I just discovered this channel, I am going to binge watch the podcasts!

@izzy2900

2 ай бұрын

binge 🙃

@zoe.tsekas

2 ай бұрын

@@izzy2900 typing too fast!

I was thinking about that like the day before I heard this clip. It's easy to overlook the point of what he's saying, but it's actually quite significant. The text data we have is produced by humans and obviously it applies to video content (think youtube) in some manner. But video can also be objective. That's where it can provide grounding because it doesn't have this human subjectivity to it. Like the example of deriving physics and relations of cause to effect. And, these AI systems will interact with people and the world and learn from it. We'll also have them spin up experiments both in simulations and the real world.

We are going to need more GPUs

rap god

Patels are evolving

The human brain isn't multimodal though. There's a pattern recognizer for audio, and a different one for visual. Why not follow evolution's lead and create an ensemble of transformers, not a composition.

@2false637

2 ай бұрын

It is tho. You can listen to something and imagine an image of it, or smell something and describe its taste etc

@anthonybongo94

2 ай бұрын

​@@2false637 don't conflate your conscious experience with the actual neural wiring. you could just as easily explain that experience through different parts of the brain working at the same time as you could one part working together

@user-cg7gd5pw5b

2 ай бұрын

Because it's inefficient? Sure our brain is still way better, but the results we get from multimodality far exceed the ones we have without it.

@anthonybongo94

2 ай бұрын

@@user-cg7gd5pw5b idk man. I'm betting that in the long run the ensemble will win. Google is beating the metrics with their models rn, but that could be better data... More params... We'll see

@spookyrays2816

2 ай бұрын

You said it wasn’t multimodal then described multimodality

Gemini aint doing so well 😂