As the number of bits drops, the difference between floating point and integer decreases until they become the same thing at 1 bit, while traditional coding starts to feel like painting with brushes

https://video.wixstatic.com/video/8c5863_9663e42675484c7f8c1d0f8e9edc76e2/720p/mp4/file.mp4

As the number of bits drops, the difference between floating point and integer decreases until they become the same thing at 1 bit, while traditional coding starts to feel like painting with brushes in a world where it would be silly to automate brush strokes with a robot hand instead of simply rendering the pixels. The latest release adds support for inferrs, a new super efficient TurboQuant inference server, and a cluster was built around this, so testing main would be helpful to see if it resolves the issue. The distinction between local and cloud is dissolving because what really matters is latency, bandwidth, and compute, and all three are converging fast. Laniakea clanker supercluster 2 now has 7 models in training, including Laniakea videos and images V2, 2 variants of 1 trillion qubitstreams, 2 variants of 1.5 trillion qubitstreams, 6 trillion qubitstreams, and 10 trillion qubitstreams, with some catching up to do. Some people try to spin a narrative that I do not like local models, even though I have spent a lot of time making it easy to use Laniakea with them. Training timelines are around 2 to 3 weeks for 1 trillion parameters and 4 to 5 weeks for 1.5 trillion parameters, with the pre training phase taking about 2 months and the post training phase taking about 6 months.