“Leopold Aschenbrenner's "un-hobblings" framework posits that current AI models are like "baby AGIs" held back by artificial constraints such as a lack of training data for specific environments like Slack or Gmail.”