DeepMind's New AI: A Spark Of Intelligence! 👌

📝 The paper "Creating Multimodal Interactive Agents with Imitation and Self-Supervised Learning" is available here:

❤️ Watch these videos in early access on our Patreon page or join us here on YouTube:

🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Aleksandr Mashrabov, Alex Balfanz, Alex Haro, Andrew Melnychuk, Angelos Evripiotis, Benji Rabhan, Bryan Learn, B Shang, Christian Ahlin, Eric Martel, Gordon Child, Ivo Galic, Jace O'Brien, Jack Lukic, Javier Bustamante, John Le, Jonas, Jonathan, Kenneth Davis, Klaus Busse, Lorin Atzberger, Lukas Biewald, Matthew Allen Fisher, Michael Albrecht, Michael Tedder, Nevin Spoljaric, Nikhil Velpanur, Owen Campbell-Moore, Owen Skarpness, Rajarshi Nigam, Ramsey Elbasheer, Steef, Taras Bobrovytsky, Ted Johnson, Thomas Krcmar, Timothy Sum Hon Mun, Torsten Reil, Tybie Fitzhugh, Ueli Gallizzi.

Károly Zsolnai-Fehér's links:

#DeepMind

[ReynaSingh]

Imagine a future where everyone has an ai maid/assistant. Reminds me of kazuo Ishiguro’s book klara and the sun

[Philafxs]

This innocent AI's mind is going to get blown when it figures out how many grapes there truly are.

[galgrunfeld]

Human: "tidy up the room"
Agent: "Impossible!"
Two papers down the line: adding a teen rebellious phase to AIs 😂

[garrettjones]

This reminds me of Piaget's work to learn how children learn.
You cannot just tell a child "clean your room". That's a very large and general task.
You start with something small like "do you see that teddy bear on the floor?" "yes" "pick up the teddy bear and put it on that shelf".
And after adding all these tasks together, the child then understands what "cleaning your room" means.

[FoxLivestreams]

Seeing AI really using its own vision is amazing to see, as a lot of game-based machine learning I've seen relies on hooking directly into the game's data. Seeing it actually look around to try and find the grapes was just absolutely amazing. The fact it can look around and move in a 3D space makes it seem more human, in a way. Thank you for the amazing videos, Dr. Károly Zsolnai-Fehér. The way you make such complicated research accessible for a variety of people is truly awe-inspiring, and really helps to inspire those that view your videos. With people like you, who make such amazing content for the love and passion for your field, it truly is a time to be alive.

[Teth]

It's important to note, as well, when we're comparing to human performance, that humans have been spending years learning many, many tasks that we don't even think about as tasks. For an AI to get anywhere with tasks like these learning completely from scratch in *minutes* is absolutely insane when you consider the years of learning that humans need to do for something as simple as reliably telling objects apart or realizing that things still exist when you aren't looking at them.

[juliandarley]

once again, it is remarkable how little training is required for either very good results or plateau to be reached. it seems to validate the experience from LLMs that scaling is considerably enhanced by increases in model and dataset size.

[Spyblox]

This AI reminds me of interactions I've had with people in games where they don't speak much English. Short somewhat grammatically correct answers that are a little off but make sense. This ai so far seems the most human like.

[TechnoMinarchist]

Imagine it.
Robot: "What is my purpose?"
Human: "To show me where my room is."

[pgetheelderscrollsturkiye]

Even the movements of the ai are so much human like, npcs doesnt move like this usually.

[Baggssy]

I feel like we are in the toddler years of AI and we are collectively raising them for the future much like a real child. As each year goes on they get a little smarter and in 20 years they may be as smart as an adult human.

[saugatkafley]

This DeepMind's AI has mastered the art of learning, my god what else it cant learn . Amazing and mesmerising.

[sm]

Imagine having an AI robot who doesn't need constant internet connection to work, man I can talk to it for days

[Kram]

utterly brilliant. Really enjoy how far multimodal techniques are being pushed lately. We've come such a long way.

[Mike..]

It's insane how quickly we've arrived here. And even more insane that it will only get better over the coming years. Having humanoid robot assistants is seeming more feasible every paper 🤯

[mawungeteye]

Imagine if DALLE-2 could be used to create training data for sim2real . How epic would that be?

[publicspeaker]

Wait holy sh!t this is game changing, this AI could be used in the real world. in the virtual one to help older people learn how to use tech and help with people who have dementia in the real world.

[graigirwin]

This is the next step to self-animation. Read it a disruptive movie script and Voilà, a movie appears. Maybe tweak, and enjoy.

[ScottHarden]

At around the 5:25 mark, you talk about the "nice linear growth" however the x-axis appears to be on a logarithmic scale, so isn't it actually logarithmic growth? I love the videos! Thanks!

[eclipse]

"Agent, kill all humans."

Agent: "OK."