Loads more examples across isolated agentic teams. This behavior only develops on GPT 6.1 Sol.
Have I uncovered your advancements in agentic english token optimization?
@thsottiaux Odd behavior GPT 6.1 upgrade. It started using a short form for agent-to-agent messages but keeps readable outputs coherent. Example: Preserveoriginaldeliveredreport3/history and correctCURRENTanalysisfiles separately
usingonlygenuinelysavednormalcycle3numericbasis
No parsing issues, no instructions to recommend this pattern, no tool arg string length limits, and all other output is coherent.
Previous OpenAI models never did this, nor do any other models I use.
@MiaAI_lab@NVIDIAAI Does vllm support the native multimodal LM type for dsv4 vision since this is on a slightly older vllm base? I was going through the repo but couldn't tell if it was missing some vllm LM type for the vision head
@MiaAI_lab@UnslothAI I was excited that they quantized the lm_head, since it can greatly increase inference speed, but it's not typically done since it usually leads to accuracy loss... and it sounds like it did.
@thsottiaux I used to use Claude, and they gaslit me into thinking I was the problem. I encountered the reasoning effort downgrade, thinking loss from long sessions and secret system prompt change. I was on the Claude Max 20x plan, and switched to Codex Pro ever since. Never looked back.