Anthropic is increasingly sounding like the dudes ~40ya who were were absolutely certain they could get to any arbitrary degree of compression by just iterating. Thankfully, they couldn’t get funding…

You cannot replicate a 300b LLM onto my washing machine’s 32MB of flash. It cannot run in my router’s 256MB of RAM. The concept that an advanced model can somehow make itself small enough to run on tiny devices is absurd. LLM smears the code/data boundary and encode sthe information of both.