There’s this gold rush going on, you’re right, any B without RLHF is meh. The things getting published as “on device LLM” focus on bitcrushing the lowest B mod

There’s this gold rush going on, you’re right, any B without RLHF is meh. The things getting published as “on device LLM” focus on bitcrushing the lowest B model with minimal RLHF and then pronouncing we have on device LLMs. We’ll definitely get there but signal >>> noise currently. First person to admit this and write their blog post with A / B tests vs. a Markov chain deserves the gold.