No chapters are available for this episode.
Part 3 of 5: Training and emergence, or the birth of an intelligence.
We force-fed a machine 10 million books. Then we taught it humanity.
The result? GPT-3 started doing advanced maths. Without being taught. It translated between programming languages. Detected invisible emotions. Wrote sonnets under impossible constraints. Then GPT-4.5 arrived. Worse than its predecessor on some tests. For no apparent reason.
As if these models had their own cognitive weather. Between 2018 and 2022, behind closed lab doors, we created something we do not fully understand. Something that develops abilities nobody taught it. And sometimes loses them, without anyone knowing why.
Cost of the operation: 200 million dollars. 25,000 GPUs. 20,000 hours of human work to teach it manners. If their creators discover abilities after the fact, what other surprises do they have in store for us?
Find the full transcript of this episode, along with all the sources and plenty more, at: How ChatGPT revolutionized the world (3/5)