Anonymous
6/14/2025, 5:31:26 AM
No.105588149
>>105588126
Yeah 2.5's context is crazy. I was writing it off and only using Claude but it was having trouble and I switched to Gemini and it ate 300k tokens but it was actually able to navigate the codebase and succeed, and this was on the cheap and fast model too
The issues with improvements are just data and training costs. There's no excuse to be using t5xxl still for how heavy and old it is but good luck being the mad scientist to replace it. In 3 years copilot might even be able to help you assemble those datasets
Yeah 2.5's context is crazy. I was writing it off and only using Claude but it was having trouble and I switched to Gemini and it ate 300k tokens but it was actually able to navigate the codebase and succeed, and this was on the cheap and fast model too
The issues with improvements are just data and training costs. There's no excuse to be using t5xxl still for how heavy and old it is but good luck being the mad scientist to replace it. In 3 years copilot might even be able to help you assemble those datasets