Raw LLM Responses

Yeah, I don’t buy the fact that Anthropic is giving the military tech that’s 1 year ahead of consumers. A pretrain run is tens of millions of dollars of compute. That’s the difference between Deepseek V3 and Deepseek V4. Anthropic would almost certainly give consumers the output of a successful pretrain run as soon as possible, not 1 year later. In fact, OpenAI is getting shit for using the same pretrain run for over a year. GPT-5.2 is the same pretrain as 4o and that’s very out of date. Anthropic is NOT giving the military a finetune- they don’t need to finetune their models for them- but rather just need to do a separate posttrain run for the military, sans safety training RLHF. This may very well be better than public Claude but it won’t be one full year better.