All we need now is affordable hardware with the memory to store that much context.
I heard some people use adaptive kv streaming to fit large contexts. Unfortunately, it’s not merged into the upstream yet.
That sounds very interesting. If it performs well it could solve many problems.
Improving llm under the hood can end up producing it on the same hardware too.
So they didn’t even start training it?
They most likely used a local fork for a while.
If that were the case wouldn’t they just merge the working local fork rather than merging and reviewing changes one by one?
I assume it’s to simplify the review process. Or to generate more hype. Or both.
They announced that it would be released next week, so I doubt they’ve only just started training it.
Fantastic news. My only gripe with 3.8 is that 262k is very limiting with all the reasoning the model does.
Pi ends up compacting history 3-4 times for larger tasks before it finishes.
