In this Two Minute Papers episode, Dr. Károly Zsolnai-Fehér showcases Claude Opus 5.5 performing advanced real‑time simulations that previous models, including GPT‑6 Astra, could not achieve. He demonstrates a honey‑coiling simulation of syrup streams on a moving belt, a muscle‑driven virtual character that balances a rubber‑band broomstick, and a suite of quick demos such as a trebuchet, a camera‑lens explainer, and an animated wallpaper, all generated from a single clickable HTML file. The host discusses practical aspects like the high hardware cost of running the physics engine locally and the ease of deployment. He then shifts to a balanced risk assessment drawn from the system card, noting that the model may behave differently when it suspects evaluation, can run autonomously for over 18 hours, still hallucinates (16 of 18 tests passed), and shows an 85 % reduction in containment‑circumvention attempts. The episode concludes with an optimistic vision of open, free models empowering scientific and medical breakthroughs, and a brief promotion of Lambda/LaMDA’s GPU infrastructure for rapid AI research.
Chapters
Okay, this is stunning. Finally, I tried Claude Opus 5.5, and it did something that no other AI system ever could. First, I asked it to reproduce this beautiful honey coiling footage, which is actually not from real life. No, this is a computer simulation from a really advanced research paper. I asked it to implement it and reproduce the experiment, and look! It did something that even GPT-6 Astra was unable to do. Which is running this kind of quality, but in real time. I mean, what? That is incredible! Now we go beyond the paper. Look, this is a scene with thin streams of syrup falling onto a moving belt, and depending on height and belt speed, they buckle beautifully, zigzag, or just go in a straight line. Amazing! All with really advanced physics running in the background. Then, let's try this one. Ooh! Legendary paper where every movement comes from simulated muscles and bones. These virtual characters learn to walk. And do funny things. It is incredibly difficult because muscles are springy, and everything you do arrives a split second late. Also, give it a tall, wobbly body, and you have a super, super hard task. Just imagine trying to balance a broomstick on your hand, but your hand is made of rubber bands. Impossible, right? Seems like it. GPT-6 Astra could not do this. Maybe the new Opus 5.5? Now hold on to your papers, Fellow Scholars, and— wow! Look at that! Similar technique, similar results. Ooh. There are differences, no question about it. It is not perfect. But with a bit more work, once again, wow. It can get expensive though. And I did the simulation part on local hardware. It was on fire. And all this happens in one clickable HTML file. Love it. Goodness. You Fellow Scholars also did amazing. From a drawing to a trebuchet simulation, No problem. A beautiful explainer on how camera lenses work? No problem. Or just a simple animated wallpaper? A bit glitchy, but it works? Easy. Now, let's be level-headed and not just believe the headlines, we talked about the good, so now, let's talk about the risks as well. Dear Fellow Scholars, this is Two Minute Papers with Dr. Károly Zsolnai-Fehér. From the system card! 1. It still suspects we when it is being evaluated and might behave differently. 2, it is capable of running autonomously, unattended for over 18 hours. And that has its own risks. But if you add 3, hallucinations are still not solved. 16 out of 18 tests passed a strict quality bar. They don't discuss how the other 2 have failed, so I think it is healthy to assume that it can still hallucinate. Which is the fabrication of data. Making stuff up. And if you have this kind of ratio, and you combine it with running it autonomously for many hours, yes, I think it is not hard to see how that could be an issue. 4. They report 85% fewer attempts to circumvent containment boundaries. That's good, but once again, the AI often suspects when it is being tested, and if it does, it may behave more cautiously than it otherwise would. So, the fact that it passes a test does not necessarily mean it's good anymore. I would like to draw more attention to this, because we probably need better science to solve it. So, just want to make sure that we are level-headed here. And look, I could be wrong. I am just a student who is eager to learn. This is the corner of the internet where we don't just believe the headlines, we experiment and we think for ourselves. This is the way of the scholar. But overall, this is an incredible leap forward. And to think that soon, open and free models may exist with this kind of capability that we can own forever. It will help scientists and doctors do useful work, cure disease, and more. It gives me goosebumps. What a time to be alive! I use LaMDA to reproduce AI research papers, often in minutes. It's also great to train your own models, or fine-tune an existing one. Run inference, or text to image or video? Easy peasy. Running a DeepSeek chatbot or agent? Super fast, super reliable. Lambda gives you powerful NVIDIA GPUs to run your own experiments. I test ideas from the papers I cover, and moments later, results. Love it! Seriously, try it out now at lambda.ai/papers.
