I use this method of agentic workflow for coding in school, with brainstorming, and detailed pseudo steps, then review it, then stich somethings together, and have it revise itself. Then after a few cycles, I give it a request using words like "magnificent," or "swe professional point of view," and other things along those lines. Which ends up give me a better product by the final iteration. I am a Sophomore CS student, and I'm glad I am doing the method right, when I'm practicing data structures. Thanks for explaining the method of what I have been doing Andrew Ng.
@@jatingupta4708 I would, but I feel I would need to give like a step process, that would have other bubbles next to bubbles for things to consider. If you give me a few days, Ill give you a generic image generated host paint of my workflow.
@@hl236 I only use UA-cam and disc for 99% of interactions. I have a git hub with student work I've done, but its nothing special, as its just college freshman and sophomore projects, keep building yall.
Hi guys, This is an excellent talk. I watched the Andrej one as well. And it seems that these talks are really focused and talk about real problems. So kudos to the team for executing on this. Its WAY WAY BETTER than these large conference interviews where you don't learn anything
Not limited to this event, I wonder what other talks stand out to you. I am trying to learn as much as I can about this space from experts. Many thanks
It's nice to see Andrew finally be impressed by something! A lot of his previous talks were along the lines of how AI can't do anything yet and it's not nearly capable, etc.
This is going to change development radically. Imagine being able to just write workflows to write tons of code based on what functionality we want while modulating interconnectedness, dependencies and other finer details and nuances which humans understand and A.I. does not. Just the best of human and ai combining to increase productivity. Excited for the future.
Brilliant. The only thing I would add is that with inference speeds offered by Groq, it’s not necessary to wait minutes anymore. Fast inference speeds have the most value when humans aren’t reading the output.
Insightful discussion on the future of AI agentic workflows with Andrew Ng. Excited to see what innovations are on the horizon. #AIAgenticWorkflows #AI #Innovation
I just did a quick test of this by customizing GPTs how I would like it to respond , its a good fame work , but i dont think its truly agentic . "Define the Goal: Be specific and measurable. What is the desired outcome of this workflow? Identify Key Milestones: Break down the overall goal into major checkpoints or phases. Flexible Action Steps: For each milestone, brainstorm a variety of potential actions that could move you forward. Avoid a rigid, sequential task list - these actions are options to choose from. Evaluation and Iteration: Determine how you'll measure progress towards milestones and the overall goal. Build in regular checkpoints for assessment. Be prepared to adapt actions, milestones, and even the goal itself based on results. Key Principles Outcome Focus: Prioritize reaching your goal over following a precise plan. Adaptability: Embrace change and be willing to pivot your actions as needed. Empowerment: This framework aims to give you agency and decision-making power within the process."
I am 54 this year coding for 3 decades, been using prompt engineering to create my code for last 2 months using multiple GPTs and thought it was the best, but this agentic loop will make programmers legacy, being slow and expensive In the next 6 months these workflows will improve, the rise of AI agents will be here whether we like it or not I advice knowledge workers to prepare for this financially, physically, mentally and emotionally This will be fast, we will be like deer looking at headlights
I LOVE AI Agents... but I'm left wondering: Why hasn't anybody developed a system/app that takes the API's from the top LLMs, created agents for each, and then have these agents all work together to brainstorm, debate, review, and solve problems? I often get 4 different answers from 4 LLMs, so why not have them all setup as agents "in one room" working together to come up with the "best" solution. I can't find anybody that's tried this... why not? Wouldn't having the "top minds" (LLMs) working together produce better results?
I think that the only think difficult is to make a good back-end for budget control because it would cost a lot on API and would be unavailable for people that are not willing to pay for this kind of website.
@@husnainarshed7806 that's a really great (and important) point. I think it could be mitigated by: 1) Having the smartest model be the "Manager" of many less-expensive "worker-bee" agents. 2) Setting a "max budget" for the query / project... e.g., "I need this task completed in x days and am willing to allocate $y budget. Do the best you can with the budget / time allocated!" Then the boss agent would be smart enough to budget (time and $$) efficiently. Say it has 10 days... it could use off-peak model times (I assume models will eventually start having time-of-use pricing tiers).
Currently, seems like "AI agents" are just a sequence of LLM requests. Take output and throw it back into LLM, with tools! Conditional loops seems to be the differentiating factor between these chains and real "agents" as I would define them. Any thoughts?
We think that this shift towards agentic workflow will lead to the need of rapidly and consistently train routers to quickly, and smartly route queries to squeeze to most performance of every of the agents. This is exactly what we aim to do here at Plurally. We are currently in an early access stage, but to anyone looking for such a solution or simply want to try it out - reach out to us (we are very responsive), would love to hear about your specific use cases!
Reflection is the correct word. Reflexion is a term created by the researchers, with a self-reflection methodology where the LLM reflects on the previous answers to improve it.
The core challenge here is that LLM fails at multi-step planning - and there is no way to guarantee that iterative Reflection bring correct solution (and optimal) in short period of time and money
To me this approach feels a bit like computer vision before neural networks took over. Hand coding feature detectors etc. That’s what you’re doing when you hand-design a workflow like “do web search, gather sources, write first draft, critique first draft” etc. These hardcoded agent flows are too rigid to produce good results generally. The models will learn to construct their own flows just like a person can. So while people might have some success building agents like this now, I think it’s a bit of a dead end that will be overtaken by foundational models.
Prompt instructions vs zero shot will always be a balancing act that depends on use case. Give too many instructions and you'll nerf the llm. Don't give instructions or use rag and you it will output responses that are low value and not actionable. However the value of an llm will always depend on a user's ability to ask the right questions. If Elon musk gave you 30min of his time the value will depends on what you ask him.
Americans hate to admit it but we've been under such a closed system for so long to the point our public education stopped enriching young minds 40 yrs ago and flipped into recruiting for agency & institutions taking away the most productive years from our workforce. It's been 90 years since we was in a very open private sector individual owners and operator creative posture. Everyone is a cog in the wheel more or less. We have so many grandfathered in economic middle men over the many phases of steam engine until today. Unlike most of the world that only individualized the past 50 years after the transitor age it leaves the west and America with many extra left over obstacles. It also leaves debts paid up front that have helped us get to this technology. Since I've retired and lived through all that 1900s, structuralism costs classical American decendants its only fair to remind everyone what these are from random Joe's perspective. Family birthrates ,18-30 year Olds trained up and entering workforce at the most creative and productive ages ( which by default tends to marry & help Maintain the elusive American prosperity) and the lack or loss of 31 -50 year old owner operators of local American infrastructure. Yes it's an unsustainable theme here that's been a very hefty price in building out our world over the 80 years of the transitor age. If any sectors are handed advantages in this new paradigm infrastructure, these are the ones who have paid the ultimate cost in my lifetime. New paradigm infrastructure where balance is there for better quality of life and at minimum restoration of all that's been compromised. We have so antiquated ways of doing things . Our city's are still under top down rule prohibition era reformed control mechanisms
Very good talk. I was only distracted by the guy with the restless leg syndrome in the front row. He must have been even more distracting to Andrew, who despite that delivered an awesome talk!
This was useless. No real content. Just guesstimations of very vague trends. Just say "Agent Models can wow you, sometimes.", without providing any details or how to build them or what tools to use. - saved your 13 minutes.
I love Andrew Ng but he has a tendency to speak with an “uptalk” inflection which makes him sound immature and lose his credibility and gravitas. Please Dr. Ng, uptalking is cringey unless you’re an insecure teenager.
its so soothing to hear andrew ng's voice. brings me back to my coursera ML and DL courses
starting from forecasting the house price🤣always in my mind.
The misery of Qwiklabs
Same, what a great presentation and such an inspiring talk.
The AI legend, I can't forget how easy was learning complext deep learning stuff by just taking Andrew Ng's courses
00:01:03 AI agents: Iterative, agentic workflows enhance performance.
00:04:29 Reflective agents: Self-assessment improves code quality iteratively.
00:09:22 Multi-agent collaboration: Diverse agents boost complex program generation.
00:12:01 Fast token generation: Rapid token output crucial for iterative workflows.
00:13:21 AGI journey: Agent workflows propel progress in AI development.
Prompt your AI to be more concise
They have a tendency for verbosity
I use this method of agentic workflow for coding in school, with brainstorming, and detailed pseudo steps, then review it, then stich somethings together, and have it revise itself. Then after a few cycles, I give it a request using words like "magnificent," or "swe professional point of view," and other things along those lines. Which ends up give me a better product by the final iteration. I am a Sophomore CS student, and I'm glad I am doing the method right, when I'm practicing data structures. Thanks for explaining the method of what I have been doing Andrew Ng.
Can u share your workflow
Do you have a blog, youtube channel or x account?
@@jatingupta4708 I would, but I feel I would need to give like a step process, that would have other bubbles next to bubbles for things to consider. If you give me a few days, Ill give you a generic image generated host paint of my workflow.
@@hl236 I only use UA-cam and disc for 99% of interactions. I have a git hub with student work I've done, but its nothing special, as its just college freshman and sophomore projects, keep building yall.
Excellent humble brag!
Hi guys, This is an excellent talk. I watched the Andrej one as well. And it seems that these talks are really focused and talk about real problems. So kudos to the team for executing on this. Its WAY WAY BETTER than these large conference interviews where you don't learn anything
They're also short and straight to the point! This would normally get stretched out to 30m
the anthropic one was not good imo, but the mistral one was good as well
@@thesadboxman true
@@cesarromerop ohh is it...i didnt checkout the anthropic one....will do
Not limited to this event, I wonder what other talks stand out to you. I am trying to learn as much as I can about this space from experts. Many thanks
Clip #2 [4:26-6:16]
I love how Dr Ng humbly describes his work as laying down another brick on the golden road to AGI.
Thank you Andrew Ng for certifying me 🙌
It's nice to see Andrew finally be impressed by something! A lot of his previous talks were along the lines of how AI can't do anything yet and it's not nearly capable, etc.
This is going to change development radically. Imagine being able to just write workflows to write tons of code based on what functionality we want while modulating interconnectedness, dependencies and other finer details and nuances which humans understand and A.I. does not. Just the best of human and ai combining to increase productivity. Excited for the future.
Absolutely. There will always be a human in charge, but insanely more productive. (We are helping LLM devs with that human intervention part)
Not as easy in enterprises.
Brilliant. The only thing I would add is that with inference speeds offered by Groq, it’s not necessary to wait minutes anymore. Fast inference speeds have the most value when humans aren’t reading the output.
Groq is on ZeroBot if you ever wanna share some thoughts 😉
Yup, I was wondering what the value is with Groq (like do we really need inference that fast) but agentic worfklows have provided a solid use-case
Groq is still bad
@@mayavi93 how come? I've never used it
It is always a pleasure to hear Andrew!! :)
I'm watching you from age 16, GOAT AI teacher!!!!
Insightful discussion on the future of AI agentic workflows with Andrew Ng. Excited to see what innovations are on the horizon. #AIAgenticWorkflows #AI #Innovation
Agree 100%. To reach AGI/ASI will require agentic workflows.
Great insights as always and I admire AndrewNG explain the advance tech explained in super simple manner !
I just did a quick test of this by customizing GPTs how I would like it to respond , its a good fame work , but i dont think its truly agentic .
"Define the Goal:
Be specific and measurable. What is the desired outcome of this workflow?
Identify Key Milestones:
Break down the overall goal into major checkpoints or phases.
Flexible Action Steps:
For each milestone, brainstorm a variety of potential actions that could move you forward.
Avoid a rigid, sequential task list - these actions are options to choose from.
Evaluation and Iteration:
Determine how you'll measure progress towards milestones and the overall goal.
Build in regular checkpoints for assessment.
Be prepared to adapt actions, milestones, and even the goal itself based on results.
Key Principles
Outcome Focus: Prioritize reaching your goal over following a precise plan.
Adaptability: Embrace change and be willing to pivot your actions as needed.
Empowerment: This framework aims to give you agency and decision-making power within the process."
Great talk - thanx for sharing!
Are the slides available for download?
Andrew Ng as always great!
Interesting talk on AI Agent, Andrew Ng. More updates on this subject will help. 🌞👏👏
I am 54 this year coding for 3 decades, been using prompt engineering to create my code for last 2 months using multiple GPTs and thought it was the best, but this agentic loop will make programmers legacy, being slow and expensive
In the next 6 months these workflows will improve, the rise of AI agents will be here whether we like it or not
I advice knowledge workers to prepare for this financially, physically, mentally and emotionally
This will be fast, we will be like deer looking at headlights
most knowledge worker are not even aware of AI at all. They will just wake up one day and their job is gone.
no pls..
Fascinating. I'll experiment with agentic workflows for my product. I really like the idea of getting more out of GPT3 using rag and these techniques.
Love learning with Andrew !
Absolutely outstanding talk! Outstanding!
Practical overview and feedback of what works and doesn't, very nice
I LOVE AI Agents... but I'm left wondering: Why hasn't anybody developed a system/app that takes the API's from the top LLMs, created agents for each, and then have these agents all work together to brainstorm, debate, review, and solve problems? I often get 4 different answers from 4 LLMs, so why not have them all setup as agents "in one room" working together to come up with the "best" solution. I can't find anybody that's tried this... why not? Wouldn't having the "top minds" (LLMs) working together produce better results?
I ask myself the same question. I'd like to see real case application bc it seens so unreliable yet
do it
@@christophscholz7484 already studying and trying
I think that the only think difficult is to make a good back-end for budget control because it would cost a lot on API and would be unavailable for people that are not willing to pay for this kind of website.
@@husnainarshed7806 that's a really great (and important) point. I think it could be mitigated by: 1) Having the smartest model be the "Manager" of many less-expensive "worker-bee" agents. 2) Setting a "max budget" for the query / project... e.g., "I need this task completed in x days and am willing to allocate $y budget. Do the best you can with the budget / time allocated!" Then the boss agent would be smart enough to budget (time and $$) efficiently. Say it has 10 days... it could use off-peak model times (I assume models will eventually start having time-of-use pricing tiers).
Mind-blowing... THANK YOU, Andrew! Great talk.
Very nice. Now I want to see how I can use agentic workflows.
Great talk! More videos like this please!
Guys! Make sure to check out the papers he lists at 11:07! It is required reading for the exam 🙂
Love the analogy of the proverbial manager checking every 5 mins after assigning a task 🙃
I am a big Fan Of Sir Andrew Ng❤
Very good. I’m currently playing around with the concept of the slow approach.
How do I get started?
great talk!
Currently, seems like "AI agents" are just a sequence of LLM requests. Take output and throw it back into LLM, with tools! Conditional loops seems to be the differentiating factor between these chains and real "agents" as I would define them. Any thoughts?
"Maybe you do, I can't do that" 2:39 😆😆
The AI's content recommendation engine helps brands discover new trends and topics, ensuring their social media presence remains fresh and relevant.
We think that this shift towards agentic workflow will lead to the need of rapidly and consistently train routers to quickly, and smartly route queries to squeeze to most performance of every of the agents.
This is exactly what we aim to do here at Plurally.
We are currently in an early access stage, but to anyone looking for such a solution or simply want to try it out - reach out to us (we are very responsive), would love to hear about your specific use cases!
Andrew is AI legend
agentic workflows will be come the best way to AGI
would be curious to get some opinion on running AI agent on the edge without using LLM on premise which we believe is the future.
The world has made AI loud enough but still its directions should be geared to gain the right progress.
Take a shot every time he says "sometimes it doesn't work, sometimes it's amazing" 😅
Great presenatation -)
in a way , conceptually, the critic and coder roles are some sort of MoE for LLMs....
3:08 - what's the difference between "Reflection" and "Reflexion"? English is not my first language.
Reflection is the correct word. Reflexion is a term created by the researchers, with a self-reflection methodology where the LLM reflects on the previous answers to improve it.
@@rafaelfigueroa2479 merci beaucoup
you can also reflect on something. Reflect on the past and reevaluate thighs etc
That’s gold!
Really enjoyed this
The core challenge here is that LLM fails at multi-step planning - and there is no way to guarantee that iterative Reflection bring correct solution (and optimal) in short period of time and money
To me this approach feels a bit like computer vision before neural networks took over. Hand coding feature detectors etc. That’s what you’re doing when you hand-design a workflow like “do web search, gather sources, write first draft, critique first draft” etc. These hardcoded agent flows are too rigid to produce good results generally. The models will learn to construct their own flows just like a person can. So while people might have some success building agents like this now, I think it’s a bit of a dead end that will be overtaken by foundational models.
Prompt instructions vs zero shot will always be a balancing act that depends on use case. Give too many instructions and you'll nerf the llm. Don't give instructions or use rag and you it will output responses that are low value and not actionable.
However the value of an llm will always depend on a user's ability to ask the right questions. If Elon musk gave you 30min of his time the value will depends on what you ask him.
What he's saying is AI agents of the future will barely resemble the LLMs we have now
They'll be more autonomous and not just chatbots we can prompt@@whoislewys3546
Great video
Thanks for sharing :)
Americans hate to admit it but we've been under such a closed system for so long to the point our public education stopped enriching young minds 40 yrs ago and flipped into recruiting for agency & institutions taking away the most productive years from our workforce.
It's been 90 years since we was in a very open private sector individual owners and operator creative posture.
Everyone is a cog in the wheel more or less.
We have so many grandfathered in economic middle men over the many phases of steam engine until today. Unlike most of the world that only individualized the past 50 years after the transitor age it leaves the west and America with many extra left over obstacles.
It also leaves debts paid up front that have helped us get to this technology.
Since I've retired and lived through all that 1900s, structuralism costs classical American decendants its only fair to remind everyone what these are from random Joe's perspective.
Family birthrates ,18-30 year Olds trained up and entering workforce at the most creative and productive ages ( which by default tends to marry & help Maintain the elusive American prosperity) and the lack or loss of 31 -50 year old owner operators of local American infrastructure.
Yes it's an unsustainable theme here that's been a very hefty price in building out our world over the 80 years of the transitor age.
If any sectors are handed advantages in this new paradigm infrastructure, these are the ones who have paid the ultimate cost in my lifetime.
New paradigm infrastructure where balance is there for better quality of life and at minimum restoration of all that's been compromised.
We have so antiquated ways of doing things . Our city's are still under top down rule prohibition era reformed control mechanisms
What's the difference between "Agent" and "Agentic workflow"? Will Agent include agentic workflow?
Great stuff.
The Legend!
How might agentic workflows transform industries beyond coding, like semiconductors?
none, other than further increasing the demand for semiconductor products in data centers
You could hook up an agent to Verisim and a 3D printer and tell it build better chips for itself
Very good talk. I was only distracted by the guy with the restless leg syndrome in the front row. He must have been even more distracting to Andrew, who despite that delivered an awesome talk!
That’s so funny that you get distracted by the guy. I hadn’t noticed it. Now I will when I re-watch this now lol 😂
Industries will be transformed by AI agentic workflows. These technologies are now powerful and accessible thanks to SmythOS. #FutureOfWork
i thought GPT-3.5 is dumb.
then Andrew pulled up chart agentic 3.5 beat GPT-4 .
Andrew should've been mayor :(
you want that instant gratif... euuhhh search result :D
Been doing AI agents operating in async workflows commercially for about 30 years. About time peeps catch up
"sometimes it works, sometimes it does not"
: whats your real name?
Pretty funny in 2024 how "the cutting edge of computing technology" and "writing scripts for cute AI NPCs" are more or less the same thing :)
Ng
10/10
别忘了投资我们
I will be happy if I get a B 😂
He still not explain why he gave b to that poor kid
He used an agent reflection to give him B. First pass the gent had given the stundent C-
What a lame introduction!
Jai hind
dont uae what you just heard
why
This was useless. No real content. Just guesstimations of very vague trends. Just say "Agent Models can wow you, sometimes.", without providing any details or how to build them or what tools to use. - saved your 13 minutes.
I love Andrew Ng but he has a tendency to speak with an “uptalk” inflection which makes him sound immature and lose his credibility and gravitas. Please Dr. Ng, uptalking is cringey unless you’re an insecure teenager.
Wait, who sounds like an "insecure teenager"? Yikes
What is up talking?
Deep.learning
Watching these videos doesn't mean i am joining you! Just for the knowledge sake!