Got fable to hand-write weights of a 100k param model to recognize real versus fake legal citations.
Basically one-shotted a full NN that’s better than training it
For the first time, firms can build compounding engines of institutional knowledge. Those that succeed will become massively more valuable than those that don't.
Advancements are constant, surprising, and uneven. There is no best model
@ashwingop @sequoia @kleinerperkins @generalcatalyst @EngramLab @harvey Putting memory, preferences etc outside the weights just makes way more sense. Apart from accessing frontier intelligence w every new model, it’s easy to understand, tweak, and audit. A firm fine tune is unauditable and near impossible to finely tweak behavior.
Show more
Performance and $$ depend hugely on harness design. In vertical AI, the gains from a great specialized harness >>> gains from more expensive model
inb4 kimi has quirks - @runta can you try with a diff model also
Show more
