I have decided that Andrew Ng is, to me, the coolest dude on the planet. The walking definition of Theodore Roosevel's "speak softly and carry a big stick", but in the world of AI.
"Intro": "00:00:00", "Faster ML Model Development": "00:01:48", "Gen AI vs Supervised Learning": "00:02:05", "Consequences of fast development": "00:03:12", "Agentic AI Workflows": "00:06:40", "Agentic Reasoning Design Patterns": "00:09:56", "Reflection With LLMs": "00:10:31", "Tool Use": "00:12:23", "Planning/Reasoning": "00:12:53", "Multiagent Collaboration": "00:13:31", "Demo: Agentic Workflow and Video Q&A App": "00:16:13"
I am a coursera alumni of his ML intro course. He really got me hooked on ML & from that I got the confidence to build several ventures using ML. Dude is the Joda of ML/AI.
one of the few voices out there actually inspiring people to Build with Ai ... its like when the app store started and there was only a few devs there.
*AI as the New Electricity* (00:00:00) Andrew opens by comparing AI to electricity as a general-purpose technology, setting the stage for discussing AI's widespread opportunities and applications. *The AI Stack and Opportunities* (00:00:54) *Fast Machine Learning Development* (00:01:51) *Agentic AI Workflows* (00:06:32) *Visual AI Applications & Demos* (00:15:08) *Vision Agent Demo Applications* (00:20:00) *AI Stack and Agentic Orchestration* (00:22:30) *Four Important AI Trends* (00:23:25) *Conclusion and Next Steps* (00:25:56)
Gold not giving me free hard work here after some buying gold for kids future no one gets no one gives and not demand any one last 19years old they're decided me know but everyone needs i give them always New ideas i am not servant of any them normal person we are not Big amunt only some money for family kids are better life i am not rich and not demand any one thank you 🙏
Excellent overview Andrew. I'm always amazed at how much you know, and how you can explain what you know in ways that the rest of us can understand! Thanks for the overview!
The exploration of AI agents and agentic reasoning highlights an exciting paradigm shift. Frameworks like KaibanJS are helping to simplify the implementation and management of multi-agent systems, especially for handling complex workflows. Looking forward to how this evolves.
AI stocks will dominate 2024. Why I prefer NVIDIA is that they are better placed to maintain long term growth potential, and provide a platform for other AI companies. I know someone who has made more than 200% from NVIDIA. I'll also take any other recommendations you make.
I think the next big thing will be A.I. For enduring growth akin to META, it's vital to avoid impulsive decisions driven by short-term fluctuations. Prioritize patience and a long-term perspective consider financial advisory for informed buying and selling decisions.
Facing a similar situation, I sought advice from an advisor. Through portfolio restructuring and diversification with good ETFs, S&P 500 and growth stocks, I've turned my portfolio around from $200k to over $800k in a few years.
@@hasede-lg9hj Your advisor must be really good, I hope it's okay to inquire if you're still collaborating with the same advisor and how I can get in touch with them?
Rebecca Lynne Buie is the licensed coach I use. Just research the name. You'd find necessary details to work with a correspondence to set up an appointment.
it took me weeks hen I started to build a basic chatbot with ai, now I just built a ai code editor than can one shot it with custom models and API keys! I love this stuff.
I could not understand the difference between using 1 A.I. and splitting it into agents vs using multiple A.I. acting as agents until I delved into the fine tuning process. Now I completely understand it.
At 11:19 lot of the times by my own experience: LLM confuses itself with a additional question and starts creating even more complicated code or fixes things that are working. Basically loops itself and cannot find the right answer anymore. LLM are awesome, its always the last 20% that is the most painful.
@ i guess full runtime env to trial and error and set of validator agents that check for increase in complexity and others track if solutions do not repeat itself. That can get us to last 2-5%. That 5% will be hardest.
isn’t agentic method the same as the chain of thought, but instead of a thought experiment (think out loud), it gets the empirical result as the (work out loud)
Andrew, can we have multi agent collaboration with Gen ai agents with embedded open source LLMs like Llama 3 or API calls to commercial LLMs for infrequent calls, and Deep Reinforcement Learning agents for very frequent sequential decision making with very low latency in order to build intelligent autonomous systems (with graph workflows to model complex processes). An example could be like urban traffic management systems with each intersection being managed by an agent in conversation with other agents at other intersections. Alternatively, we could use orchestration for structured workflows like industrial automation.
Managing unstructured conversations between traffic agents in the above example would, I think, involve techniques similar to network optimization for the internet to optimize urban traffic management system performance, efficiency, and reliability. Structured workflows like industrial automation, in contrast (consisting of API calls, function calls, scripts, and orchestration), would be relatively simpler in comparison.
20:57 Conveniently sliding surveillance use case into a demo. If the metadata is human bio data and the context videos are CCTV footages, what does it say about the state of privacy in the age of AI?
agentic workflows are connected to others models to do tasks like execute code in the terminal, surf the web, etc. COT alone is just within the LLM without access to other models
AI Agents scare me a little, they are very efficient in narrow domains at the cost of the big picture. LLMs can do correlations across domains that AI agents can't, although great in their domain they will lack in discovery and creativity needed for the progress of science. Basically losing the big picture.
I disagree. Agents are just one part of the puzzle, and you could realistically build a whole ecosystem combining both (see LangChain, LlamaIndex, etc) to get the best of both worlds. I think the future is essentially going to be a huge spider web of agents + llms doing their thing in concert. It's an exciting time to be a tinkerer!
@@VikBrummer Thank you for your interesting perspective. You may be right as it mirros with LLMs and the transformers they are made of. Wishing you are right
@@PierreH1968 you're welcome! Most agents will just be small hyper specialised llms, but there's nothing stopping them from communicating with larger llms to get answers, run flows, etc. You'll find as we move forward in this space we have orchestrating llms that will control agents and llms together, getting what we need from both
@@VikBrummer I really hope you are right as transformers were already often biased. (mistaking humans and animals, (google) becoming biased from the reinforcement of bad actors (Microsoft bot) ...Still the mistakes were out of bad training sets easily spotted by experts. It allowed the creation of LLMs to forge humanistic visions. But Mini LLMs in the form of agents from diverse origins, easier to own and reinforce with smaller hardware could be much more insidious and opiniated even with supervision. The fake open sourcing we have already on AI (since we don't have access to the training set) is making me weary of the Agents, that rich opportunists could turn into Mr Hyde ... You seem to have a good insight on this. I'd love to know more about the pillars of your optimism.
@@PierreH1968 you're absolutely right on all of that, but the same could be said for any of these llms or otherwise. We're going to need to ensure we build the governance, ethics and guard rails in early to temper / mitigate these things. One thing I'm working on is to build out a few small agents who will handle this type of stuff at work so that the question or action never reaches the end before going through the agents. At the end of the day agents are going to be nothing more than tools we use, so you just gotta make sure use / build the right ones
VERIDA $VDA is the same thing, but stronger All in in $VDA 100x potential (ranked 2600th= marketcap of $2m) $0,014/token, this will pump violently during next altcoin season, making eyes on it
chatgpt wasn't even getting simple maths right the other day. felt totally let down. and glimpsed a future where we rely waaaay too much on AI and don't notice the mistakes it makes till way too late
I am surprised that he hasn't focused on any analytical methods to evaluate these tools. I think these tools are amazing but data scientists like him should focus on building confidence in these tools and not on the happy paths only.
@dusanbosnjakovic6588 I think his and others' expertise has led us right here. Llms are at the moment like a black-box. You simply poke it and see what comes. Plus, I think the end goal for ai is to have a system that remarkably understands natural human language.
@@KamyaSamuelElder he did help get us here, but now the skills that are needed to do what he is suggesting gave morning to do with his expertise. He is talking about modern app development, not data science.
He is definitely same accent, same pace, same speaking style as i watched in Coursera learning. I like him !!!
I have decided that Andrew Ng is, to me, the coolest dude on the planet. The walking definition of Theodore Roosevel's "speak softly and carry a big stick", but in the world of AI.
"Intro": "00:00:00",
"Faster ML Model Development": "00:01:48",
"Gen AI vs Supervised Learning": "00:02:05",
"Consequences of fast development": "00:03:12",
"Agentic AI Workflows": "00:06:40",
"Agentic Reasoning Design Patterns": "00:09:56",
"Reflection With LLMs": "00:10:31",
"Tool Use": "00:12:23",
"Planning/Reasoning": "00:12:53",
"Multiagent Collaboration": "00:13:31",
"Demo: Agentic Workflow and Video Q&A App": "00:16:13"
Thanks for sharing this!
Andrew Ng is the person whom I follow, trust, and look for new information on AI. Thanks.
Me too. All least he was. However there is not much about this moment that has to do with his expertise any more.
Me too! He IS the source of the Truth about AI
I am a coursera alumni of his ML intro course. He really got me hooked on ML & from that I got the confidence to build several ventures using ML. Dude is the Joda of ML/AI.
one of the few voices out there actually inspiring people to Build with Ai ... its like when the app store started and there was only a few devs there.
I like Ng because he gives practical information everyone in the field can use, in a very humble way.
Andrew is inspirational. I find his talks thought provoking.
This guy is making a true contribution to the world in this era. Cheers, Andrew!
*AI as the New Electricity* (00:00:00)
Andrew opens by comparing AI to electricity as a general-purpose technology, setting the stage for discussing AI's widespread opportunities and applications.
*The AI Stack and Opportunities* (00:00:54)
*Fast Machine Learning Development* (00:01:51)
*Agentic AI Workflows* (00:06:32)
*Visual AI Applications & Demos* (00:15:08)
*Vision Agent Demo Applications* (00:20:00)
*AI Stack and Agentic Orchestration* (00:22:30)
*Four Important AI Trends* (00:23:25)
*Conclusion and Next Steps* (00:25:56)
There's so much gold in this short video that I have to watch it multiple times. Thank you Andrew Ng!
Gold not giving me free hard work here after some buying gold for kids future no one gets no one gives and not demand any one last 19years old they're decided me know but everyone needs i give them always New ideas i am not servant of any them normal person we are not Big amunt only some money for family kids are better life i am not rich and not demand any one thank you 🙏
Excellent overview Andrew. I'm always amazed at how much you know, and how you can explain what you know in ways that the rest of us can understand! Thanks for the overview!
This was so a great talk. Love the little demo site at the end too, very inspiring.
Andrew is GOAT into LLM- AI world
The exploration of AI agents and agentic reasoning highlights an exciting paradigm shift. Frameworks like KaibanJS are helping to simplify the implementation and management of multi-agent systems, especially for handling complex workflows. Looking forward to how this evolves.
The insightful and thorough discussion on Ai agentic make this video content intriguing. 😎💯💪🏾👍🏾
Yet another great video - Thanks Andrew!
Andrew Ng is the living Demi God Of AI on earth. This man & his brilliance is truly incredible.
Great talk! Andrew Ng
Great talk, awesome content, Andrew is that "guy"
AI stocks will dominate 2024. Why I prefer NVIDIA is that they are better placed to maintain long term growth potential, and provide a platform for other AI companies. I know someone who has made more than 200% from NVIDIA. I'll also take any other recommendations you make.
I think the next big thing will be A.I. For enduring growth akin to META, it's vital to avoid impulsive decisions driven by short-term fluctuations. Prioritize patience and a long-term perspective consider financial advisory for informed buying and selling decisions.
Facing a similar situation, I sought advice from an advisor. Through portfolio restructuring and diversification with good ETFs, S&P 500 and growth stocks, I've turned my portfolio around from $200k to over $800k in a few years.
@@hasede-lg9hj Your advisor must be really good, I hope it's okay to inquire if you're still collaborating with the same advisor and how I can get in touch with them?
Rebecca Lynne Buie is the licensed coach I use. Just research the name. You'd find necessary details to work with a correspondence to set up an appointment.
I just looked her up on the web and I would say she really has an impressive background in investing. I will write her an email shortly.
This is Genius! I know someone who needs help transcribing podcasts, but AI could do that and store it in pandas
going crazy about this videio, i really think the word need to see this!!!!!
world 🗺
it took me weeks hen I started to build a basic chatbot with ai, now I just built a ai code editor than can one shot it with custom models and API keys! I love this stuff.
Such an Informational and Insightful Lecture.
Kudos & Thank you for this incredibly insightful and thought-provoking video❤
Andrew Ng 🎉🎉🎉
Love it! Thx for sharing!
Very good!
Amazing talk
This man is incredible!
Great NG
AWESOME! THIS VIDEO = GOLD
I could not understand the difference between using 1 A.I. and splitting it into agents vs using multiple A.I. acting as agents until I delved into the fine tuning process. Now I completely understand it.
Can you explain how fine tuning helped you understand it? Any examples you can share you did with fine tuning?
The Goat 🔥
That baggage demo is super cool. Can anyone suggest any blog post or repo on how to implement that?
Let me know if you've found something on it!
Merci pou tubi wi
Wonderful, thank you!
His calmness is always brilliant . Such a wonderful insight as always
At 11:19 lot of the times by my own experience: LLM confuses itself with a additional question and starts creating even more complicated code or fixes things that are working. Basically loops itself and cannot find the right answer anymore. LLM are awesome, its always the last 20% that is the most painful.
@@benmaxinm well said. That problem a framework hasn't yet addressed. What's the path to optimization from 80% to 100%?
@ i guess full runtime env to trial and error and set of validator agents that check for increase in complexity and others track if solutions do not repeat itself. That can get us to last 2-5%. That 5% will be hardest.
@@benmaxinm The true next level is having them learn to not have to take many iterations but to learn from previous mistakes
@ true, but that remaining 5% requires inteligence. Thats still miles away.
Great teacher😀
Such a good video
Awesome talk!
Great talk by Andrew. But please add English subtitles.
our techy ip man!
It wasn't lost on me, particularly the football examples and that yellow world atlas - keep up the good fight XXXXX[]
Excellent Andrew Ng, Where does Inference like GROQ will fit in this stack ?
Nice
I’m really interested in ai agents and am looking to connect other people also getting into this. ❤
isn’t agentic method the same as the chain of thought, but instead of a thought experiment (think out loud), it gets the empirical result as the (work out loud)
Wow these visual agents can solve so many crimes so fast!! Amazing...
Can it count the spectators in that image?
This is from May, 2024 - 6 Months Old. You shoud note this in the description.
He mentioned anthropic claude computer use at 24:35
Andrew, can we have multi agent collaboration with Gen ai agents with embedded open source LLMs like Llama 3 or API calls to commercial LLMs for infrequent calls, and Deep Reinforcement Learning agents for very frequent sequential decision making with very low latency in order to build intelligent autonomous systems (with graph workflows to model complex processes). An example could be like urban traffic management systems with each intersection being managed by an agent in conversation with other agents at other intersections. Alternatively, we could use orchestration for structured workflows like industrial automation.
Managing unstructured conversations between traffic agents in the above example would, I think, involve techniques similar to network optimization for the internet to optimize urban traffic management system performance, efficiency, and reliability. Structured workflows like industrial automation, in contrast (consisting of API calls, function calls, scripts, and orchestration), would be relatively simpler in comparison.
20:57 Conveniently sliding surveillance use case into a demo. If the metadata is human bio data and the context videos are CCTV footages, what does it say about the state of privacy in the age of AI?
How is chainofthought pattern is different than this Agentic workflow?
agentic workflows are connected to others models to do tasks like execute code in the terminal, surf the web, etc. COT alone is just within the LLM without access to other models
@@Geeisjudied would you say structure wise it's similar to microservices except we have agents instead of microservices?
@@dreadserpant yes, but with an LLM as the orchestrator of the micro services
AI Agents scare me a little, they are very efficient in narrow domains at the cost of the big picture. LLMs can do correlations across domains that AI agents can't, although great in their domain they will lack in discovery and creativity needed for the progress of science. Basically losing the big picture.
I disagree. Agents are just one part of the puzzle, and you could realistically build a whole ecosystem combining both (see LangChain, LlamaIndex, etc) to get the best of both worlds. I think the future is essentially going to be a huge spider web of agents + llms doing their thing in concert. It's an exciting time to be a tinkerer!
@@VikBrummer Thank you for your interesting perspective. You may be right as it mirros with LLMs and the transformers they are made of. Wishing you are right
@@PierreH1968 you're welcome! Most agents will just be small hyper specialised llms, but there's nothing stopping them from communicating with larger llms to get answers, run flows, etc.
You'll find as we move forward in this space we have orchestrating llms that will control agents and llms together, getting what we need from both
@@VikBrummer I really hope you are right as transformers were already often biased. (mistaking humans and animals, (google) becoming biased from the reinforcement of bad actors (Microsoft bot) ...Still the mistakes were out of bad training sets easily spotted by experts. It allowed the creation of LLMs to forge humanistic visions. But Mini LLMs in the form of agents from diverse origins, easier to own and reinforce with smaller hardware could be much more insidious and opiniated even with supervision. The fake open sourcing we have already on AI (since we don't have access to the training set) is making me weary of the Agents, that rich opportunists could turn into Mr Hyde ... You seem to have a good insight on this. I'd love to know more about the pillars of your optimism.
@@PierreH1968 you're absolutely right on all of that, but the same could be said for any of these llms or otherwise. We're going to need to ensure we build the governance, ethics and guard rails in early to temper / mitigate these things.
One thing I'm working on is to build out a few small agents who will handle this type of stuff at work so that the question or action never reaches the end before going through the agents.
At the end of the day agents are going to be nothing more than tools we use, so you just gotta make sure use / build the right ones
Sure iterative multi-agent AI will give better results but what about the ICER? Probably only makes sense for complex tasks?
I would put this in propotion to market share. Nvidia has 99% of the stack. Perhaps OpenAI has a similar lead in the foundation model
Building a smarter world with AI
VERIDA $VDA is the same thing, but stronger All in in $VDA 100x potential (ranked 2600th= marketcap of $2m) $0,014/token, this will pump violently during next altcoin season, making eyes on it
As time passes, Andrew himself is behaving more like an AI Agent!. Is he even real ?! 😉
chatgpt wasn't even getting simple maths right the other day. felt totally let down. and glimpsed a future where we rely waaaay too much on AI and don't notice the mistakes it makes till way too late
Has he written any books on ai
and now I know how to pronounce Ng.
His name reminds me angular dependencies.
!
🙌🙌🤲🤲🙌🙌🙏🙏😢😢
gah, too much asmr in the microphone... good speaker tho
I am surprised that he hasn't focused on any analytical methods to evaluate these tools. I think these tools are amazing but data scientists like him should focus on building confidence in these tools and not on the happy paths only.
It's currently an art to get proper prompting
@KamyaSamuelElder that's right, and not a science. Why would Andrew Ng be an expert in this field we listen to?
@dusanbosnjakovic6588 I think his and others' expertise has led us right here. Llms are at the moment like a black-box. You simply poke it and see what comes.
Plus, I think the end goal for ai is to have a system that remarkably understands natural human language.
@@KamyaSamuelElder he did help get us here, but now the skills that are needed to do what he is suggesting gave morning to do with his expertise. He is talking about modern app development, not data science.
AI ≠ ML