DALL·E 2 and Moravec's Paradox 🤖🧠🖼️

Hi guys,

I am following a discussion on Twitter, where people are sharing interesting takes on intelligence.

You can find said Twitter thread by clicking on this Screenshot, however I can provide you with a little context myself:

Context - this week is packed with impressive AI publications:

I personally had a hard time catching some rest this week because my Twitter feed informed me about the release of new, astounding publications literally every day.

It started on Monday with the release of a large language model (LLM) based on Googles Pathways system, PaLM. It is a 540-Billion parameter language model and claims to have advanced several capabilities on language and reasoning:

On Tuesday, robotics and AI researchers from that same company release a paper, SayCan, which uses language modeling for helping robots to understand, plan and execute on long-horizon natural language instructions. Here are some examples:

After Tuesday came Wednesday - and let me tell you, Wednesday was intense. OpenAI released DALL·E 2, which they describe as an “AI system that can create realistic images and art from description in natural language”. I just want to show you some AI generated pictures with their respective prompts:

You can find many more examples on their project website or on social media! These images look incredible to me and started a lot of discussions around the state of artificial intelligence, deep learning “breaking walls” and connections to human intelligence.

Finally we arrive at today, Thursday - the week has not even gone by yet! Today researchers introduced Socratic Models, a method that allows individual foundational models to talk to each other in order to do multimodal tasks like zero-shot Q&A or image captioning. This seems really promising for further scaling the impact of large machine learning models!

People now wonder - what does this mean for the future of artificial intelligence?

This leads us back to the initial question, where people discuss what kind of skills are easier to learn for an AI. Unfortunately I cannot give credits to whoever created this meme (found here), but it kind of started off like this after people shared images generated by DALL·E 2:

Also Sam Altman from OpenAI released a blogpost on DALL·E 2 where he stated the following:

It’s a reminder that predictions about AI are very difficult to make. A decade ago, the conventional wisdom was that AI would first impact physical labor, and then cognitive labor, and then maybe someday it could do creative work. It now looks like it’s going to go in the opposite order.

In response to that I found the mentioned discussion on Twitter, which goes as follows:

Feel free to read through the thread yourself here.

My question: is this statement true?

I am wondering how people from different sciences would respond to this. Is mastering motor intelligence harder than neocortical intelligence?

Does motor intelligence actually require (approx.) 10x more neurons and did it take 100-1000x longer to evolve?

More precisely, I would be curious to learn how we can measure or estimate the amount of neurons required to learn or solve a particular task?

Is there more literature on this?

To give credit, I read this question from Eric and started wondering what the folks at ResearchHub can add to this discussion.

Preprint

Preprint

DALL·E 2 and Moravec's Paradox 🤖🧠🖼️