Every so often you meet a founder so sharp that every conversation lights up new pathways in your thinking and quickly deepens your grasp of their business and technology. Working with people like that is a privilege.

We’re excited to sponsor @a16z speedrun’s AI Faire at @SFTechWeek on October 9. Runware founder Ioana will be presenting, alongside live demos from AI startups across the speedrun community. Request a spot!

🎓 We've added deep-dive guides for @PrunaAI's excellent, H3-based, P-Video-2 Pro! If you're just jumping into P-Video-2 Pro, start here: - Writing prompts for full shots and sequences: - Camera moves, physics and full-body motion: - Dialogue, lip sync and native audio: - Starting and ending on frames you choose: - Speed, quality, size and length:

It was only a few weeks ago that I was posting how AI isn't very good at 3D model generation and likely won't be until we have multi-modal models. Well, Astra changed my mind. The trick is to give it a set of reference images from each side of the object you are trying to create and give it an environment where it can run the comparisons. This wasn't a one-shot, but the first pass was 90% there and only required a little fine-tuning. It also generated 10K/1K/100 triangle LODs for me and rigged it 🤯 There's a few modeling issues and self-collisions with the animation, but honestly this is good enough for what I need.

Back to working on Three.js Fire Pro for a bit. Doing some explorations on the tooling right now. I added some debug views that let you view different properties of the fluid such as heat, vorticity, smoke, oxygen, etc. You can view these by slicing the domain and viewing in 2D or as a volumetric rendering. These are mainly to help me in debugging but I will leave in since they are helpful when authoring the scenes themselves.

If you want to see how crazy these new models are, you'll want to read this. I built a wooden crane for my kids, but was struggling with the actual mechanical design of the carriage and lift mechanisms. I spent a few weeks trying to figure out the design and bought a bunch of parts off Amazon I knew I would need (bearings, timing belt, etc.), but designing the whole system was a bit above my pay grade. I decided to hand all of the measurements to Astra, gave it a rough idea of what I wanted and the Amazon links to the parts I bought. It thought for a bit and then asked me a few questions on the design I had failed to provide. It figured out the part numbers, looked up the datasheets to grab the dimensions, and asked me to provide dimensions that it couldn't find. We iterated a few times, aligning the design with what I had pictured in my head. The initial design was more complex than I needed and too bulky. But I hadn't given it any constraints on the size of the assembly, so I wasn't surprised. It took each piece of feedback in stride and updated the design in a sensible way. In the end, it generated some PDFs documenting the entire assembly and generated STEP files for all of the 3D parts I needed to print. The level of spatial and engineering understanding needed to do this is probably beyond what the average graduate coming out of a 4 year engineering school is capable of. I think we are on the cusp of an engineering revolution in the same way that CAD tools replaced drafting tables. This is a general purpose model that isn't even specifically trained to do this task and I gave it some horrible, underspecified prompts. Imagine what could be achieved with a custom-tailored model?