zlacker

Claude responds differently to "think", "think hard", and "think very hard". Just because it's hidden to you, doesn't mean a user doesn't have a choice.

Saying gpt-3.5-turbo is better than gpt-5.2 makes me think something you got some of them hidden motives.

https://code.claude.com/docs/en/common-workflows#use-extende...

>>prodig+(OP)
’’’Phrases like “think”, “think hard”, “ultrathink”, and “think more” are interpreted as regular prompt instructions and don’t allocate thinking tokens.’’’

>>Comput+Y
They dont allocate thinking tokens but they do change model behavior.

>>prodig+a1
I was getting this in my Claude code app, it seems clear to me that they didn’t want users to do that anymore and it was deprecated. https://i.redd.it/jvemmk1wdndg1.jpeg

>>Comput+q1
Thx for the correction. Changed a couple weeks ago. https://decodeclaude.com/ultrathink-deprecated/

>>prodig+C1
Nice blog, this post is interesting: https://decodeclaude.com/compaction-deep-dive/ Didn't know about Microcompaction!

>>sunaoo+Qm
If you're a big context/compaction fan and want another fun fact, did you know that instead of doing regular compaction (prompting the agent to summarize the conversation in a particular way and starting the new conversation with that), Codex passes around a compressed, encrypted object that supposedly preserves the latent space of the previous conversation in the new conversation.

https://openai.com/index/unrolling-the-codex-agent-loop/ https://platform.openai.com/docs/guides/conversation-state#c...

Context management is the new frontier for these labs.