Latest / Logosofia no Dia a Dia
Who Can Actually Train an LLM From Scratch?
Daniel asked a simple question: how many organizations worldwide have trained a large language model completely from scratch? The answer turns out to be two questions wearing the same coat. At the frontier, the count is unverifiable — labs have stopped disclosing training compute. At the 1–70B parameter scale, the count is rising fast as sovereign initiatives in Korea, Switzerland, Brazil, Denmark, Norway, Portugal, and Iran ship from-scratch models. Meanwhile, the eight largest training runs since January 2025 are all American, and the top one is roughly 25x the compute of the largest…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.