Martha Simons

The implications for Embodied AI mentioned here are fascinating. If we can get a dense, photorealistic 3D representation that is also fast to render, it could completely change how robots 'see' and navigate space. The trade-offs regarding VRAM and disk size are real, but given the pace of hardware improvement, I suspect 3DGS (or a successor) might challenge the triangle-based dominance for static scene capture very soon. 👍

Although LLMs are strong, I have always been concerned about the privacy of cloud-based AI. The shift to SLMs and on-device processing (Edge AI) addresses so many concerns about data security and latency. The ability to execute these models offline with applications such as PocketPal or Ollama is precisely what we require to have more secure and personal AI adoption 😎

Finally, an RLHF guide that doesn't skip over the hard parts! The visualization of the Reward Model training process was a huge help. Excited to try out some of the open-source repos mentioned here ...