Leaked for maximum power of suggestion because a blog post without strict definitions for Reasoning would cop a lot of flack. This way you are hyped to think of their progress as totally encompassing your widest definition of reasoning that no one else has yet achieved.
Interpretation of "reasoning" looks to be a matter of opinion, like AI and AGI. If it's sequential application of some rules, then there is a lot of 'reasoning' software around. Almost all of it. There is no strict boundary where it becomes 'AI'. However, even stochastic parrots like ChatGPT can be extremely useful and outperform humans on some tasks. So, OpenAI talking about reasoning actually didn't say much. Their GPT is already doing it at some level, it can even spell it out step by step.
Is it named after the “create a copy of the strawberry on this plate, identical at the cell level but not at the molecular level” task (“at the cell level” meaning, has a cell of the same kind at corresponding locations, but ‘what molecules are where’ does not matter to the task.) ?
"... with the aim of enabling the company’s AI... to plan ahead enough to navigate the internet autonomously and reliably to perform what OpenAI terms “deep research...”
Nice... let the AGI loose on the internet... what could go wrong? :-P
Prompt: "ChatGPT, can you give me the phone and text records of all AT&T customers?"
I still have no idea how they practically want to solve the logic problem with their simplistic approach, without something like cyc.
They have no way to prove or disprove partial factoids in their token streams. Even if they manage to find those factoid regions, as explained in their paper. How is their feedback loop driven? Who gives the correct answers?
They can simplify and abstract speech, OK, but reasoning needs some logically driven feedback loop. Knowledge.
And no, Wikipedia is not a source of truth. Highly manipulated.
I was kinda hoping to incite discussion about reasoning AIs and how they could work, hence my STaR callout, but I think we're all equally in the dark here.
Previous HN discussion about "Quiet-STaR", 280 pts with 265 comments https://news.ycombinator.com/item?id=39713634