specific teaches the general As AI handles more entry-level tasks, the path to mastery is changing. Credentials matter less. Proof of work—what you've actually built, shipped, failed at—matters more. This isn't anti-AI nostalgia. Willem uses AI constantly. But he's intentional about preserving friction where it counts: sketching by hand before designing, writing full drafts before refining, going to primary sources when learning. What productive friction are you deliberately preserving in your work?