Andrej Karpathy on Why you should work on AI AGENTS!

This is not my video, but it was so good and compelling I decided to share with the community.

Andrej Karpathy, a legend in building AI systems encourage people to work on AI Agents and why small timers like us might have better moat at it than big corps like OpenAI.

❤️ If you want to support the channel ❤️
Support here:

[leo.budimir]

"There is a large class of problems that are very easy to imagine, very easy to build demos for, but are actually extremely hard to make products out of. For example, self-driving cars: it is easy to imagine, very easy to build a demo of a car driving around the block, but making it into a product takes a decade. Same for VR. AI agents are probably like that as well."

[seekingtroooth]

🎯 Key Takeaways for quick navigation:

00:24 *🕹️ In 2016, AI focus was on RL agents, especially in gaming contexts like Atari.*
01:58 *🔄 Shift in approach: From AI agents to language models; unforeseen trajectory.*
02:42 *🌐 AGI likely as AI agents in diverse forms, part of organizations or digital civilizations.*
03:25 *⚙️ Building AI agents challenging; requires long-term commitment beyond exciting demos.*
03:52 *🧠 Drawing inspiration from Neuroscience for cognitive tools in AI agents.*
05:29 *🏛️ AI agents builders at the forefront; labs like OpenAI focus more on language models.*
06:27 *👏 Building AI agents is inspiring, being at the edge of capability in a transformational field.*

[ZenMasterNahc]

I'm on the same page. Going into the future people will need their own Ai agents to combat or defend against rogue Ai agents from both known and unknown sources and origins.

[mohsendowlatshah]

Thank you for the video. It is amazing listening Andrej's talks on AI

[MaJetiGizzle]

Great insights from Andrej! Thanks for sharing this video!

[eck]

Cool, so that's what Andrej is trying to create now. Real AI brains. That's kinda cool!

[ReddSpark]

Thanka for sharing this. I'm one if the entrepreneur/hacker thsts busy experimenting with Agents. Felt like im foing really cutting edge work but didnt realise others also realised this. Hopefully i get my solution to work so i can share it more broadly!

[majesticglue]

if this talk was done by a researcher not from OpenAI, i'd trust it more. He knows the best way to create AI Agents right now is through gpt since currently it has the best capabilities, and AI Agents are going to be a huge money maker for them if any tool does something good enough considering how you'd have to spam the crap out of their api to utilize an AI Agent. Much incentives for him to promote AI Agents.

[TheGaussFan]

Andrej knows Tesla FSD better than most everyone. He says its a 10 year problem. FSD was announced by Elon in 2016, so solved by 2026. Sounds about right. My model 3 is driving without intervention most of the time now.

[AndyPayne]

The hippocampus DIRECTS memories to regions but does not store them. Very important.

[denisblack]

Truly inspiring! Kinda shifted me from “omg we all gonna be replaced by an artificial ai economy” panic to “ok, let’s build some stuff for this ai economy”
Thank you

[Nick_With_A_Stick]

Awesome! Valuable information! Thanks for posting this!

[dzlfiqar]

saying got distracted for self driving is a next level of distraction :))

[werwardas]

Thank you for sharing!
Anyone knows where to find the newest papers regarding agents and transformers to keep up?

[krunalkshirsagar]

Don't know if his talk is leading to spiking neural networks or something. Since, he mentioned about brain and hippocampus. Idk i could be wrong

[robertmazurowski]

Making Chat GPT solutions is easy :) the limitations are in the models themselves, which you are depended on big tech companies. Someone with 10 Milion USD and will always make a more successful business than someone with zero, especially if it is easy. Someone with 1 billion will make a way more popular app than 10 milion one. If something is technically easy, then money decides.

If you do a startup as a small team, concentrate on something that requires a very specific set of knowledge, skills or experience. Something like this is technically hard to replicate even from someone with 10 milion.

[tvm]

Could someone post a link to the book he referenced?

[alienkishorekumar]

Clearly it's just chance. No one can predict which model would pick off. Of course, also seen only from Hindsight.

[mostlazydisciplinedperson]

We currently building this hippocampus of ai agent

[wanghaitong]

I wonder what does AI Agent really mean? Reinforcement Learning? Or a new concept?