Some of the AI adoption pitches landing in my inbox could really use a little AI adoption themselves. 😁 Lately, it feels like 90% of the DMs and connection requests I get are offering an AI workshop or help with “agentic transformation” etc. Somehow, all that intelligence seems to disappear when it comes to checking who they’re contacting. I wrote my first neural network more than 20 years ago while still studying. At Reply, we’re working with most clients on sophisticated agentic AI projects that deliver substantial cost savings at scale. Between my posts, public talks and podcasts, there’s plenty of context available to work out what might actually be relevant. A quick look would make for a very different pitch. Given the sheer volume, I’d assume much more of this outreach is automated now. Folks, when you’re trying to sell AI expertise, researching the person and company you’re contacting seems like a reasonable first use case for the very agentic AI you’re pitching. You have the tools to make that research much more efficient. Skipping that step makes it quite hard to trust the offering, and often the first sentence is enough to give it away. Is it just me, or are you also seeing more AI slop outreach in your inbox ?

Parents with several kids might have an unfair advantage when it comes to working with AI agents. With five kids at home, we’ve had plenty of experience keeping several things moving at once. Add years of juggling meetings, a half-finished slide deck and some code waiting for attention, and apparently I’ve been training for this for quite a while. With AI agents, those parallel tasks can actually keep progressing while we focus on something else. One agent researches a topic, another writes code and a third prepares a draft. Getting that to work takes good multi-tasking, a sense of which tasks can run independently and enough context for each agent to produce something useful. Good delegation becomes a pretty practical AI skill, including recognizing when someone needs a clearer explanation of what you meant. Of course, five agents finishing at once also means five results waiting for your attention. That’s where the bottleneck can move, because checking the work still takes time and an understanding of the subject. And an AI agent saying “done” may deserve the same follow-up as a child announcing that they’ve put their shoes on. 😅

Exciting! It feels like Mixed Reality in a small form factor is finally happening with Meta’s new VR Glasses. 📷 I got particularly excited about the “hologram” calls: photorealistic avatars streamed as real-time stereoscopic video. 10 years ago we started working on HoloBeam at IdentityMine, also with the good ol’ Kinect. 📷 We captured people as live volumetric video and streamed them into HoloLens over a normal internet connection. Meta is taking a different approach here, animating a digital likeness instead of streaming the live depth-camera capture we used for HoloBeam. Also impressive specs: 📷 Around 100g on your face, with compute + battery in a separate tethered puck 📷 Dual micro-OLED displays, 2,412 × 2,288 pixels per eye, 37 PPD and up to 120 Hz 📷 Snapdragon Reality Elite + 12 GB RAM 📷 Eye tracking, hand tracking and full-color passthrough apparently without distortion 📷 Multiple spatially anchored virtual screens for getting some work done and can be used as simple usb-c display 📷 Up to 3 hours of video playback, with 45 W charging The 70° horizontal FOV and the cable to the puck are still compromises. And $1299 is steep with shipping in spring 2027. Can’t wait. Are you excited as well?

How many times have you checked your GPT-6 weekly usage limits this week? 😅 I've checked mine many times, and I'm approaching 10% remaining with a reset coming tomorrow. Should be enough to get me through the rest of the day. Phew! I guess token maxing has evolved into usage maxing: trying to make the weekly limit and the reset timer reach zero together. 🤣 Have a great weekend ahead! 🙌

Our oldest daughter with me and standing room only during my talk today. Both are great, but having her there was the best part of it. Had a great time talking about scalable Agentic AI we develop at Reply at the Brains on Silicon conference. The talk was about what it takes to move agents from a demo to something that holds up in production, and where Physical AI plays a role once the agents leave the screen. Our oldest daughter is on semester break and studies something related, so it worked out nicely that she could come along. Also great to meet many people again I haven't seen in a while and to make new connections. Thanks for having us.

Agentic AI. Ja, ich weiß. Darüber spricht gerade wirklich jeder. 🙄 Am Montag halte ich einen Vortrag bei der Brains on Silicon und werde zeigen, was davon inzwischen tatsächlich im Unternehmen funktioniert. Mein Fokus liegt auf skalierbaren Enterprise-Lösungen und Cases, die bereits in Produktion laufen. Natürlich schauen wir auch ein Stück weiter nach vorn: Was passiert gerade an der AI Frontier? Welche Trends werden relevant? Und was tut sich bei Themen wie Self-Improvement, etc.? Und bei einer Konferenz namens Brains on Silicon müssen die Brains auch den Körper bekommen: Deshalb geht es auch um Physical AI und humanoide Robotik, ein Thema, an dem wir bei uns im Lab inzwischen ziemlich intensiv arbeiten. 🕺 Ich freue mich auf die Session und vor allem auf den Austausch danach. Wer von euch ist am Montag auch bei Brains on Silicon? 📍 Montag, 10:30 Uhr, Stage 2 „Agentic Workflows“

Yesterday I gave myself an early birthday present and upgraded my ChatGPT subscription from Plus to Pro 20x. That gives me 20x the usage of Plus, which sounded like quite a lot. Well, less than a day later I have 40% of my weekly limit left. 😅 I’ve been using GPT-6 Astra pretty heavily for some more complex work, and the results are very impressive. It’s really exciting to work with. It is also quite the token burner, apparently. What I find particularly interesting is the trajectory, though. We’ve seen this with o3, for example, where API pricing dropped by 80% within weeks. Inference gets optimized, smaller models catch up quickly, and costs come down. Looking at the pace of open models as well, I wouldn’t be surprised if capabilities that are very expensive today become available there within a couple of months at a fraction of the cost. So, exciting times for sure. And I guess my Pro 20x usage meter will provide some real-time evidence of just how much fun I’m having with Astra. 😉

Surely it’s not AGI, but after using GPT-6 Astra Pro/Ultra for a few more complex things, I feel it’s the first LLM that actually gets and executes one-shot prompts correctly pretty much every time. Of course it’s also thanks to the Desktop harness and testing capabilities but still very impressive and for sure deserves the hype around it. But AGI, c’mon. 🙄

🖖 60 years ago #StarTrek debuted on TV! I am celebrating #StarTrekDay with a 3D model of the USS Enterprise popcorn bowl. How did I make it? I found an image of this on the Paramount website (see screenshot). I gave this image to Astra and asked it to use Blender with its Computer Use Agent to turn it into a 3D model. The result is not 100% perfect with the details but really good and mind blowing considering I simply gave it a website screenshot, a simple one-shot prompt and spend ~15 EUR in token cost. BTW, I love how it's using Blender's Python interface to basically write code on-the-fly to create the 3D model and render it. Have a great #StarTrek60th today! 🌌

Wir streiten über KI-Verbote an Schulen, während ein Foto vom Arbeitsblatt reicht, um ein Physiklabor zu bauen. Interaktive Lerninhalte zu bauen, war bisher vor allem eine Frage des Aufwands, und genau das ändert sich gerade. Moderne Sprachmodelle wie GPT-6 Astra machen aus einem schlecht fotografierten Physik-Arbeitsblatt ein interaktives 3D-Lernlabor. Ausprobiert habe ich das mit den Physikaufgaben eines meiner Kinder. Ich habe einfach das Foto hochgeladen und darum gebeten, daraus eine Website zu bauen, auf der wir Parameter verändern, Experimente durchführen und die Physik direkt in 3D beobachten können. Etwa 45 Minuten später hatte ich sieben interaktive Experimente mit 3D-Visualisierung, Live-Diagrammen und Erklärungen auf Deutsch und Englisch. Die geschätzten API-Tokenkosten dafür: rund 10 Euro. Man kann zum Beispiel die Masse und Beschleunigung eines Autos verändern, einen Fußballschuss untersuchen, einen Kolbenmechanismus in Bewegung betrachten oder sehen, wie sich das Abfedern mit den Knien auf die Kräfte bei einer Landung auswirkt. In der Bildschirmaufnahme seht ihr das ursprüngliche Arbeitsblatt und das Ergebnis. Ziemlich beeindruckend, wenn man bedenkt, dass alles mit einem Foto angefangen hat. Das weckt bei mir auch Erinnerungen. Vor mehr als 20 Jahren haben wir im Studium kleine interaktive Physikexperimente gebaut, damals noch mit Macromedia Flash, und einige von euch werden sich daran erinnern. Solche Lernmodule zu entwickeln, war eine Menge Arbeit. Ich setze mich seit Langem für Virtual Reality (VR) und 3D-Visualisierung in der Bildung ein, weil viele abstrakte Zusammenhänge zugänglicher werden, wenn man sie selbst erkunden kann. Vektoren und Matrixtransformationen, chemische Moleküle oder physikalische Experimente sind gute Beispiele dafür. Die große Hürde war dabei immer der Aufwand für die Inhalte, und heute reicht dafür das Foto eines Arbeitsblatts als Startpunkt. Stellt euch vor, Lehrkräfte könnten ihre Aufgaben mit wenig Aufwand auf diese Weise ergänzen, und Schülerinnen und Schüler könnten selbst Werte verändern, Vermutungen überprüfen und Zusammenhänge entdecken. Das Ergebnis ist dabei ein Startpunkt und kein fertiges Unterrichtsmaterial, das gehört zur ehrlichen Betrachtung dazu. Dennoch gehören genau diese positiven Einsatzmöglichkeiten für mich unbedingt in die Diskussion über KI an Schulen. Wenn Lehrkräfte die passenden Werkzeuge und den nötigen Freiraum bekommen, wird aus einem Foto in Zukunft öfter ein Lernlabor.

Education should never look the same again. Now with the latest generation like GPT-6 Astra, even a badly taken photo of physics homework become a perfect interactive 3D learning lab in one-shot. I tried it with some exercises from one of my kids. I simply uploaded the photo and asked it to build a website where we could change parameters, run experiments and see the physics in action. About 45 minutes later, and for roughly €10 in estimated API token costs, I had seven interactive experiments with 3D visualization, live graphs and explanations. You can, for example, change a car’s mass and acceleration, explore what happens during a football kick, look inside a moving piston mechanism or see how bending your knees changes the forces during landing. Check out the demo video for a screen recording showing it all in-action and where I started from. Quite impressive for something that started with a photo. This also brings back memories. More than 20 years ago, we built little interactive physics experiments as a student project for upcoming teachers, using Macromedia Flash (some of you will remember that). Creating those learning modules took a lot of work but the opportunities are huge and I’ve long advocated to use VR and 3D visualization in education, because so many abstract concepts become more approachable when you can actually explore them. Think about vectors and matrix transformations, chemical molecules or physics experiments. A major blocker was always the effort required to create the content. Now we can start with a photo of a worksheet and ask AI to build an interactive learning experience around it. Can you imagine teachers turning their exercises into experiments that students can explore themselves, changing values and testing their own predictions? The conversation about AI in schools needs to include these positive uses. Teachers and parents can help kids use AI as a learning tutor, ask better questions and investigate why something happens. Of course, the explanations and simulations still need checking, and the kids still need to understand the subject when they sit down for an exam. And these novel AI tools can be a tutor and help anyone to prepare by making the learning itself more visual, interactive and engaging. There’s so much potential here for education, and I’d love to see more schools, teachers and parents exploring it.

Voice + LLMs are the future of the user interface. I’ve been thinking about this for quite a while, and the more and more I am convinced that this is where we’re heading. The idea is simple: we speak naturally, Speech-to-Text captures it, an LLM understands what we actually want and translates this so the system acts on that intent. This is much closer to how humans interact with each other. We tell someone what we want, add context when needed, and expect them to understand our intent rather than giving them a precise sequence of commands. And we see more and more evidence of this already. The GPT-6 Astra demo video clearly shows this type of multimodal, conversational interaction. Omarchy, which I posted about a few days ago, treats dictation like a first-class citizen. And on my main machine I use Handy all the time, the open-source Speech-to-Text tool that turns any text field into a voice interface. This has also changed how I interact with LLMs since last year. I often dictate long prompts now, simply speaking out all the context and thoughts that come to mind. The transcription doesn’t even need to be perfect, because language models are very good at understanding incomplete sentences, missed words, accents, and all the other imperfections of natural speech. Also, the visual UI won’t disappear either. I think it will become much more dynamic, with just-in-time UI adapting to whatever we’re trying to accomplish and showing info, data, controls, when we need them. And although I believe voice will become the most important input mechanism, keyboards, mice, touch, and other controls will obviously remain. Sometimes they simply make more sense. But a large part of what we currently type into computers can move toward voice, with intelligence sitting behind the input instead of treating it as just another text field. Voice + LLM + JIT UI is becoming the user interface layer across operating systems, applications, cars, robots, Spatial Computing, and many other devices. I’m pretty convinced this is where we’re heading. You?