AGI?

AGI?

AGI?

Tuesday, April 18, 2023

AI will plateau until the next breakthrough, which may take at least a decade.

Sam Altman suggests that we are reaching the end of the era of enormous AI models, meaning that simply creating larger models like "GPT-5" may not lead to the improvements that social media platforms anticipate. Instead, memory and vision could be the keys to unlocking AI's potential.

Memory and vision, mon ami!

Let's begin with vision.

Humans experience the world through our five senses, which provide the information we use to form our understanding of reality. Vision represent 80%.

Current models have been train on our output which a compress version of reality.

For Ai to 10x from here, we will need to train it on reality, with vision ability, not human output.

When editing an image you want a raw version. Same thing here.

Right now, DINOv2 from Meta is the most promising breakthrough, Tesla being too secretive to tell.

We all thought the world needed iRobot but we really want is 32k Token.

Token are what define memory for GPT.

We, humans are short and long term memory that help us think in a very large spectrum.

Without large memory, Ai will be very limited when it come to learning on a new environment.

The tricky part? More memory needs a ton of computing power.

We might need some breakthroughs, maybe even quantum computing, to create a true AGI – that's Artificial General Intelligence.

So, don't stress too much, d'accord? We've got some time before AI takes over the world.


Voilà!

Personal letters

Most of these letters are intimate and can't be find anywhere else

Personal letters

Most of these letters are intimate and can't be find anywhere else

Personal letters

Most of these letters are intimate and can't be find anywhere else