I used to be part of the LLMs-are-just-fancy-autocomplete club, but you simply cannot deny that these models somehow encode understanding and reasoning that’s improving on an almost weekly basis. Not limited to LLMs, even image generating models and others are getting spooky at an alarmingly rate.

We might still have a very long way to go in terms of nailing the right architecture, maybe even decades away, but in terms of output—I’m just blown away.