[washadjeffmad]

That's kind of a stretch when Bard has been barely relevant since its launch.

It's somewhere between a 7B and 13B for most tasks, which can be run on a cellphone now, and we were already able to run better models locally on commodity hardware. Model complexity and capability are improving so quickly that there's no feasible way for Google to respond.

Sure, Google don't want to look like they're late to a party and under-dressed, but Meta just got a decade of advancement for free in three months with the LLaMa leak. They're looking for anything they can say to sound like they're making progress.

Benchmark this stuff yourself. Use logic problems, have the local model talk through them, and if it doesn't understand what to do, you'll be shocked how a single hint or seed change can lead it to a solution. You can even ask them to generate contexts for themselves that would help them solve those types of problem in the future, which work incredibly well.

Combined with arbitrary length context and long term memory projects, finetuning almost isn't necessary with sufficiently large models. It's really exciting.