@elonmusk and @xai definitely solved the compression problem having more density per megawatt of compute through energy efficiencies and innovating engineering built off of existing XAI engineering but also likely years of compounding bottlenecks from past problems solved at various companies.
@ItakGol@DownhourJeff Google has not sold its SpaceX shares. Since its initial $900 million investment in 2015 for a roughly 7.5% stake, do you think they will sell their shares though?
@elonmusk Scaling laws are imminent, would like to see a 10x in user base growth north of 100m+ users over the next few years for SpaceX to really have a grounding footprint in global ISP user base.
@scaling01 The smarter play is to skip one giant monolithic model. Instead, orchestrate a converged cluster of specialized 1-5T experts trained in parallel, then fused for coherency via routing & alignment but each 1-5T model expert has to have a high intelligence density to begin with.
@ArtificialAnlys If NVIDIA focuses in and continues this, the strategic play would be open source IoT physical AI vision in a new world of AI intelligent devices where every individual can tinker with an out of the box configurable AI.
I have no doubt @SpaceX will make radical improvements in efficiency in the AI infrastructure domain, Elon also stated long ago that C language was the most fundamental and besides i'm always reminded of the Raptor V3 evolution as the same company will apply first principles methodology for AI training and inference
@nextbigfuture@elonmusk If this is true, the implications are enormous and escape velocity is near and models will catch up to Frontier Labs all while having different variants.
@AwokeKnowing@alexwg Exactly 🎯 eliminates the need for cooling & ground based resources that are usually needed making it much more efficient. Why? Because you’re exchanging heat for light.