Personal AI agents are great, but there is a BUT… I'm a big fan of personal AI agents. They can save us hours, take care of tedious tasks, and let us focus on what actually matters. But when it comes to young people, we need to be much more careful about how we use them. If you're an adult, you've already gone through years of developing your thinking, your reasoning, your ability to work through hard problems on your own. You've built the foundation. Delegating tasks to an agent at that point is a tool that multiplies what you can already do. But if you're still in school, that foundation is exactly what you're building. And it gets built through the work itself: struggling with a math problem, figuring out how to structure an essay, getting something wrong and understanding why. An agent that does your homework for you doesn't just skip the boring part. It skips the learning. The finished assignment was never the point. The thinking you did to get there was. That's why, for young people, the role of AI should be completely different. It shouldn't be an agent that does things instead of you. It should be a tutor that helps YOU learn. One that asks you questions, guides you when you're stuck, and makes sure you actually understand, not one that hands you the answer. This is exactly why we built @Astra__AI, and it's the core goal behind everything we do. Astra helps students learn through a clear, structured learning process. It's designed so that students do the thinking themselves, with AI supporting them along the way, so they come out the other side with real knowledge and the ability to reason on their own. I genuinely believe this is one of the most important things we need to get right as a society. If an entire generation grows up outsourcing their thinking before they ever learn how to think, we lose something that's very hard to get back. AI can be the best tutor any student has ever had. Or it can be the thing that quietly stops them from learning. The difference is in how we build it. Let's not mess this up.

I encourage all entrepreneurs by heart: GO BUILD. There are two great places to stand in this economy as a business: - You're a giant. You own the data centers, the GPUs, the energy contracts. Compute is your edge. - You're a small, fast team. A small team of people with AI can now ship what used to take a 100-person org. Design, code, marketing, support, research, all with leverage that pre-AI founders could only dream of. The danger zone is the middle: companies too big to move fast, too small to own the infrastructure. If you're small, this is your moment. Speed was always your edge.

For those worried about margins in the AI application layer, take a closer look at this chart and how AI inference costs are falling. Like any other big technology breakthrough, AI is a new PLATFORM. As the famous quote goes: "A platform is when the economic value of everybody that uses it exceeds the value of the company that creates it"

Voice is quickly becoming a growing way people interact with AI. The CEO of Instinct (@noahrshinn) recently shared that 90% of interactions with their personal AI agent happen by voice. At @Astra__AI, serving more than 15M users, we are also seeing the growing importance of voice interactions Voice does bring some unique standpoints that one should take into consideration: 1. Engagement is higher. Talking is faster and easier than typing. Lower friction means more interactions per day, and frequency is what builds retention. 2. Bigger addressable market. Voice opens AI to people who don't type comfortably or don't sit at a screen all day. That includes older users, people on the move, and mobile-first markets. 3. New hardware surfaces. Earbuds, smart glasses, cars, and wearables have little or no screen. Voice is the natural interface for all of them, and AI agents will make them more usable. 4. Getting voice right in AI is hard. Low latency, natural turn-taking, handling interruptions, and keeping costs per minute down. Teams that solve these will build a strong advantage in my view (at Astra AI we are already working hard on this). 5. Richer context. Spoken conversations carry more nuance than short typed prompts, which lets agents become more personal and more useful over time.

At @Astra__AI we now have more than 15M registered users. $0 in VC funding; bootstrapped from the ground up. Our growth curve keeps getting steeper and steeper, and we are now not only one of the fastest-growing startups in Europe but in the World. Thank you to our users and to our team, who continue to ship and focus relentlessly on bringing the best version of a personal tutor accessible to everyone.

One of the strangest things about working in AI right now is the gap between how fast the technology moves and how slowly opinions about it change. Every few weeks there's a new model that's meaningfully better at reasoning, writing, coding, or explaining. But when I talk to teachers, parents, and even business leaders, I often hear the same thing: "We tried it, it wasn't good enough." When I ask when, the answer is usually "a couple of months ago." First impressions stick, especially disappointing ones. But in AI, a month-old experience is basically a historical document. For those of us building AI products, this is part of the job. Shipping great technology and products isn't enough. We also have to help people update their mental picture of what's possible. Sometimes the most valuable thing you can do is get someone to try again.

At @Astra__AI we are now adding +150.000 new users each DAY! Our growth curve continues to look steep, and we've done it bootstrapped from the start, with zero VC investment so far. To put that in perspective, in the first 2.5 years, Instagram grew by +110k users per day.

Vercel's CEO: "When a team adopts AI, they always adopt multiple models. The average enterprise adopts over 5 AI models." This goes against the arguments that we'll end up with a monopoly or duopoly in the AI model layer.

In our conversation, Vercel CEO @rauchg explains his recent statement that open models now account for 80% of Vercel’s token traffic. He says models from OpenAI, Google, and Anthropic have steadily been losing ground to open models, with openness and speed emerging as key trends.

Investors undervalue vertically integrated AI companies. For months now, raw model performance is no longer the moat, and it is becoming increasingly less important. On top of that, switching between models and providers now takes hours, while SaaS in the past might have taken months.

As a UX designer by heart, I love that even in AI we are coming back to the basics of what makes a product or service valuable. Not every use case needs math-olympiad-level intelligence. Most don't. That's the moment the application layer gets interesting, and we are in this moment right now. $META's Muse is an example of that. It is not winning on raw model performance, it's winning on product: distribution and UX are doing the work.

People are overestimating the AI model layer and underestimating the AI application layer. Many in the industry have recently asked me what we will do if Frontier Labs lowers the pace of new model releases because of the recent pacing calls from Anthropic and OpenAI. At a higher level, people still vastly underestimate the value creation happening in the AI application layer and overestimate the value of the AI model layer. The AI model has become much less important to successful task completion and client satisfaction than it was just last year. Most gains and product improvements are increasingly coming from our harnesses, workflows & integrations, and the usage data patterns we have. To return to the question: if frontier model release cadence slows substantially, we will continue to grow at the same pace or faster than before.

CALLOUT to the AI industry: Public evals are becoming irrelevant. Whether it's intentional or not, models are increasingly benchmaxxed. The AI Application layer companies need their own internal evals, built on real outcomes and real task completion. At Astra AI we process over a trillion tokens a month. Every new model that ships gets run through our internal benchmarks before anything else. That is the real eval. The delta between public scores and internal results is bigger than most people would guess. If you're building on top of these models and you don't have your own evals yet, that should be priority one.

At @Astra__AI we now have 77k people signing up every day (up 712% YoY). Our fully bootstrapped 3-year-old startup now gets daily sign-ups that are larger than the number of installs WhatsApp gets in the US per day, and almost as many as Instagram does.

Fun week to be the CEO of Astra AI. Back to school, stronger than ever. While we had a strong year last year, the stakes are even higher this year: daily user registrations in the first 10 days of September are up +4,680% YoY (46x), with over 12M users already using Astra. We did this without a single VC investment. The team spent the summer laser-focused on improving the product and being ready for exactly this moment. Thank you. Now we have to earn the trust of every one of those sign-ups. Also, to our partners, yes, we (the world) will need more compute.

Our @Astra__AI is now the 3rd and 5th most-downloaded edu app in Argentina 📷 (Android & iOS). We are coming for that num 1 spot, Owl. (@duolingo)

AI progress is truly remarkable, but at the same time I believe AI will not drive human ingenuity and value down but rather up, because the idea that AI keeps progressing while we humans stay still is wrong. Humans have always progressed, learned, and adapted to our environment and the tools and technology that we have at our disposal, becoming more creative and discovering more things; this time will be no different.

@Astra__AI was selected as App of the Day on the Apple @AppStore 🎉 Behind this are 3 years of improvements, tough decisions, and people who care about what they’re building.