from @AravSrinivas (perplexity CEO): hybrid agentic inference is how we maximise token value per watt per user local model on intel core ultra series 3 -> handles confidential files