Rise of the AI Wrappers

Was this newsletter forwarded to you? Sign up to get it in your inbox.

We're at a rare moment: Some of the world’s richest companies are practically giving away access to their most powerful AI models in an effort to become kings of the next technological frontier.

If you're an expert in anything—law, medicine, energy, construction, agriculture, you name it—the opportunity has never been clearer: take your subject-matter expertise, combine it with the increasingly powerful capabilities of AI, and build a tool that helps you achieve your goals. And these tools, also known as AI wrappers, have become tidy businesses.

It happened in software engineering first, naturally, because so many AI experts are expert software engineers. Windsurf dramatically speeds up software development with AI, and now OpenAI wants to buy the company for about $3 billion (more on this below). Cursor, which does the same, reportedly turned down OpenAI last year (it’s now valued at $9 billion). The same trend is playing out in adjacent fields. Figma’s new features bring AI to user interaction and experience design by turning designs into working sites and creating content in your style. Descript, which makes video-editing software, is using AI agents to speed up content creation.

And we're still at the early stage of a much broader wave. Big winners will sprout anywhere people have domain expertise and problems worth solving.

There's a built-in conflict, though: Model providers profit when AI wrappers maximize token use (hello, reasoning models). The companies that run wrappers have the opposite incentive: Token use is a cost, so they look to save money by minimizing the amount of information (tokens) sent to and generated by models, and/or relying on cheaper models whenever they can. In the short term, both sides benefit—wrappers get affordable, powerful AI; model providers gain market share and scale. But this arrangement can’t last forever. One resolution is already occurring in the form of market consolidation, like OpenAI's reported acquisition of Windsurf, or wrapper companies like Perplexity choosing to build custom AI models on the cheap.

What's the smartest move right now? For most teams, building your own AI model isn't practical. Instead, the strategic play is speed: Rapidly build a product, capture your market, and—crucially—collect and control your data. Excellent niche-specific data collected today becomes the key ingredient for fine-tuning cost-effective models later, giving you a future-proof competitive edge. As I wrote last week, benchmarks become self-fulfilling prophecies. Data feeds good models, creating a virtuous cycle.

Looking ahead, there's an even bigger opportunity: what I’m calling "meta-apps." A meta-app is an AI-powered tool designed to work across multiple contexts or services. Granola, for instance, lets you record conversations whether you're on Zoom, Slack, or meeting in person, and ask questions of the transcripts—becoming a verifiable second brain. WisprFlow saves time everywhere you type with great dictation since its floating interface follows you everywhere you go on your Mac. Devin acts like an AI software engineer you can ping on Slack or tag directly on GitHub.

These meta-apps have emerged because AI has dramatically simplified connecting multiple APIs, and new capabilities like image understanding let AI navigate complex multimodal (text, images, and audio) tasks independently for the first time. They’re incredibly powerful because they become essential, embedding themselves deeply into their users' lives, sometimes creating more value than any single-purpose app. This is the elusive promise of Apple Intelligence, an AI that uses all of your data to always provide the perfect response.

Building a meta-app is hard. It requires deep integration, smart handling of data privacy, and solving tricky interoperability problems. The lessons from this first wave of AI wrappers, and the consumer markets they develop, will be critical for those building the meta-apps of tomorrow. I can’t predict the future, but I’m betting on teams with a clear vision who can get there, fast.

Release notes

Speed wins: Windsurf ships Wave 8 and gets a $3 billion call from OpenAI

In just 174 days, Windsurf went from launch to eight “Wave” drops—roughly one big release every three weeks. Wave 8 hit Tuesday with huge improvements to working in Windsurf with your team, the ability to create fully autonomous workflows, and the power to run multiple AI assistants (Cascade) at the same time. On the same day, Bloomberg broke news of OpenAI’s $3 billion offer for the still-scrappy dev-tool maker. Velocity is leverage.

A timeline of Windsurf’s development. Every illustration.

Figma fires back with Make, Buzz, Draw, and Sites

Figma is starting to write its AI story: Designers get an AI coding copilot (Make), a brand-safe AI-powered content engine (Buzz), quick-sketch vector tools (Draw), and push-button deploy to the web (Sites)—all inside the canvas.

Anthropic keeps the pedal down—and might be teeing up a new model

Within 10 days, the Claude team did the following: added real-time web search to the API, undercutting Google and potentially starting a search price war; folded Claude Code into its premium Max plan; dropped native integrations for Jira, Confluence, Slack, and more; and—per Bloomberg—started prototyping a “vibe-coding” tool with Apple.

Benchmarks of the week: Impressive sprints, stubborn marathons

1. Drug-discovery agents design end-to-end pipelines (from design to synthesis) that beat most human teams on novelty and viability.
2. GeoGuessr—AI knows where you live, or can figure it out: An o3-powered bot nailed pro-level location guesses, often within meters.
3. Factorio v0.2 is the newest video game at which AI is trying its hand. It’s a game in which you build and maintain factories, and requires coordination and optimization. A group of AI agents playing the game has shown some early success.
4. But lots of challenges remain for agents: New research shows that their capability drops exponentially as task length grows. Translation: Today’s agents shine in bursts but still stumble on full-day projects—a hurdle that will likely be hard to clear. At least for a while.

Knowledge base

“Vibe Check: Gemini 2.5 Pro and Gemini 2.5 Flash” by Katie Parrott/Context Window: More than 4 million developers are using Google's Gemini models while tools like Cursor and Windsurf are powered by Gemini 2.5 Pro. With its massive context window and significantly lower costs (eight times cheaper than -o3), Gemini 2.5 Pro handles complex tasks while Flash offers flexible speed-vs-thinking trade-offs that make it ideal for production workflows. Read this to understand why Google's strategic advantage might not be in flashy consumer AI interfaces but in becoming the essential infrastructure powering the next generation of AI applications.

"Why o3 Is the Best Model Yet for Real-world Learning" by Rhea Purohit/Learning Curve: AI is supposed to be a good teacher. Rhea set out to put that idea to the test by challenging OpenAI's models to help her achieve a seemingly impossible fitness goal. Read this if you want to understand why o3 might be the breakthrough learning tool we've been waiting for.

🎧 🖥 "What Jhana Meditation Can Teach Us About LLMs" by Rhea Purohit/AI & I: Nadia Asparouhova didn't set out for spiritual enlightenment, but her deep dive into Jhana meditation led her to a provocative conclusion: maybe we're more like LLMs than we think. This episode will make you question everything you thought you knew about what separates humans from machines. 🎧 🖥 Watch on X or YouTube, or listen on Spotify or Apple Podcasts.

"How One Writer Uses Epic Walks as a Creative ‘Operating System’" by Scott Nover/Superorganizers: Craig Mod, a writer and photographer, organizes his life and work around epic month-long walks across Japan that fuel his books, newsletters, and membership programs. These aren't your average "touch grass" breaks—they've spawned walking salons with Wired co-founder Kevin Kelly and even accidentally made a quiet Japanese city go viral. Read this to discover how he built a creative life based on perpetual motion.

🔏 "AI Focus Groups—And Soon AI Copywriters—Will Make Ads Superhuman" by Michael Taylor/Also True for Humans: Michael Taylor thinks that AI will transform advertising from basic roleplay to fully autonomous ad engines that test millions of variations on virtual audiences before a single human sees them. His own tool, Rally, already lets marketers test ads on AI personas with 85 percent accuracy. Read this if you want to know whether your marketing career has a future (spoiler: it might, but not in the way you think).

From Every Studio

Smarter summaries coming to a Brief near you

Over the past two weeks, we’ve rolled out major improvements to our email summarization system, making Cora, our inbox management tool, even better at surfacing what matters most.

What’s changed:

1. Essence over excess: Cora cuts to what matters to you. Whether it's a newsletter or a project update, you'll get the bottom line, like “TechAI raised $75 million” or “John asked for an intro to Jordan,” without the fluff.
2. Context-aware clarity: Summaries adapt to match the email. Technical newsletters keep their lingo, while project updates stay conversational and follow a crisp three-part structure: bottom line, context, and next steps.
3. Feel the tone: Emails aren’t just facts. Cora catches emotional cues, too, so you can follow up with joy, urgency, or empathy.

Sparkle’s shipping smarter, lighter, and experimenting with local models

Sparkle general manager Yash Poojary’s latest deep dive into open-source models shows just how far small, fast, and local models can go. After testing across text, image, and audio/video categories for file processing, he found that Meta’s LLaMA 3.2B hit a sweet spot: powerful enough to rival Claude Sonnet 3.7, and light enough to run locally.

Why experiment with open-source? You can download open-source models like LLaMA 3.2B and run them directly on your computer with no internet connection or data leaving your device. That means better privacy and faster performance for you. Imagine a cluttered Downloads folder full of meeting notes, receipts, and PDFs. While Sparkle currently only organizes by file name, we want to build a version where it understands a file using its type, content, and creation date. Using local open-source models, Sparkle could scan, summarize, and organize those files automatically, without sending anything to the cloud.

LLaMA3.2B currently only supports text. In the meantime, we’re watching this space.

Meanwhile, Sparkle’s gotten a few upgrades:

1. Faster performance
2. Custom file organization timing (hourly or daily)
3. Menu bar icon is now optional

These tweaks are small but mighty. They’re built to make Sparkle feel lighter, more personal, and more in sync with how you work. Try Sparkle today.—Vivian Meng

Alignment

Good mistakes. About two days ago, I tweeted something riddled with typos. Normally, I’d cringe myself into oblivion and obsess over imaginary judgments of my intellectual—and moral—failure. (Seriously, why do we always spot these mistakes three seconds after hitting “publish”?) But this time, I didn’t care. Actually, I liked my post better. The mistakes felt more human. Perhaps it’s because we’re increasingly surrounded by artificially perfect prose: clean, crisp, emotionally sterile words that practically scream AI. Maybe you’ve sensed this too. Your trust evaporates the instant you encounter suspiciously spotless grammar and overly precise clarity. Typos and awkward phrasing, by contrast, restore faith. They’re like faint but meaningful signals of life. Tiny reassurances that someone, somewhere is real. Of course, soon enough someone will train an AI to sprinkle imperfections into writing and simulate this authenticity. Maybe that moment will mark the beginning of something stranger, a world where everything becomes suspect. Until then, I'll embrace the occasional typo. A small rebellion, a good mistake—proof that at least for now, a human is still at the keyboard.—Ashwin Sharma

Collaborative filtering

Director’s cut. Frame Set is hiring a senior software engineer to help build the next generation of creative tools for filmmakers. The company, which helps directors find inspiration, is expanding its vision to support the entire creative development process using AI—treatments, storyboards, animatics, and more. This is a full-time, remote role with competitive pay, equity, and benefits, plus the chance to make a big impact on a small, fast-moving team. If you're excited about building tools that empower creative professionals, get in touch at [email protected].

Alex Duffy is the head of AI training at Every Consulting and a staff writer. You can follow him on X at @theheroshep and on LinkedIn, and Every on X at @every and on LinkedIn.

We build AI tools for readers like you. Automate repeat writing with Spiral. Organize files automatically with Sparkle. Write something great with Lex. Deliver yourself from email with Cora.

We also do AI training, adoption, and innovation for companies. Work with us to bring AI into your organization.

Get paid for sharing Every with your friends. Join our referral program.