Muse Glimmer is out, and we're opening the weights! Genuinely proud of what we shipped here, and even prouder of the people I got to work with on this model. Lots still to figure out, which is the fun part 🎉
More here: https://t.co/FjiqN8QdrG
@christianelton_ 1) heterogenous compute complexity/incompatibility
2) latency/routing/coordination complexity costs
3) there is a real cost to the user (electricity, using up MTBF)
The benefit you get is not even close to worth it
Most prestigious institutions are talent magnets first. They dont _cause_ the success of their members, but attract members that already have max propensity for success in a self-reinforcing loop. True for countries, companies, schools ... something that should inform policy.
Really attractive for founders to think they are being copied. Rarely the case in practice, people usually come to similar solutions independently with the same market signal
@joebradford Headline is misleading since “free chatgpt” is meaningless. Free default switches the model on the fly. Results are better controlling for a single model (even an old one). Wouldnt expect other frontiers to be that different. Either way classical arabic isnt quite there yet.