Old Gemini 1.5 punching way above its weight by borrowing the thinking process from deepseek-reasoner. 🧠

it's like RAG but for thinking... RAT

LMAO

Seamlessly switch between AI models like ChatGPT and Gemini. Why limit yourself? Explore the power of choice with #AI #Flexibility #AdvancedFeatures

I have a new idea - combining the thought process of "gemini-2.0-flash-thinking" with gemini exp 1206 could be a powerful combination. Plus, it has multimodal capabilities!

You can also extract purely the thought content from Gemini Thinking models!

You can get improvement on 3.5 by using 3.5 as a reasoner with some good prompting including an example or two. Any non-reasoning model will improve with any reasoning steps. Interestingly, I used 1.5 to generate probably the world's first AI reasoning traces. See pinned thread.

Is anyone doing benchmarking on this approach? Seems like a super hack.

is juice a an openrouter chat applet you made?

Unreal 👏👏

Do you have a GitHub to see this process of extracting the reasoning from it?

I may release this repo! Keep notifications on.