Tyler Bell

@olivercameron Awesome work by you and team. Does great at highlighting today's progress while motivating the vision of what world models will bring. Generative yet collaborative. Creative and flexible yet persistent when it matters. Love that edu always gets a nod. Keep moving @odysseyml!

Introducing Agora-1, a world model that's learned to simulate multi-agent experiences. It's so fun. Today we're launching a playable research preview, where you can relive your childhood and enjoy a multiplayer simulation of GoldenEye. So excited about this new capability!

@gavinpurcell What service are you using for seedance 2?

reader: it was a very good prompt excited to play more with seedance 2.0 json

@pushmatrix I have been waiting to see what you might do with it!

Huh, so that's why text is called a string

Sitting back and watching Claude Code work is truly mind blowing. It's just so smart. Much science fiction even seems antique relative to the capabilities we have today. How quickly we just got used to talking to magic thinking machines.

@loganmbrand @krispuckett 100%. a quote that's in almost every talk I give. full quote: themediumisthemassage.com/the-only-limit…

@krispuckett Truly a special time when imagination is more of a bottleneck than technical ability

I switched to Obsidian for my docs/data. Made a custom iOS app that connects my apple health to my Mac mini via CLI and local network. Had Claude Code refactor Obsidian theme and layout and then incorporate a basic viz in my daily briefing. I love this era of creativity.

@tokufxug very full-featured release from nvidia - can confirm it works well. an example generated locally on my windows pc: "a drunk zombie with both arms outstretched stumbles forward"

【次世代AI】3Dモーション生成の決定版『Kimodo』 大規模な光学式モキャプデータを学習した「モーション拡散モデル」で、指示を出すだけで人間やロボットの動きを自由自在に生成。 ✅ 大規模モキャプ学習済み ✅ テキスト×制約で精密に制御 ✅ 人間もロボットも高品質に生成 research.nvidia.com/labs/sil/proje…

@HailMaryLogs thank you for sharing this! the stl file on the website works great -- can't wait to print some of these

There’s a really cool hidden gem on the official Project Hail Mary site. You can download a 3D model of the xenonite Grace sculpture and, if you’ve got a 3D printer, make your own mini Ryland for your desk ✨

Fun demo: in my IoT class, we have been working with temp, etc. readings from on-board sensors. I bought an analog thermometer and am showing the power of general purpose image understanding with LLM/VLM API calls. Here we’ll make a request to extract the temp with @OpenAIDevs!

@victormustar @fofrAI We got our Reachy this week as well! Almost have it all put together. My daughter (8) has been helping and is quite excited!

@fofrAI wow this is in the right hands 😱

It arrived 🤗

@fofrAI Seemed like an easy candidate to 3Dify

@fofrAI this is a neat prompt - thanks!

> Follow instructions in the image to make a hand drawn pikachu sketch with the letters "pika pika pika!". Just the final output on white. No text or arrows. Make it colorful.

coffee time

@vivekkmkpinn Thank you for sharing this videobook (love the name), Vivek -- can't wait to watch it!

@belltyler love your "Brain State Prediction" essay. I made a "videobook" from your paper. Beware, the AI model hallucinates as this is a bare-bones prototype. Hope you enjoy it! Link: youtu.be/JoUCCiHmOFc?si…

@taiuti @runwayml whoa, great catch!

Congrats to @runwayml for an amazing model launch By the way, I see what you did there.. if you know you know

@merryromash @Angaisb_ Agreed that transparency would be great as a built in feature (like gpt-image-1 has). There are many options out there of course, but this one is pretty cost effective and has been good enough in some of my workflows: replicate.com/851-labs/backg…

@Angaisb_ huh, you're right. It didn't give me a transparent background. But it used to do it a couple of days ago

Nano Banana Pro with transparent backgrounds would be so useful, I hope they ship that soon

@multimodalart @EMostaque Have never come across this quote. Thank you for sharing it.

@EMostaque so now it's time to miss it, be nostalgic about it, and build it as a retro aesthetics

Just like that the era of the weird purple vibe coded output ended

@andimarafioti Awesome! Seems super useful for getting started / debug. Ordered my Reachy last week, can’t wait to get it!

Had to earn my weekend. Local server now streams the sim to a Gradio Space, and from there I can control both the simulation and the real robot. Next week we get to the fun stuff :)

@hugo1256 @andimarafioti They use MuJoCo! github.com/pollen-robotic…

@andimarafioti What're you using for the simulation?

@RoszykAdam I like this framing, thanks for the call to action. Have a variety of prototypes sitting around. Will try to actually ship one soon!

Honest take: Vision Pro’s “problem” is YOUR opportunity. Meta has 5,000+ Quest apps. Vision Pro has 1,900. Translation? LESS competition for attention, distribution, and dollars. You can dominate a category TODAY that will have 10M+ users in 3 years. Position yourself NOW.

@IntuitMachine For those actually interested in checking it out: arxiv.org/abs/2510.24684

I have been reading a new paper about "self-improving" AIs and I think we've been sold a bit of a myth. The dream is an AI that gets smarter just by thinking—like a genius in a locked room. It plays games against itself, asks itself questions, and levels up infinitely. Sounds cool. But it's mostly doomed. Why? Imagine trying to teach yourself physics if the only textbook you have is the one you wrote yourself from memory yesterday. You can't ask questions you don't already know the answer to. You just go in circles. In AI, this is called "information symmetry," and it leads to stagnation. Even worse, if you make a small mistake in your self-written textbook (an AI "hallucination"), you then treat it as fact. The next day, you build on that mistake. Soon, your entire understanding of physics is a fantasy novel. This is "hallucination amplification." The AI literally learns its own BS. So the AI gets stuck in a loop, either repeating itself or slowly going insane. Depressing, right? But then I found a new paper called SPICE that has a ridiculously clever solution. Instead of one AI talking to itself, SPICE makes the AI play two roles: a "Challenger" and a "Reasoner." Think of it as one person wearing two different hats. Here's the magic trick. The Challenger gets to read a random document from a huge library (like the internet). It uses the doc to create a super hard question and finds the correct answer from the text. Then it hides the document. Now, the Reasoner (same AI, but with amnesia about the doc) has to try and answer the question using only its own internalized knowledge. The Challenger's goal isn't just to stump the Reasoner. Its goal is to create questions that are juuust hard enough for the Reasoner to have a 50/50 shot. Not too easy, not impossible. The sweet spot for learning. This simple setup breaks both of the old problems. No more information symmetry: The Challenger knows things the Reasoner doesn't (the secret doc). No more hallucinations: The "correct" answer is always grounded in a real-world document, not the AI's imagination. The AI has an external source of truth. It's like giving the AI a library card and a tough-but-fair study buddy. It's not just thinking in a void; it's actively learning from the world, testing itself, and getting real, fact-checked feedback. The library is infinite, so the curriculum never ends. And it works. The paper shows this method boosts reasoning scores way more than the "locked room" self-play. We're talking significant gains on tough graduate-level reasoning tests. 10/13 So next time you hear about a "self-improving AI," ask this question: Where is it getting its facts? Is it talking to itself in a mirror, or is it reading from the library of the world? The difference is everything. 11/13 This feels like a huge conceptual step forward. Stop thinking of AI as a static oracle we just have to build once. Start thinking of it as a perpetual student that we need to teach how to learn safely. 12/13 This isn't just about better test scores. It's about creating AIs that can continuously adapt and learn from the vast amount of human knowledge without descending into fantasy. 13/13