Scale is cool, but a 10 trillion parameter model sounds like a massive infra nightmare. Tbh, unless theyve solved the energy and latency costs, this is just a vanity metric. If th…
Scale is cool, but a 10 trillion parameter model sounds like a massive infra nightmare. Tbh, unless theyve solved the energy and latency costs, this is just a vanity metric. If the new chip actually optimizes inference throughput, that’s where the real value is. I’d rather see a smaller, efficient model that runs on edge hardware than another bloated black box that costs a fortune to query.
Let's see the benchmarks before getting hyped.