The next frontier: AI that can plan ahead

Via Superhuman AI

“Most chatbots can tell you what the weather is like. They can even contextualize climate trends from the past few decades. But ask them something more complicated — like how economic and social patterns might affect the climate in the future — and they’ll struggle to come up with an answer.

That’s because AI models aren’t very good at anticipating things: We take it for granted that we can predict how different scenarios might play out before making a decision. How do we do it? By recalling past experiences.

GPT-4 and its competitors, on the other hand, are only capable of “thinking” a few steps ahead. They especially struggle with math equations and logic puzzles, since those tasks require a lot of trial and error.

That could soon change:

Execs at both OpenAI and Meta recently told the Financial Times they’re working on new models that show glimmers of reasoning capabilities
Both companies have been hazy on the details, but the basic takeaway is that each model will see improvements in their ability to remember what happened in the past
We won’t have to wait long: OpenAI’s long-awaited GPT-5 is reportedly coming “soon,” while Meta’s Llama 3 should be out in a few weeks

We’re already seeing a lot of improvement: For example, ChatGPT used to be pretty bad at chess. It sometimes tried to make moves that aren’t even allowed. But between GPT 3.5 and GPT 4, something changed. Now, ChatGPT can beat about 50% of ranked players.

If the new models live up to the hype, it would make us one step closer to achieving Artificial General Intelligence (AGI) — the fabled point when AI can compete with humans in any cognitive task.”

Posted on: April 17, 2024, 9:40 am Category: Uncategorized

By: Stephen Abram

rev="post-52647" No comments

0 Responses

Stay in touch with the conversation, subscribe to the RSS feed for comments on this post.

The next frontier: AI that can plan ahead