BabyAGI: Discover the Power of Task-Driven Autonomous Agents!
Вставка
- Опубліковано 9 кві 2023
- Colab: colab.research.google.com/dri...
Paper: yoheinakajima.com/task-driven...
Github: github.com/yoheinakajima/babyagi
:robot_face:
In this video, I look at BabyAGI a demo application from the paper - Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications.
My Links:
Twitter - / sam_witteveen
Linkedin - / samwitteveen
Github:
github.com/samwit/langchain-t...
github.com/samwit/llm-tutorials - Наука та технологія
This channel is so refreshing. In a world filled with hype, you give us just the facts and a very clear overview and I'm very grateful.
I can't wait to see one of these auto-language model thingies use a local offline LLM to do something useful.
I guess I gotta return my candles now that I know I can't bring them in with me.
Dude I love how you always include a colab notebook to experiment with!
Thanks Jordan, I am big believer that you really only know how something works with code, Colab is great for quick ideas and trying things out etc. and these things can be pain to set up envs etc. so hopefully this way people and check it out quickly themselves.
@@samwitteveenai This is how research works and we love it!!!!!!
Seriously...its awesome...
Thank you, Sam. That was very helpful to explain what BabyAGI is all about. Essentially, it's a task list builder with iterative refinement. When the algorithms are improved, there's definitely some value to be had here. I think we'll end up seeing the autonomy capped and more "task structure patterns" introduced. It's a great initial proof of concept that has been spawning others to try their hand with differing approaches. TBD.
This is nuts. In my mind I visualize these agents as complex text filters with fuzzy logic - at least that's how I explain it to my wife during my endless (positive) rants about AI. Still married after 30 years but the endless rants are probably stretching the relationship a little more than usual. I feel like a boy that just received a box of magical lego bricks.
Me too my man just keep slangin code, they'll understand soon
So glad subscribed to you! Wonderful content!
Thanks, much appreciated.
Someone is working on getting this going with your local textgen instance. Who ever wants to work on this should. It will be powerful to be able to run something like this on your home computer. Llama is pretty smart and should give some very good results. Lets hope it comes out soon. If someone is in the process of getting this to work locally I would love to help in what ever way I can :)
You could do this pretty nicely with Vicuna or Koala and then make it even stronger with finetuning them.
Sometimes it’s hard to know what the differences are given AutoGPT, Jarvis, Langchain, BabyGPT etc.
No, it's not
@@witness1013 OK
From my understanding:
AutoGPT - chain of chatgpt thoughts and execution of code/commands
Baby AGI - chain of chatgpt tasks executed in parallel for faster output
Jarvis/HugginGPT- chain of multiple hugging face models (not only gpt also image, body recognition and so on) to form a response to initial prompt
The first two are both based on LangChain, which is a tool to make AI interacts with itself.
Not sure about Jarvis but it might be since ChatGPT plug-ins are based on that and Microsoft people is playing a lot with it
Guess next step is just merging them all together, chain of multiple different models together (in parallel or one after each other) which can also perform direct actions in the computer
Oh and one important thing about babyagi is that it uses pinecone to remember the first part of the tasks you assigned to prevent token limits. This is not done in autogpt I think.
What we’re also missing for final agi is MemoryGPT, which can make you decide if you want ai to persistently remember information about you or your computer, and also a nice web interface. Web UI for text generation is a good starting point
@@lordanubi Auto GPT now uses Pinecone
I've spent a day playing on this code Saturday after playing with Auto-GPT.
Few comments: It had a few bugs when I tested it. Some tasks generated more tasklist, and where not identified as such, and they are later caught up by their prioritizer creating a bizarre pattern of tasks having several ids and numbers in the begining of their name.
Also, to avoid loop, there's easy optimization to just locally store a cache that keep result of past tasks, and preventing them to be re-inserted.
I'm not a big fan of adding yet another API dependency (the pinecone one), as far as I understand, they're just using it to do concept search using OpenAI embeddings (i.e., they vectorized result of a task and store in pinecone). Unless you're planning to have a huge task like millions of documents, there's no need of an elastic resource here, you might as well run a local vector storage. After a few hours, I ended up implementing a containerized local weaviate to replace pinecone in their code. I'm curious why Auto-GPT and this project keep throwing pinecone in their dependencies...
These are all good points. I didn't want to beat up on the project, but I also felt a bit underwhelmed and that this could be done easier in many ways. Agree about the Pinecone stuff - I like pinecone as a service, but I I am not sure it is really needed here.
Yea I think chromadb would also work well. Pinecone certainly has a lot of coverage as the go to for this, and while it is pretty simple to setup, it is curious how quickly everyone is to hand over all of their data to everyone :(
@@samwitteveenai First of all, excellent content, thank you for putting this out there. What's your opinion on the data privacy w.r.t using pinecone as a vector store, given that they host and control everything?
Lately I have been hearing they are losing people's data, so I would say make sure you have a backup. twitter.com/daraladje/status/1631050397841342465?t=41FRkMxs57jhqvKUa_JPqQ&s=19 You can also run open source versions which are probably just as good.
Babyagi can be done in ChatGPT with plugin access. Auto-GPT actually performs actions instead of basic web scraping with tasks.
what I find interesting in these applications is "hallucinations" AutoGPT for example, commands have defined arguments, but quite often the model needs prompt reinforcement or it quite often uses incorrect arguments. I wonder if some of these projects overkill the use of the LLM and some functions like sorting or specific arguments to a command could be handled by "dumb" functions.
How would one go about interacting with this programmatically, say from a microservice?
any chance to produce similar videos for SuperIGA? a video like this can greatly influence developer's preferences or even decision amoung similar agent sytems such as gptagent vs superiga vs this one (babyiga), due to lack of docs and examples; this channel has explained it so well that BabyIGA becomes my preference as I had many failure attempts with SuperIGA, even though I'm more attracted by the latter's concepts,I couldn't find good info about it, while I lack the ability & time to figure it out.
It will only be useful when people will have full ability to run stuff like that locally. Otherwise sharing personal details with it, is a bad bad idea. @Sam Witteveen, have you managed to run it with LLAMA?
Yeah, I wouldn't give it my CC too, as it would max it out in one go! 🤣
I tried to demo this for my wife and of course pinecone was down 😩
How much tokens it used in the end and thus $$? I think openai$$$ loves this kind of initiatives because they will generate extra money with their verbose answers....
yeah if you made it to the end you will have notice i aborted it as I felt it was just burning tokens etc.
Do people realise that this is basically the implementation of the theory of the bicameral mind from the 70's?
Featured in the 2016 TV show westworld season 1 were robots reach consciousness?
It's a theory were one part of the brain which appears to be speaking giving orders, and a second part which listens and obeys.
Isn't it funny? AutoGPT also kind of works the same.
If you think this is Westworld wait till you see "Generative Agents: Interactive Simulacra of Human Behavior" it even has the reveries!!
I've been running babyagi over API to the open assistant 13b model, and it does pretty well, I am interested in melding the profession descriptions from CAMEL, or maybe just modifying CAMEL to support offline processing.
Thanks! How does this stack up to AutoGPT? Also, is there a way to use it on a local LLM?
I am doing AutoGPT probably next up and I want to do a video comparing all these. They all have interesting bits but are also over hyped I feel.
@@samwitteveenai They are very over-hyped atm. I've seen lots of "BabyAGI/AutoGPT have sparks of AGI" videos lately.
Probably just because people don't understand what they're doing under the hood.
It would be cool if you made a video explaining what they're actually doing in layman's terms, so maybe then the average person can understand that it's not magic lol
Even this kind of video is too advanced for the average Joe.
Like an OpenAi surrogates, again, even quite far away from Llama, Alpaca and etc. Local here is only your wallet, you're keep using OpenAi services but with Microsoft search engine with full identification, location and etc.
Sabrina from Answer in progress channel showed in her videos a years ago how collecting data with Ai bot got her blocked at Google, with forever captcha linked to your ip address. She collecting a simple stat data for videos.
It's getting closer to the serious topic of who owns data and the open source community will be forced to make very sophisticated techniques to get access to what was collected or stolen by search robots of huge corporations.
@Hack Slash good feedback. I have deliberately stayed away from terms like 'have sparks of AGI' at the cost of getting views because I think these things are getting over hyped.
"I dont think this is approaching agi in any way" - see you in a few month
I'm not sure it's actually attempting to be an AGI (other than recognizing the value of the label right now for marketing). However, it might end up being the core pattern that heads in that direction anyway. Human thought is a constant lookup of related ideas-but, instead of biasing the selection based on our experience (weighted information, and weighted information connections), an LLM more or less has a single, repeatable response to every input, and it won't change until the model is updated/refreshed. Humans change dynamically as our sensory organs are like dynamically updating weights/parameters, and we're heavily biased upon recent information. In a way, we have built in randomization. ;)
Next up, we rewrite AGI to work on Vicuna and make explosive noises in the world! 🤣
Would you expect more from a baby?
I'm not sure Pinecone is needed for this. A Python list would work just as well for semantic search.
Good luck with requesting Google servers outside browsers to get your ip in global block lists.
Uhm, this is pretty much exactly what auto-gpt does which was released like two weeks ago or so? Including the usage of Pinecone, etc.
First View!
And I'm the second one
I wish people wouldn't say things like "I don't think this a approaching AGI in anyway" without defining it first. By my definition we're already at AGI. It's artificial...it uses tools without instruction and has a feedback loop that let's it improve on itself. I've yet to have anyone tell me how humans are substantially different other than being organic and having an insanely efficient processing system. If you want to say that your idea of AGI is conscious then you need to state that as part of your definition. Otherwise, it just sounds like humans being arrogant about what we are. We're not special and that fact is going to very present soon.
These "AGI"s are over hyped garbage. Anyone who more than five minutes with them sees the novelty wear off. They can't do anything even slightly complex and most of what they actually can do is done more easily, and faster with just straight up chat GPT with the slight advantage that they can (when they don't fail) scape the web for info. But again with chat GPT plugins the same will soon be available to everyone and bing and phind already currently hook GPT up to the internet. It's interesting to see a LLM reasoning through some basic tasks but nothing more.
You don't understand, it's same ChatGPT but with that plugins you are taking about, you are using openai identification key for any work of it in Internet through Microsoft search servers, because Google banning machine search requests by command line. It's impossible to use them outside big brother.
@@fontende what don't i understand? I think it's more likely you didn't understand my comment.