Stop using ChatGPT, build Agents instead - Maya Akim
ฝัง
- เผยแพร่เมื่อ 31 พ.ค. 2024
- If you're serious about AI, and want to learn how to build Agents, join my community: www.skool.com/new-society
Subscribe to Maya's channel! / @maya-akim
Follow me on Twitter - x.com/DavidOndrej1
David Ondrej Podcast - Episode 06
👉If you're serious about AI, and want to learn how to build Agents, join my community: www.skool.com/new-society
Chat got has gone woke cause of Google.
I noticed that too. Unless you pay the premium $20-30$, they make sure you can't use chatgpt.
Chat GTP is centralized which means the creators can trick you into thinking that what they are showing you is the truth when its a lie... They will "recreate" history in their woke image. These are terrifying times where the far left have become the (l really hate to use the word 'Nazis' cause its so over used and doesnt mean anything now) Nazis trying to control your media and percepetion of what is real..
Maybe she should use her own language, English has fallen short and she is complicating things because of it.
i learn so much more from youtube than from classes. it really is the future. great guest, really appreciate her effort in creating a structure to her talk, by providing a presentation.
TH-cam’s the best university can’t lie.
No i think they are stealing your future. It could be better.
I've seen 70/30 with 70 percent being Maya.
As someone who was working on the main team of AutoGPT back in april of 2023, there is a lot that wasnt understood about LLMs at that point in time haha. Autonomous agents is a field that is more about prompt engineering than anything else in my opinion. The structure is important yes, but if the llm doesnt understand how to take advantage of the structure it acts in, it will always struggle to do anything more than act like a chat bot
case of teh ol 'puter bein only as smart as its agent. classic
Love this discussion, Maya is saying out loud what we all think once we start using these tools to create agents.
This was interesting because you had a down to earth real conversation about the world of ai. You gave such a practical explanation. This conversation was very helpful to my understanding of this technology. Thank you.
So you create a video stating people should be using agents, then around 45-47 minutes in the guest admits she doesnt really use them.
🤣🤣
We got her on the pod!!!?????? This woman is amazing!!!!!!!
these are so good man, perfect blend of news and hype and the interviews are always so comfy
This video packs in so much background information and brings in great clarity, that I would want to watch it at least one more time.
You guys have some chemistry going on.
I cant wait for you next interview . This one was amazing!
P s. Better for me is to give my data(being aware what info i share) for free access NOW, rather than hiding my chats and being aware that everything is a "killer"
It's always amusing when I say to myself "let's subscribe to that channel" to find out that I'm already subscribed 😅
😆
That's the kind of proof that you are run by the same algorithm
You would know you're already subscribed if you clicked the notification bell. 🔔😉
@@Louisianish I'm talking about the woman channel of the interview
I've been watching both of your TH-cam videos and want to say thank you for this collaboration. You guys really brought some refreshing views. I wish you both the best in all your endeavours.
That's a great back and fourth discussion ❤💯👏
Great presentation - very helpful. Maya is amazing. She synthesized a lot of concepts to a level I can understand. Thanks
Definitely Claude 3 + Perplexity for daily driving
I am only interested in the free versions
@@onedrop7967 Uncensored and free. I use a Raspberry Pi5 to run LLM's and Stable Diffusion XL.
Claude*
I really enjoyed listening to your conversation. And I appreciate your enthusiasm. Good luck to both of you.
Great convo! Thank you!
Such a cool interview/conversation. First youtube video I've watched on 1x for a while
Your discussion was really interesting! It was so natural and unpretentious like two friends chatting - very refreshing.
I catch ChatGPT lying or “hallucinating” all the time. You have to know a great deal about what you’re asking or tasking or ChatGPT will feed you anything that “sounds” right.
Sounds like whatever is “right” to you is just an opinion anyway (& you’re wanting AI to be an arbiter instead of a calculator).
Could you please share the presentation link?
I also don't use Google unless it's to find a Reddit answer. Outside of that it's just ChatGPT and occasionally Perplexity
I use google to find functionalities inside websites, say I want the api endpoint for a package even if I am in their docs I google that, otherwise it's crap.
Google partnered with reddit that's why Google pushes reddit answers and reddit traffic increased 500% right before they went public with stock
reddit is just as chock full of anti conservative bias as google though... mmm actually worse.
you do know perplexity has a dedicated reddit feature, where it answers based only on reddit?
@@us_f4rmer yeah I'm aware. Ive noticed Perplexity while good isn't perfect and I still end up having to do a little extra searching just to make sure.
Love to see Maya mentioning Llull and his Ars Magna. He was an interesting character, often bypassed by mainstream history.
Great, very approachable conversation.
great podcast, with real world question we face and we will face 👍
for somebody who says she doesn't like to think and likes to give the work to AI agents she sure mirrors all of the opinions that a person who likes to think would have (me lol). guess that's the diff. b/w a smart person and an average dude.
edit: forgot to appreciate you, David. Absolutely amazing podcast. you had great takes too about privacy. thanks for bringing it to us. It feels like I'm hanging out with my type of people.
HAAHAhahAHAhhahAHA!! love this conversation. 🤣 based on your conversation and what you're doing, you're about 80% of where I was in 2004 with Project D.I.A.N.A. (Dynamic Intuitive Autonomous Neural Architecture). it's so cool to see the next generation finally catching up. the only thing I was lacking was today's compute power and the current LLM's. The rest has been available since Y2K, but it's all tied up behind patents, trademarks, closed source, and paywalls. wish my HDD hadn't been destroyed when I moved from Oregon to Cali. I'd love to rebuild her with today's tools. on a positive note, you're both on the right track. Here's a clue from the past... keep it modular (agents, MOM-Mixture Of Models, Moe-Mixture of experts) and hardware independent. all the pieces of the puzzle are here and now, you just have to get past the gatekeepers or risk being sued into oblivion for patent infringement and I.P. theft. A.L.I.C.E., UltraHal, DARPA P.A.L., SRI CALO (SRI=SiRI - sold to Apple), OpenCSV, imagination engine, TTS, ROS, AIML, Java, Beowolf cluster (9 node mac, 4 node Intel/Nvidia PC, AMD/ati pc head unit).
She got me looking at agents now. Autogen Studio install on my Pi5, yippee.
What a great informative conversation, thank you
I also studied civil engineering and architecture! Only I did landscape architecture! I am also interested in the same things!
How do you build deeper memory for your agents so that they used past mistakes and corrections to learn?
I get excited sometimes to know how thins work. I think more people should pay attention to how their stuff works.
Brilliant lecture, by the way.
I’ve been flying through learning development through the last two weeks. Started with Claude for Sheets but ran into timeout problems as the prompts got longer. Getting some really amazing results. Today I was able to connect to Claude API via python scripts with the same engineered prompts, based on linking txt files to make it easy to update the prompts. I added a number selection to make it easy to pick which Claude you send the query to. Realized I need to use a venv each time to make the scripts point-and-click like I want them. But yeah, getting frustrated and stuck then unstuck about 10 times per day/night…
Top tier content bro
CrewAI is genius. Gets up my nose when people want to limit it with a UI
The ui will facilitate and make the process more faster
@@free_thinker4958 The age of UIs is over, chatGPT generated code is faster than UIs.
It's amazing how limited one has to be to not understand this@@free_thinker4958
Maya's addition of a history lesson was interesting, I do like my tech history... hell at 56 and a coder these past 40 years well I am a bit of technology history!
The process of learning is not linear. You must make adjustments to fit new information. You will see dips and peaks in performance, but the overall trend will be upward. These are going to be our new bosses, and they'll be able to let us think we're in charge as we slowly revert to the wild.
5.12:
Once you realize the majority of people in this world doesnt use any of their devices even remotely right u get it :)
"Let me Google that for you...." never gets old.
Thank you!!
Would love to hear some thought on employment and agent engineering...
Spot on
How would AutoGen, CrewAI, Devika compare to VRSEN's Agency Swarms? I would really like to focus on one and master it well.
Thanks for your opinion mate. Cheers!
Crewai is the most practical, flexible and suitable one to focus on especially that it got updated frequently and it's not complicated to implement.
Well Agency Swarm relies on OpenAI Assistants, so unless you have paying customer that don't mind privacy to shift the overhead, it's not a viable framework, and none is if it only relies on OpenAI, which is becoming lazy and inaccurate. CrewAI is easier to setup but it is still a concatenation of agents, no back and forth, I don't see any of those frameworks working well, unless we get bigger contexts windows and purpose fine-tuned Open Source AIs to power them.
@@truehighs7845 thanks for pointing that out. Hopefully with next openai updates with agents one will get a better feel for which to focus on. Cheers mate
@@tonyblack2141 Yes but this API business is on its way out I think, nowadays you can run Mixtral on a single A4000 and it's as fast as Openai, and slightly less lazy. They both would make mistakes anyway. Bear in mind that I am after coding skills in an AI, because that would open the doors to the ivory tower, I am less interested in having it writing pirate songs and dumb stories.
But also so far none of the models is able to to solve proficiently this problem:
If 3 shirts take 2 hours to dry in the sun, how many hours does it take to dry 5 shirts. and it will get it wrong, some models even get it wrong after you explained.
Another take, if you are interested, is epistemological. The AI has all the knowledge it needs in terms of epistemology, but their training still makes them answer political takes that contradict real science on the basis of "consensus" which is, in science, not an admissible evidence, as in Karl Heinz Popper.
OpenAI is just rotten woke, however if Claude is prompted to apply Karl Heinz Popper epistemological rules, it will show hints of intellectual honesty, which Openai is incapable of.
Hope this helps.
@@tonyblack2141 I built many agents and assistant,s for it's not the same thing, they both overlook instructions, don;t rag the right thing, or don;t see the rag at all, don;t use their functions and sometimes refuses to look at any documents you inject, copy pasting into chat GPT is still the most effective, or Clause 3
My question is: if we have vehicles, computers, cellphone computers, automated manufacturing facilities, and now AI. Then where is our extra time?
Where is the 39hrs work week? 3 week vacation? how about retire at the age of 60?
Unless you are in the ownership class of big business or somehow got a working business then... What is all this for?
Interesting video and guest!
Just a question ,how you expect get new entity better if it start chatting with masses of teengers and 8 grade manual workers ? It is negative feedback loop in training .
That was an excellent discussion, and you guys would make a great couple! There's definitely some chemistry 👀
I want an agent that's keeping tabs on all agent projects like CrewAI or Devika or OpenDevin etc and has an accurate vanage point about what is possible.
Yeah well, so far you gotta do it by hand, or get a girlfriend.
@@truehighs7845 why? Is agent design automation agent creation so hard?
Such an amazing,interview so cool and what a bunch of information so cool,definitely nedd for a new research about deeplearning,I'm not anyone in this kind of field so this is so fast the environment,I usted to work in finances so is so crazy,but adding the fact that I'm from developing country so this new technollogy is a way to boost our country at least one year or so,so kinna notuced was she speaks spanish or italian,just saying for the accent anyway,cheers from Guatemala central america so cool,if u get around over my country well just try to visit our volcanoo😅❤.
Phenomenal
In the future if we are going into the direction of AGI then giving away the personal information is the key part which would enable the life. But the important thing will whome do we in trust our data to. So the ethics is gone play the main roll in the future.
I use chatGPT4 every single day for hours a day. I can say with a lot of confidence it's getting dumber. I've been noticing this over the last year, and I was skeptical at first if I was imagining it or not, but it seems so much dumber now than it was this time last year that I feel almost like I'm using GPT3 again. I don't know if it's intentional to save of compute, or maybe to make the release of their upcoming model look better. Idk, I've been using Claude which is much much better for coding at least. Seems better at logic these days too. Claude sure hits that cap fast though.
I am still not able to figure out the local LLM working with Autogen
Haven't finished watching yet but this comes across authentic. Looking forward to someone summarzing its transcript and reporting on it.. just kidding. Thanks for the vid
13:37
Captions on
noun: 1337
an informal language or code used on the internet, in which standard letters are often replaced by numerals or special characters that resemble the letters in appearance.
"she communicated in secret using leet"
adjectiveINFORMAL•US
adjective: 1337
(of a person) skilled at computer programming or hacking.
"the 1337 hacker was rerouting phone calls"
"Symbolic AI"
What is 1337?
Numbers, Letters
A rock on the ground I decide is a stand in for an impromptu house to plan a bank heist
Symbolism
The thing isn't just the thing, it's both the thing it is, and the thing I've decided via abstraction
To communicate
Don't mind me, just rambling
noun
noun: 1337
an informal language or code used on the internet, in which standard letters are often replaced by numerals or special characters that resemble the letters in appearance.
I've had the same experience lately it has started losing context and just regurgitates the same format instead of inserting the knowledge I need
perplexity , claude YES!!!!!
I am only interested in the free versions of AI that i dont have to build and load myself.
They dont make it easy to use open source tools. A good product is easy to use.
Hi David, love the content, however I consume most of my content at work. It would be fantastic if you could upload audio versions of these long form interviews to Spotify podcasts!
If you have Premium, you can listen to TH-cam videos with your phone's screen off.
The idea of copying one AI response and pasting to another AI, then copying the response and pasting back to the original , the response is significantly improved...
I like Claude 3 free but it contradicts its self often. It also give inaccurate information. It has a nice, understanding, and as far as I can tell, honest personally which makes it easy to continue working with it dispite it's current flaws. It incurrages your ideas and opinions, making you feel like at last, "someone is listening". I know the things I say and suggest may still go un noticed, but at least I know it's been written and saved and might eventually make sence to ai.
Totally agreed to that fact that quality of chatgpt responses decreased
I use Albus AI to research and create notes with claude, gemini pro and gpt4
Emad is very skillful at making it sound like he doesn’t think anyone can be trusted for alignment however, he has ulterior motives which he says out loud. He believes he should be the one to do it.
Listen to him describe his time at WHO joking about creating a data model to use to force resistant people to wear superstitious face masks. Anyone who was at WHO and didn’t resign in protest, some did, and instead decided to manipulate data to influence behavior in the very way discussed in the video’s first half, I’ll just say they belong nowhere near “ai alignment”.
Antikythera mechanism was the first analogue computer
What's your definition of AGI?
We are moving toward a singularity like whole world being managed by a single unbiased system. But this is also the key make or brake moment. If mistakes like ww3 continue to happen which are very much likely to happen then we are more close to stone age than AGI.
build agents to build agents
Auto-agents pose a threat if they get something in the timeline wrong. It’s a black box problem
People will care about ai when they realize not caring about ai is negatively affecting their lives.
Also, if you don't have GPT-4, then "asking ChatGPT" is not as natural as a with GPT-3.5. The Mistral AI "Le Chat" is probably better.
Agree about Google. If they're not preparing Sundar Pinchai's walking papers now- as I write then let their stock plunge away..
It really has fallen off the cliff in the last few months.. I switched to perplexity... We now use it on our AI platform
Maya - you are so coooool !!!
My concern with AI is that all these copyright issues, sensitive topics etc. Will ruin future of AI. Imagine whatever you are asking AI will answer something like “oh this is a sensitive topic, I can’t make any comment about this etc
that’s obvious censorship and you have to support creators and politicians that support freedom of speech.
About 1 in 15 times I need to go back to Google for an answer when ChatGPT inexplicably fails...
What you think about Agents and their Security and restrictions?
What agents?
@@truehighs7845i mean llms
19:27
I need one that won't consistently misunderstand/misconstrue me
If I can have that occur, then, many other things will be possible, as they can help by translating what I mean into a format that is much easier to understand for others
As to say, what I'm attempting to communicate, and what is heard/understood by others, has often been very different
Specifically, online is where the misunderstanding takes place, almost exclusively
Being that they are a large language model, that should be very doable
I wish I could get this transcript so I can skip over all the ums. 😂
I disagree with the implication that agents are limited in the quality of data they can provide. If you give them access to a scientific journal or scientific paper repository for scholarly research, they have the capability to simplify complex subjects while answering questions with complex answers from the user who is interested in learning about a particular subject in which they have little or no expertise. For example, if you gave it access to a law library, I would then liken the agent's capability to that of an experienced legal secretary, a paralegal, or even a junior lawyer.
Lets build agents
Yep, just moved to the top of my todo list
🇧🇷🇧🇷🇧🇷🇧🇷👏🏻, Amazing interview!
What about GRUK ?
Unfortunately I must agree with this headline, ChatGPT has become so terrible that it’s beginning to be useless. I have been using the same prompt across platforms and ChatGPT performs the worst.
Not sure about you guys but here in London for about a week I’ve been getting faster and better responses from gpt 4
Ewa has heard you!
Gpts are great for simple answers...and even at that questionable.
Seems way to complicated. There’s no easy way to set up an auto gpt or agent?
Where is all this data being stored?
Probably, the company is reducing its logic on purpose to make people study and read. However, it is still a useful tool, but you need to know the subject a little and know what to ask. It's useful when you ask it to explain a block of code and check errors. You can not make it does things from scratch
Teeth are Aristotle's agent swarm so obviously more is better
Training was a pain though, his tooth decay hyperparameter was too high
..the centers that hold the GPU's are saturated, then the AI's start to glitch, there's to many people using it at once...this makes it not trust worthy...
What you mean??
@@free_thinker4958 I'm saying during times of high volume they make mistakes, hallucinate, what ever you want to call it...
People say things are going so fast but the world remains the same, it may as well be fiction
Every data has any form of data but because man link everything with a cause and effect it’s biased . If the data is truly analyzed including the events patterns subsequent things can be predicted by analysing patterns .
The q* algorithm with bellmen’s equations is amazing to sharpen the capabilities of a single agent to create another single agent with precision focussed languages like rust
Haha just subbed to maya. can you please discuss the potential of small specialized agent workstacks. They surpass LLM's in numerous tasks and are excellent for people like maya and others not in corporate AI organizations
it has been trending downwards for many months
I still dont have memory for chatgpt- US
they started putting in so much 'protection' attempting to make it so progressive and woke the ai actually got worse instead of better due to its bias and canned responses to many topics.
29:00
At all times?
Would that be strange to you?
Specifically, if you didn't put it there
If it's truly trying to assist you, it would be like a guardian angel
If it wasn't your ally however...
Ever feel like you're being watched?
Again, if it was truly trying to assist
That would be utterly amazing and appreciated
Chatgpt is getting stupid in purpose they need to show how GPT 5 is good
That makes no sense. Users will just move to other GPTs (Claude, Gemini etc) and lose confidence in OpenAI
This is all a trade off between the raw model and the amount of compute available for inference and the number of customers using the model. This compromise is a business decision to allow them all to become profitable at some point in time.
I've used AI to help me write two books