ChatGPT skeptics picked the wrong camp: GPT learned logic. Period.

Here’s the take homes from an April 2023 no-cameras Berkeley Microsoft talk on GPT’s capabilities. And how this confirms and expands our collective thinking on emergence of intelligence in LLMs.

Heather West just posted this on my LinkedIn feed. CREDIT | LinkedIn / Heather West

Let’s start with Heather West’s (Execs on Deck) notes, posted on LI, from this talk on GPT by Microsoft’s Sebastian Bubeck at University of Berkeley :

• GPT-4’s ability to reason, discuss and interact “is the essence of intelligence.”
• GPT-4 is “absolutely superhuman at coding” and “its writing capabilities are astounding.”
• Bubeck used GPT-4 to artfully resolve a disagreement with his wife.
• GPT-3 has the intelligence level of a 7-year-old, while GPT-4 is indistinguishable from an adult.
• While ML is about teaching a machine a well-defined task, this is not about mimicking human learning. Given the scale of the data, “We’re trying to mimic something more akin to evolution.”
• Tools that have been used in ML are now almost irrelevant in light of GPT-4, Bubeck argues.

‘Bottled cognition’

This is fascinating and confirms all our thinking and testing reported here and by others. And the findings outlined in last week’s Microsoft preprint…