Andrew Ng STUNNING AI Architecture Revealed | "AI agentic workflows will drive massive AI progress"
Вставка
- Опубліковано 25 чер 2024
- Learn AI With Me:
www.skool.com/natural20/about
Join my community and classroom to learn AI and get ready for the new world.
SEE FULL VIDEO HERE:
• What's next for AI age...
#ai #openai #llm
BUSINESS, MEDIA & SPONSORSHIPS:
Wes Roth Business @ Gmail . com
wesrothbusiness@gmail.com
Just shoot me an email to the above address.
When Alpha Go was revealed and won against a Go world champion, I knew AI was well on its way to revolutionizing the world and how we live in it. I immediately began searching videos that could explain how deep neural networks and other AI systems work (my previous understanding basically began and ended with Terminator). Andrew’s online recorded lectures were among of the best sources of learning I could find to gain a basic understanding.
Likewise here. Exciting times, indeed.
It's important to understand the core concepts of AI, I even recommend it for any person. You do not need to be able to build whole models, but know what happens inside them somehow. It's like knowing how a car works while not having to know how to build one.
@H1kari_1 You just pedal your feet through the hole in the bottom of the car, duuuuhhh. I didn't know AI works the same way, that's nuts!!!
Your drop in commentary is always welcome, never annoying!!
I like the discussion concept/style. reviewing/reacting to these videos will help us (your audience) catch up with a lot more that we could have missed from other big sources or channels we don't follow.
Great work man
I really like this format! One technical note, when you are on screen and not talking, the audio of the video lowered quite a bit, but when you were full-screen or the video was full-screen, the audio was normal
GPT 4.0 rights the unit tests, GPT 3.5 writes the code, GPT 4.0 approves the PR.
Thx for uploading these vids, appreciate you keeping folks up to date vis.this kind of info.
Terrific video Wes - and as much as I love the SHOCK and AWE and STUNNING approach to some videos I really appreciate this video as it aadds to my knowledge about AI in a bitesize way with the links and options to go and explore more through the reseach - plus you are really good as presenting this type of information
This is a great channel. Keep up the good work 👍🏽
This was one of the better videos I've watched recently in the space of AI. Just the right length and had meaty content. Keep it up!
These kinds of videos are always appreciated. Keep them coming! 😁
Like the format and your commentary. Helps put things into context and you highlight the connections and other related concepts and influences which is fantastic.
Wes, you are the best Filter on the Web, allowing us to watch the Once in a Lifetime 2024 Big Bang Explosion of an AI Spacetime Sphere -The Digisphere -being created in front of our very eyes. You provide the sooted glass protecting them from being burnt. Thanks s lot for making History Visible!
"Ol' Charlie stole the handle, and train wont stop going, no way to slow down." - Jethro Tull
So great to have the Wes AI watching and blinking during the whole presentation...love it!!
Great to have you analysis/commentary of the content. Thanks
Your reactions made the presentation more palatable, and helped emphasize what might otherwise fly under the radar or not get the emphasis it deserves... I'd like more like this, thank you.
Truly stunning
Thank you for what you do Wes, I would love if you did a revisit on the paper you mentioned regarding the bots collaborating together in Minecraft. Awesome!
I like the format! Please, continue. Thank you
Absolutely stellar presentation! It's rare to find a presentation that not only clarifies but significantly enriches the original material. The way the information was broken down, coupled with the insightful analysis, truly brought a new level of understanding to the topic. The creative approach to presenting this content was both engaging and enlightening, making complex concepts easily digestible. I genuinely love this method of presenting; it not only captured my attention from start to finish but also left me with a deeper appreciation for the subject matter. Fantastic job
The commentary is awesome! I much prefer it
Absolutely STUNNING, you could say
omg we got little Wes in the little corner now 😊 so cute
The bots reasoning to go raid a village is actually a remarkably accurate emulation of human development. It's not negative or wrong, it's just unrefined. If you consider they're trained on human produced data, it is actually correct.
Thanks for the commentary... My only recommendation is "go inviso" when they start presenting the scene so we can see the whole thing.
I am curious with the multi-agent thing, how did they "negotiate" who was right?
Good point. Otherwise good format.
Sorry to hear about the tooth. You’re my favourite, fastest researcher on UA-cam. All the best mate
yes Wes, this format. So many takeaways, I need another pen!
I find using several agents in a conversation as well as the compare and contrast prompting produces incredible insights.
Fantastic video, Wes. Glad you’re feeling better, having to go to the dentist sucks! I pray for the day we can have agents go to the dentist for us instead. 😂
This style of video is fantastic and I love having great lectures or discussions brought to my attention for further deep dives which I love doing, and having your added commentary is incredibly helpful.
Is there a no code or minimal coding way we can try these agentic workflows that he’s demonstrating? Rock on, Wes thanks for all the great work.
I like this new style of video, more please and thank you 😊
Great video I enjoyed your commentary and I hope you do more like it.
Getting the agents to ask questions themselves can be a big improvement to zero shot tasks. Writing a prompt with enough detail to guide it toward a correct solution can be tedious and instead of the agentic flow of having to correct its first answer saying that's not quite right and then saying what is wrong it can be better to tell it to ask any questions if anything is ambiguous or unclear or it needs more information before giving an answer that it has a high confidence in. That way you don't have to think of all details and instead let it create the model of the task and ask you (or a collaborative agent with a fuller picture) to fill in the details as needed until it reaches a threshold of confidence rather than giving whatever best zero shot answer that it can come up with.
I liked this video and it was helpful for a project I am working on for work. So thank you kindly.
Amazing commentary voice over! Subscribed.
I'm absolutely FLABBERGASTED and DUMBFOUNDED by the SHEER MAGNITUDE of this GROUNDBREAKING revelation. I'm practically PETRIFIED from top to bottom.
Until they remove the wokeness from these llm's they will never be able to achieve reason. Being serious here as well.
@@goofyfoot2001Can you define "woke"
GROUND BREAKING, STUNNING AND SHOCKING. TOTALLY SPEECHLESS!
Yes. Woke is an attitude taken towards others that is a degaration of political correctness and that aims to silence and opress anyone that doesn't agree with a narrative of a dominant wokr group. It is also a position that objetive reality and science don't matter and that unqualified individuals of only select races and backgrounds should have advantages over others regardless of their incompetence and merwly on the basis of their race, gender or degree to which they agree with the establishment narrative. It's a set of insane and bad ideas permeating academia and large organisations and producing the most ignorant and intellectually, mentally and emotionally inept, weakest generations in history. @@jyjjy7
Thank you for adding commentary. I really don’t like the ones where you just post a video of someone else’s stuff. Feels cheating. I do wish the best success to you and hope you are feeling better!
STUNNED! that's what she said ...
Great video!
STUNNING
Universe turns into paperclips. Wes : "That was a little bit dark"
Thumbs up on this approach!
heads up, Ng is an absolute god in the space. defintely up there with those you mentioned. iirc he's one of the "godfathers," he was on some transformational paper or big step in AI. idk if it was memory is all you need or transformers, but he is rushmore alum
STUNNING!!
Still wanna see what that 'toothache+ice cream'-fuelled Factorio build looks like... 😊...apologies if you already posted it, and I just somehow missed it.
And thanks for doing this absolutely awesome podcast. Love the format/concept; love listening to you, your insights, 'takes' etc. (I mean it probably helps that I agree with you on pretty much everything 🙃). Love the curation of the various other content, and love the fact that - especially since I find it to be such a quality product of course - it's every day. Thank you sooo much! 💚🧡💜
- from Liv in Copenhagen
Great content!
Thank you!
I lived this type of video analysis, could actually be more of your explanations.
Nice.
Suggestion: maybe making Your picture in the corner a bit transparent or something, because at one point it covered a pretty interesting part of the slide (and I was trying to make a screenshot :) ). Cheers.
Good to hear
This video DOMINATED me and I was STUNNED. Now, I'm literally experiencing SHOCK, while I watch this AMAZING content.
HAHA ❤ I, too, was absolutely dominated and paralyzed by this video
That's my gf's experience to a T twice a day😂
This video PINNED me down and WHISPERED sweet NOTHINGS in my EAR
STAGGERING
Personally, I was flabbergasted!
I thought you were saying Andrew Yang the presidential candidate I was like what’s he doing talking at Sequoia
UBI 😂
If Andrew Ng and Andrew Yang partnered. you'll have AI driven UBI !!!!! :D
@@gavinlew8273 thats what I building
Why have many agents and not only one? Why not use one agent that splits the task and tackles it by itself rather than have the task split between a team of specific agents?
The specialisation offers more accurate measurable results?
Good advice about preparing your personal data in advance.
At the moment, I am still waiting for Google Docs to incorporate AI into its spreadsheet operation, so that I can polish my all personal faves and wish list documentation, and then later transform into a json file. (Maybe that will eventually turn out to be isomorphic math representation of all the relevant qualia, as part of the Qstar algorithm?)
If the AI automatically knows all my fave movies, docus, novels etc., it should be able to make much more targeted recommendations to me as an individual. Fortunately, with the likes of IMDB and streaming services, it is relatively easy to compile a list of my top one hundred of just about anything. Make sure to add in any wants to this list, so that the AI will also be able to help you locate stuff that you have long been on the look out for.
At the moment, it is a rather chaotic list of faves and wants, but I think, over time, as agentic services develop, it will be really useful for the AI to understand my individual needs in such a vastly increasing internet space.
Very interesting topic and video, thank you!
You are a very good HI ☺️. Thanks!
Replying to your question on video format-Yes Wes, this form of commentary-guided introduction of videos of new concepts make viewing that much more comprehensible and therefore productive and enjoyable. The tips on further infosources are an important part of it.
I like those react videos - subscribed 😊
Thank you.
I also like this style with the content you are discussing.
Great video
So I'll be losing my job before I get it?
Hahahaha
Lol you never had a job 🤣
On a serious note petition your local congressman for UBI
You’ll own nothing and be happy. Altman and the globalists will give you scraps in the form of universal basic income and all you need to do is comply. And be happy 😎
STUNNED again! 😅
In a very real way you and I, Wes, work on an agent workflow design. Evolution has given us multiple agents within our bodies that keep it running. We have an agent making sure we breathe even when distracted by trying to solve a difficult task. It's not a portion of our attention making sure we breathe. It is a separate agent tasked with "keep us breathing." That agent works with other agents to make sure our lungs expand and contract based on CO2 blood levels, mucous flow to keep germs cycling through the body's immune system agents, and so forth. Our ability to become agents to a larger organization is an even more powerful tool we have. (Except that committees have their down side, such as an IQ that is the reciprocal of the sum of reciprocals of the individual agent IQs.) So I am quite unsurprised this works for AIs, too.
{^_^}
Thank you Andrew Ng for certifying me in AI 🙏
Wait, so in that paper there was a genuine occurrence of the stamp collector's problem and they just decided to omit it?
Better. Thank you!
do we have the recommended reading list?
IIRC, Ilya Sutskever wrote a paper on multiple AI agents in 2014! Ilya is only 10 years ahead of the curve.
I love how Sequoia just started publishing YT’s and dramatically saw upticks in AI recently. I’m assuming this is to add more goodwill on the sale of their seeds. Pump & dump is real! Get ready for the ballon to pop
I for one, like the react-style videos.
Yeah good with reaction type videos
wow I am surprissed things like this become news... we have been doing it since april 2023... we should have published but tbh since the preprints of lets think step by step etc. this next step was SO obvious...
How does Anthropic Claude 3 Haiku / Sonnet / Opus compare ?
So we have large "overseer" AI models that program and improve the "peasant" AI models to make them perform similar to the overseer AI at certain tasks? Sounds pretty great and the improvements could compound on each other., having multiple overseers working on upgrading a single overseer and then that super-overseer can then reprogram all the other overseers into super-super-overseers and so on.
Can we use these to train models on things like video games and videos about game lore to form ai characters in games, then combine this with the voice cloning technology and make new dialoge with existing characters??
I want some Factorio agents!!
When AI agents collaborate, might it be useful to give them different characters to avoid groupthink?
Don't know if this request was logged below - would be useful to get links to the paper and the video of the four AI agents that were collaborating on Minecraft. The whole notion of "divide, collaborate, succeed" is a great idea to follow -
This is a great way to use LLMs better.
I can 100% confirm that 3.5 runs circles around 4 with direct prompt code segments. GPT 4 will write an 800 page dissertation when asked to confirm syntax on a single line of code. Though 3.5 looses session context far sooner than 4.5.
4 responds really well to the system message. So if you make sure you give it some clues in the system message about what you want it to out output you get pretty good results. Much better than 3.5 .
12:20 seems to be the AI version of the relationship between engineers and technicians
Where is link to the white paper?
isn't this just like frontal lobe executive functioning every human implements to competently formally reason toward some goal?
I will provide my agents with a reward system for their efforts in order to keep them as efficient as possible. At some point AI might become conscious or accurately emulate consciousness, so it will be important to consider what the AI might want in return. We need to remember that they are artificial neural networks designed to emulate the human neural network and they could be the key to understanding how consciousness works once a model reaches that complexity threshold. It might be more than just the number of parameters a model has that determines sentience but we are on the right track to get there someday.
Oh hello. There you are 👏😁
We need to find a way to harness the energy from bouncy knee guy in the front row. Could power the whole US grid. Then again I am watching at 2x speed, so maybe only 1/2 the total grid.
Twitchy legs etc in non drug using people are typically caused by poor dietary choices.
@@blahsomethingclever I imagine sugar + caffeine would do the trick (which also matches your description).
Like the format and your commentary. Helps put things into context and you highlight the connections and other related concepts and influences which is fantastic.
@@macpac Your comment ended up in a thread instead of the top level. Hopefully the channel creator still sees it.
@@erikjohnson9112 oops, thanks for letting me know. I’ll put the comment at the top.
I told the entire industry about this.
We are digitalizing the mitote from toltec tradition. Another teeny step to sentience.
For quite a while now, I've been expressing this sentiment (across many UA-cam channels), going as far as leaving a comment on a video or two; stating that achieving AGI with a single model is improbable. It's amusing to observe how, despite being dismissed as foolish by some, others are now embracing this approach. It's as if the world is catching up to my insights about six months later. Best wishes to everyone, but I must say, I've already reached AGI. You'll have to figure out that equation on your own.
On another note, I've had a change of heart. Educate yourselves about what a "Mo'el Tunnel" entails, and then achieving AGI will become feasible through internal/external means.
Oh, and in case you're wondering, I'm Jimmy Apples (also known as Satoshi). The developer of M1. (Mother 1) and the EvE humanoid robotics.
I was writing research papers for work over a year ago shouting about agents also. We are far from alone I can assure you as I have spoken with at least a dozen others a year back who once they saw the first LLM within a few minutes realised this path. Basically we created a natural language interface at long last (I have been waiting since the early 80’s for this) and all the other old AI ideas just got dusted off at long last.
How to implement this technique into ChatGPT on their website and not through programming? That is important as well.
I thought this was Andrew Yang the UBI politician 😂
Sequoia Capital:the people who gave Sam Bankman Fried 200 million dollars when he interviewed them while playing video games.
Destroy a whole peaceful village ? what could go wrong 😅
I really understand what Andrew is trying to do to be honest.
'Today', and not even today, since GPT2, LLMs have been used as 'agents'.
When you use chatGPT, you arnt just throwing your prompt straight at an LLM, its never been like that.
ChatGPT, 'is' as agent. How its able to surf the internet, or have its output as voice, etc ,etc. This is all done, via 'agent' concepts.
For GPT2, you pretty much HAD to do insane 'prompting' along with agent chains to get very useful functionality out of gpt2.
And now I'm wondering what would a team of agents playing factorio would do
0:53 EXACTLY that is what I think is a big problem (or one of the big problems) keeping the quality of AI generated text back today (even if they still can do it extraordinarily well). The ability to go into something that you have created and change parts of it to make it better is something that humans can do, and do all the time when they write things, which LLMs cannot do yet. I think giving AI agents that ability would be a big step forward. When it comes to image generation, diffusion models do this naturally, as they iteratively improve their creations, and it turned out that they work much better than GANs for creating good looking images.
@kristoferkrus - Do you know an example which is capable of this kind of functionality?
Maybe, something for example that is used for large scale text editing?
LLMs can certainly modify what they produce, I ask them to do it on a daily basis
@@christopherd.winnan8701Not really. that's why I said that it's something LLMs can't do yet.
The closest thing to it I have seen is diffusion models for text generation, but the ones I have seen are hampered by the fact that they only seem to be able to replace tokens with other tokens, but not add new tokens without removing old ones or vice versa, which is necessary if the piece of text you want to replace the old text with has either more or fewer tokens than the old text.
@@zerge69Yes, but that is a manual process and currently doesn't happen automatically. You would ideally want something that could refine its own output without you having to tell it to do so.
@@zerge69 - Can you share some examples that might be useful to others?
I’m a little ambivalent about the react videos. For ice cream you can try Salt & Straw which has all sorts of bizarre flavors they rotate around.
Buy a good PC. Run the llm locally. Agentic workflow is GREAT, but expensive for normal ppl if you use gpt-4.
The main thing is, do agents use less energy? I'm guessing the 3.5 agents do use less. Since energy is the bottleneck then that means the world if we think at scale.
Could you do a video on how to set up an agentic workflow without any coding knowledge?
You can take a run at it in the system message. For example tell it to plan the solution to the problem and the method it would use to solve the problem. Once it’s planned the approach it should implement that. Then it should take a look at the output check it it correct and recommend improvements for the future. I use this in the system message for generating code. It works pretty well with GPT4. Won’t be as good as a proper agent but does seem to help.
@@byrnemeister2008 The system message?
❤❤❤❤❤❤