I prompted this entire grid in the background while watching some YouTube videos. By 2026, building games will be where AI coding is today. Most people in the industry have no idea about the sweeping changes it will bring - but that’s what makes it exciting: bringing deep.

Not everything is auto/exactly differentiable in 3D AI though :D.

Back‑propagating all the way from the app to the model layer lets us ship things beyond what existing off‑the‑shelf tools offer.

This will hold even more strongly as AI minimizes the time between thought and execution. These functions have to be done over small time integral windows, otherwise it can be "overthinking".

Equation of a blissful & energetic life is elegant but hard to optimize in practice. Reward function flow is clear and measurable: # breadths/min -> # heartbeats/min -> # thoughts/min. Creative flow state reward func: Min_{other tasks} Max_{current task} thoughts/min.

Which AI frontier lab is closest in terms of getting a base multimodal bioLLM with below specs? It has been almost a year since my father suddenly passed away from Glioblastoma. In the background I have been mapping and calibrating what it would take to solve it. Stack: frozen.

A deep learning native game engine will look as follows:.- Real-time video diffusion model as the canvas.- A 3D memory system to store 3D/4D traces, which the video model will condition on to give arbitrarily complex storylines.

Current neural networks already know how to produce Pixar-level and beyond video outputs, if we solve just one problem from here: 4D meshes with easy control from prompts. This will enable the next generation of video models to deliver fully editable, production-quality animated.

How far can today's AI tools take us in terms of getting (explicit) 3D controllable AI videos? Here's a tool in our web dev app and resulting outputs. I have been waiting to create kids content for my daughter with AI :D

CSM Cube has unprecedented 3D AI workflows and this is just the tip of the ice berg. I can't wait to share more soon.

RT @CSM_ai: Cube lets you generate environmental assets with parts easily from a single image. Our web app is the fastest way to create usa….

Happy to report we never interviewed this guy . crazy to see almost everyone on my thread having some experience with him 😅.

Open problems to solve in scaling up large neural networks further: . - Meta-learning programs. Given a set of computer programs and tests, have N agents predict reusable code fragments to reconstruct the entire codebase, with the objective of minimizing Kolmogorov complexity (as.

RT @CSM_ai: [New 3D Superpowers for All]. All free users now get trial access to our 2M poly models before upgrading 🚀. Get a high-resoluti….

RT @CSM_ai: From the Image-to-3D menu, select the highest geometry settings without textures to generate 2M+ poly meshes 🔥 .

When LLMs run DevOps on Kubernetes, it’s mother-seed GPUs birthing spawn GPUs.

It is interesting to reflect on some intuitions from before gradient descent really started working in practice. If you thought about optimization, it was obvious that a neural network could learn the smooth landscape. But it seemed implausible for them to solve the discrete

Make your production code easily LLM readable to increase speed of shipping.