$5 MILLION AI for FREE
Vložit
- čas přidán 25. 06. 2024
- Imagine an AI where, all in the same model you could Translate languages, Write code, solve crossword puzzles, Be a chatbot and do a whole bunch of other crazy things.
In this video, we check out the BLOOM large language model. A free and totally open source 176B parameter LLM.
BLOOM model: huggingface.co/bigscience/bloom
Quick examples of running BLOOM locally and/or via API: github.com/Sentdex/BLOOM_Exam...
Neural Networks from Scratch book: nnfs.io
Channel membership: / @sentdex
Discord: / discord
Reddit: / sentdex
Support the content: pythonprogramming.net/support...
Twitter: / sentdex
Instagram: / sentdex
Facebook: / pythonprogramming.net
Twitch: / sentdex
Contents:
0:00 - BLOOM model basics
3:05 - What's a Large Language Model (LLM)?
4:06 - What's Prompting?
6:40 - BLOOM Training Data & Model Behavior
9:09 - Tokens & Tokenization
12:03 - Using your $5M AI (How to prompt)
16:49 - Advanced Prompt examples
21:16 - What's Next?
#deeplearning #artificialintelligence
I was one of the people working on this model. It is an honour. Thank you for posting this!
Thank you!!!
Thank you so much ❤️
Thanks
I'm a structural engineer, I've no idea what all these are about... But I KNOW this project will gives access to high level AI to freelance developers and in turn help a lot of people.
The Internet is mostly built on open source stuffs. I hope AI is too
Hats off!
this is like giving someone a 5000 horsepower engine and telling them they need to build a car for it lmao, thanks for the explanation!
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
Haha
When are you gonna use the AI to generate a story for a video? :p
analogy
The answer is the right question
19:10
Model: *throws an error*
Dev: *feed the error message to the model*
Model: wait... that's illegal.
lol, inb4 people just generate memes using this
@@jan_harald I’m convinced this has been happening for years now lmao
Basically a model that develops itself. An older version of the model which works can have the error be fed into it from which it fixes the new model and the cycle continues. An infinitely upgrading model
It's interesting how we can be 100% responsible for creating something, and then not know exactly what it does
Never thought of that its a nice thought
@@ghostbluster8900 Definitely. Will save this one.
kinda like our systems we live in.
what could go wrong, right?
any parents kids say hi
I had the chance to do some work with the OpenAI team to build a Q&A demo with GPT-3. I saw a ton of their prompt engineering tricks for producing awesome multi-part answers like what you demoed here. But another trick I learned for more specific information (e.g., less general knowledge), we could pair GPT-3 with a retrieval engine, so a question would trigger a vector search to retrieve relevant paragraphs from a big database (e.g., loads of text scraped from the internet), and then feed those retrieved paragraphs *and* the question into GPT-3. So it would be like "answer using ... etc
The quality of those results were beyond anything I could have imagined. LLMs with good prompt engineering are mind-blowing.
Oh god I need this for when someone asks me a easily google able question. “Answer (question) with (first result on google)” auto reply bot
@@them2545 yeah it's perfect for that haha, I demoed it in a video, search for "How to build next-level Q&A with OpenAI"
@@jamesbriggs Thanks for leaving this comment. Subscribed to your channel.
Dude, every video you have posted and the python book series are all so golden, thank you for uploading videos as usual.
I completely agree, also I was perplexed by the small view count on some of them, it's unfair
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
Yeah, I‘m feeling the same way.
What python book series???
١١١١١١١١
23:30 "what an incredible time you know to to be alive" Hahaha, that sounded a lot like a 2 minute papers reference right there. Nice, I agree with the sentiment.
id definitely love to see a prompt engineering video! like looking at the training data and figuring out interesting ways to extract information
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
ID=the id
Awesome, i thought this video will be about dall-e
WHAT? Get back to Factorying Trupen, no time to rest.
Dall-e free to try up to a certain amount of credits then you pay, you get 15 free credit a month. I wonder how they'll monetize this project...
Dalle is not free
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
@@waseksamin3216 researcher like me got early access with a total of 150 credits along with months of free access in the early months
Sentdex, you remain one of the best channels on youtube, thank you so much for your hard work, I'm currently making my way though your neural network from scratch book. Really appreciate everything. This is so awesome.
I respect the way you explain things in the most coherent way possible, capturing a broad audience.
Excited and motivated!
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
I really enjoy your last dozen videos that broadly cover different AI advancements.
Duuude the prompt engineering stuff!
Honestly if you make some accessible videos on how to effectively generate prompts then I think you could easily attract an incredible amount of people.
Agreed!
Your idea of a job descriptor: "Prompt Engineer" solicited a thought. I've been waiting for a job description of a knowledge engineer to become a real and common, someone who distills knowledge from information using the tools of machine learning and artificial intelligence. We live in an exciting time. The world is on the cusp of a paradigm shift, a world where AI & machine learning is as common as SQL databases are now...
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
Your idea of a job descriptor: "Prompt Engineer" solicited a thought. I've been waiting for a job description of a knowledge engineer to become a real and common, someone who distills knowledge from information using the tools of machine learning and artificial intelligence. We live in an exciting time. The world is on the cusp of a paradigm shift, a world where AI & machine learning is as common as SQL databases are now. The world is going to need a lot of knowledge engineers, and I think that the job description is going to be a lot more interesting than the one you have in mind.
there you go, I got BLOOM to finish your paragraph for you.
I had no idea that large language models could be this powerful. Holy cow. I'm definitely going to play with this like a ton.
On one hand it’s extremely elegant and beautiful, on the other it showcases the predictability of humans. Both hands hold something equally fascinating to me.
I experienced myself when using base model and instruct model of GPT-3, where instruct models are tend to generate the result based on the prompt that more careful, and gives the 'close' to expectation result, though sometimes it went off too far. But nonetheless, if we start with no examples or zero shot, don't expect too much for the model to generate the results like you want it to, instead, give them more data, give them something to follow so the model can generate like the way you expect it to do .
Wonderful video. I have already thought of some incredible uses for this and will begin once I arise from slumber! Thank you kindly for helping bring this to the masses.
Good luck!
Love to see more prompt engineering from you, best video from you, till now
Thank you very much for covering this so well. I’m blown away
I'm so glad I have a come across you videos, this has just complimented so much new knowledge and insight to my own machine learning journey. Thank you!
Loved the video! Yes, more on BLOOM application ( separate playlists if needed), please sir
Hahaha some of the AI's responses had me getting up out of my chair, and when you mentioned you did the same I had a good laugh. Thanks for the video man, all I want to do now is think of a cool way to utilize this. Amazing
Your videos are always a breath of fresh air. Straight to the good stuff, no other bs. You've helped me the most in my python journey and I will always be grateful for that.
this may be the best video you've produced! (and I've watched most of your channel!)
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
Advanced prompt examples really blew my mind.can't believe that such an amazing tool is available for free
I was 70% paying attention as I’d mostly seen and understood the chatbot thing while using GPT-3, but then the error prompt came up. That’s when I was impressed! I will definitely be trying this on GPT-3 next time I’m an idiot.
How fun! A Dynamic Chatbot that uses Sentiment Analysis to dynamically have a reply. You're awesome Sentdex, keep up the good work!
Oh...how far you´ve come. Look at you writing books and commenting the latest AI models. It´s been as interesting to see your own mental and technical progress through the years as it´s been to see the technology you use and comment on. Well done.
20:47, "I would like to try the cheese of sheep." Wow, that nicely encapsulates old Monty Python and recent events! :)
I am still in the early part of the book but man is it good. And the qr codes linking to example videos.. chefs kiss.
the video quality is just amazing, hat off to you sir
Holy crap. This is my first time back to this channel since I learned Python from you in like, idk, 2010?? Amazing evolution, never thought I'd be back here. Of course GPT3 is the thing that eventually brings me back ;D
That's how machine learning should be open source, open to public, and worked on it by public instead of big corporations propriety AIs, competing each other ultimately leading to doom of humanity.
This AI model has more chances of becoming 'general good consciousness', compared to big tech companies, although humans are selfish but tend to be kind and empathic compared to big tech companies emphasizing only greed and selfishness.
The human you're describing and the average internet user are two different things
How does making a model open-source bring us any closer to solving the alignment problem? If anything, it puts misaligned AI in the hands of the average person.
I don't see how a world wherein everyone has access to unfriendly AI is better than a world wherein only a few corporations have access to unfriendly AI.
@@The_Original_Hybrid It could mean more people being interested in AI and more funding for it, so that's a plus, though I do agree that a strong AI could very well be like nukes. Not something you want to have just available to everyone.
You are precisely wrong. Competition is what drives progress and human benefit. Creating stuff and then not monetizing is nothing but slavery.
Simply look to every communist or socialist government ever, total disasters. People should create value for others so they can get paid, anything else is slavery and degeneracy.
Honestly OpenAI's CEO proposed a future in which companies like there gatekeep AI and sell access out to secondary businesses, and that idea scared me enough that I started deep diving into projects like this seconds after.
I love your videos, man. Always so well explained and engaging and integral to me learning how to code some 5 or 6 years ago, which btw, earned you an acknowledgment in my (neuroscience) PhD thesis, haha!
Hah, awesome!
2:16 "i love you" in hindi is "tu hi meri jaan he" is so flirty because the literal meaning of "i love you" in hindi is "mein tumse pyar karta hun" and "tu hi meri jaan he" literally means "you are my life", A.I. has reached some serious heights.
@sentdex, this CZcams channel is truly such a MASSIVE gift to humanity and the universe.
I can’t wait to build some projects using this technology!
I've been trying the GPT-3 playground the last few days and Bloom looks like it's almost the exact same. It can really be useful.
Hugging face is going to make me a data scientist
Thank you so much for sharing this.
This is insane. Great video mate!
Performing sentiment analysis of user input initially and then dynamically changing the input prompt based on the sentiment will be awesome. 🤩
Hey man,
Thanks for existing! You have helped me alot throughout my college life and after that as well. Keep it up!
I have a suggestion could you please reduce the light intensity in the background and keep more light focus on your face. The background looks little fake in current lighting setup
this is insane. thanks for the explainer sentdex!
this is a great video man, great explanation
Been watching your content for a while. Happy to say I've invested in a hard copy of nnfs!
This video was fun to watch and super mind blowing.. AI is increasingly getting sophisticated!
You're a GIGACHAD for breaking down this info. Buying your book now :)
Awesome 👏 I just bought that cool neural network book as a hardcover from your site 🤓
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
One of the reasons i never used cortana or had any virtual assistant in my computer is precisely because it doesn't run on my machine and my data isn't safe so i am happy that people are making AI in an open way
{'error': 'Model is overloaded, please wait for a bit'}
It worked for about 20 min
I guess people found this video interesting
17:45 - I've started using this method in the OpenAI playground to write a book with GPT. It's just as jaw dropping as Mr. sentdex says!
This was a great video. Thank-you for sharing this knowledge with us
Hey mate!👆 Tell Max Thomas you were refferred by me for the right strategy on his new project to participate on it.Thank you for the comment ✍️
7:09 "Peruse" a word I know thanks to the show "Silicon Valley" hahaha. Great video!
I would really like to see something like the AI Dungeon which is kind of a chatbot but makes up an adventure with the user.
AI Dungeon Master!
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
Mate, with you on that ,spent many hours in common room of dorm playing DnD
Oh, that's perfect.
awesome video bro, i can feel your passion
I've spent some time with open ai's playground. I was asking it questions about something specific and it invented a technical definition, a non-English vocab word to describe it, not knowing what the word was I asked it to define the word. It gave an elaborate definition. I googled it - it's not a word at all. I wish I remember what it was however, I was busy working on making e-commerce store product descriptions and just quickly went back to work. While I'm a programmer at heart "AI prompter role" you've been describing is very much going to be a job I can already see it.
Reminds me of those Least Common Denominator arguments around the supercomputer. I also need this for vim.
It's working thanks my friend
Another mindblowing result I got. The model can "run" javascript :D
Code: function abc(a, b) { return a + b };
Run: abc(1, 3);
Result: *4*
In my first test I used "sum" as function name and it worked too. The I changed it to "abc" to make sure it is realy understanding the code.
of course it will not work with every digit.
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
What are the chances this video's script was generated through BLOOM? Good video dude!
Hey mate!👆 Tell Max Thomas you were refferred by me for the right strategy on his new project to participate on it.Thank you for the comment ✍️
This does sound fantastic. Your description of this made me think that this sounds just like long term memory in our brains. Something triggers our "memory" which then floods our mind with many different memories related to that perceived trigger. These "memories" in our long term memory would include memories from all of our senses like how it taste, what it feels (touch) like, what it looks like, the sounds it makes. These will also include facts and figures, names, passages like from literary works, etc.
As I see it, your "prompts" are memory triggers. What if this idea of BLOOM as a long term memory was used as part of a bigger whole connecting to many different parts such as sensory processing, inference building, pattern recognition, etc. Each piece of this would be a major project unto itself. But boy would that be amazing!
Thank you for sharing your thoughts on this technology. There is so much out there to be discovered, revealed, and dreamed of...
Amazing video - a lot of useful information
That recliner review was crazy
Hey mate!👆 Tell Max Thomas you were refferred by me for the right strategy on his new project to participate on it.Thank you for the comment ✍️
Oh my fricking god I'm learning machine learning from now I can't believe how epic bloom was until you showed me how one can tweak with prompts to get what one wants,
I'm quite good at linear algebra, statistics and python language, and currently working in JavaScript and node,
I'll keep working on it as a part time thingy.
Hey mate!👆 Tell Max Thomas you were refferred by me for the right strategy on his new project to participate on it.Thank you for the comment ✍️
WOW ! this is really great, can't wait to try it.
This is incredible work!
Hey mate!👆 Tell Max Thomas you were refferred by me for the right strategy on his new project to participate on it.Thank you for the comment ✍️
Really looking forward for a similar thing to happen for image generation like DALL-E or Midjourney
I suggest checking out stable diffusion if you haven't already heard about it.
@@Rctdcttecededtef I heard about it as soon as the announcement dropped. Sadly, it's still closed for now, but looking really exciting. And the code is already open-sourced
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
@@tsukuyomin I’m sure it’s not closed? I got instant access a couple of weeks ago. I used it once and went back to MidJourney though seeing it’s way faster and less messing about. Maybe I was using it wrong though.
Hugging Face released the diffusers library recently, which is maybe one step closer to this
Great Video! I just wanted to go to bed, but now i am stuck with every minute of your video :D
I wonder if anyone is applying this to Radiology? Humans are still better than AI at making image observations but not as good as prioritizing differential diagnoses. So if BLOOM has seen a corpus of radiology reports (I know Stanford has a large one) then you might be able to prompt, "I see a,b,c in an CT Scan. What is this most likely?" And see what it says.
Using AI on radiology is one of the student intern projects at the University of Queensland. But I doubt BLOOM would be applied. I reckon researchers want to patent and profit off things they developed. It's hard to patent a technology that's based on an open source. If they're selling the software, it can be easily reverse engineered too so trade secrets won't work.
@@anonymouschriz4687 Not sure this is correct. Many commercial products (Google, Facebook, etc.) make use of open source components. No issues that the lawyers can't figure out. There are ways to protect software. This is just a question of introducing Radiology researchers to a potential new tool. Worth kicking the tires.
@@anonymouschriz4687 Just make FOSS
@@DavidChannin I would imagine if it had all the relevant radiology text it could do what you're describing. I don't know how likely published medical research is to be incorrect though. I've read that a lot of psychology research isn't reproducible. I could imagine errors in papers contaminating answers to prompts about diagnoses. And maybe there are more focused, less computationally expensive models that can do the job better.
I become soooooooo happy when I see people this enthusiastic about tech like I feel you bro, indeed what a time to be alive!!! 💯💯💯
This is an ACTUAL "open" AI.
I think its disgusting that "OpenAI" COMPANY gave themselves the name "open" for marketing.
they were open at first but they found out that they cant really handle $8 million loss per year
@@user-cc5kl7qv8f
If they didn't have a business model for "open AI" they shouldn't have used the name.
If they can't operate openly, then they should CHANGE their name.
They are leveraging user trust in the word "open", for marketing purposes.
You can literally go to the chatgpt thing and interact with it for free. Ive already used it for programming a python script and let it consult me over various internet security topics.
I didnt spend a Dollar for that. Theres maybe this 3 times better thing they didnt made public available...
I dont care its already great.
@@danielbrunk9121 I don't know if you have heard but they are going to make it paid. "Costs are too high."
@@danielbrunk9121 It's free because it's in beta and the feedback is valuable for them, but the only way to make something free for everyone is making it open source
First time coming across this channel nd wow i cant believe i'm just finding u for the first time.
Second, mahn i'm afraid that once ai reduces the work of basic thinking for us that humans will regress into some problematic behaviours where they overly depend on emotions and aesthetics to guide thought rather than finding a balance between what you feel nd genuinely seeking what you dont know
The most under-rated channel in the whole digital world. This channel should have at least a 100 million views but if you already know this channel consider yourself among the chosen few.
Hey mate!👆 Tell Max Thomas you were refferred by me for the right strategy on his new project to participate on it.Thank you for the comment ✍️
Great explanation!
The AI can even make ASCII art! This is amazing!
insta-like before I even watch since I know this is the best python youtube channel :)
Your book is awesome.
I feel like these multi lingual models could be cool to use as a chatbot to help learn languages. Once you are at an intermediate level, you could have conversations where BLOOM is prompted to use low level vocab and respond in the language of your choice. Maybe even add in some good text to speech and vice versa!
Checkout Quazel. They are doing exactly that. With the text to speech. You will be able to practice any imaginable interaction. Imagine learning french by talking startegy with napoleon. The future is mind blowing. What a time to be alive.
Awesome explanation. Thanks for sharing.
'A new job enters the bar...'
I would be very much interested in how to produce paragraphs or text blocks with a given topic. That would increase my freetime a lot ( as an informatic student with a lot of irrelevant lessons, i have to write a lot of essays ) Thanks for your dedication to produce such valuable videos. Always worth to watch. PS: Love your Book :)
Outsource your essays to me for as low as $25 for 2000 words
Right, why would you want an artificial intelligence when you can have a natural one 😄
@@artyman162 Because its free :) and funny to play along with it
@@elvispiece Well, it has to be in german and we have to do it alway. s within aworkgroup so we can define our next step, details and topics to write about. So thanks for the offer :)
Writing is the second most powerful skill you can have. You only harm yourself by cheating
Great video, interested in prompting chatbot with bloom
I’m imagining something like this could be used to complete/assist developers with difficult software problems in the future if we feed the algorithm incomplete/broken code with instructions of some sort. Just one possible use case that popped into mind
Hey mate!👆 Tell Max Thomas you were refferred by me for the right strategy on his new project to participate on it.Thank you for the comment ✍️
¡Señor Sentdex, es usted un genio, muchas gracias! 😀
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
“If something online is free, you are not the customer. You are the product.”
Hey mate!👆 Tell Max Thomas you were refferred by me for the right strategy on his new project to participate on it.Thank you for the comment ✍️
Definitely want to learn more about this
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
Wow great thing great video, thank you
How do I get an API key to use this model in Google Colab?
See description for an API example. To get a key, make a huggingface.co account and you can get a key in your settings
@@sentdex Oh I see, thank you! And how do we store API keys securely using Google Colab?
this is really mind blowing
Wow, this is another step forward; these other models couldn't have persisted in a closed source setting forever; if anything, this encourages competition!
Amazing!
Love you brother
This is mind blowing!
What about employing model itself as a prompt engineer? Create a feed back loop of prompts and explaining how prompts could be improved:
"AI language models are powerful tools for problem solving, however they require correct and elaborate prompts to give out results we want. For example, we could ask how to defeat cancer once and for all with this following prompt.
Prompt: "
And feed result back to AI, specify what are the mistakes and continue this "dialogue".
I have some thoughts about these very large models (mainly because of their size and internal complexity), if you treat the prompt as kind of like "short term" memory. In the context of a chatbot, the previous prompts, the model will "think" it last outputted those tokens.
Bringing around to my actual idea. This is very similar to boltsman brains, hypothetical brains that come into existence spontaneously in the vacuum of space, with a full set of memories, it makes "predictions/thoughts" about its next action/output based off its "remembered" last thoughts, makes an output and then spontaneously disappears again.
This feels somewhat similar to how current NN models work, we feed it a input, it runs and returns an output.
Could this be how our own brains work? except our brains are a constant activation of neurons and not a single feed forward activation of neurons.
Am impressed by the development of Bloom. At the end of blog you a bit of history on development of the A.I any chance you could expand on that, particularly the scientists involved ,it seems to me they need to acknowledged ,given that have made it opensource/FOSS .
it's not open source. There are usage restrictions on BLOOM
Wtf… this man is the truest evil genius in the world. I aspire to be at your level great Sentdex
I'm kind of surprised that a "corpus callosum" AI hasn't been invented yet. As it currently stands, all of these AI are very impressive, but also very rigid in how they expect input.
Like you mentioned, this AI isn't a chatbot, so it doesn't respond to your inputs -- merely expands on them.
But a sort of higher-level "corpus callosum" AI I would imagine to be a sort of "meta AI" or a "network of networks" in a sense. It would be a program/AI that is able to take a very general abitrary input, and then determine which "sub AI" to send it to.
So if you feed it an image as an input with a prompt, it would pass the input to it's "visual lobe". Or if you asked it a question using a microphone, it would pass the input to it's "audio processing lobe" and interpet your request.
In a sense, this "meta AI" itself wouldn't really be good at any one task, merely an intention interpeter to pass inputs to it's sub-specialized AI lobes.
It would basically mimic the function of the corpus callosum, how the human brain is made up of multiple specialized lobes, with fibrous networks that connect the lobes and allow them to pass information to each other in the brains own "native language" (Neuro/chemical signals).
Currently the state of AI is that each of these models acts as one specialized lobe/sector of a human brain... But they are all disconnected.
Imagine when something like MidJourney can directly interact BLOOM, etc.
I imagine that a simple heuristics program/AI would be a good place to start. For example, think of those 20 questions games that can predict what object you're thinking of by first narrowing down if it's a mineral, animal, etc.
A heuristics meta-AI would, for example, take a prompt like "draw me a cat" and be able to understand that I want an image as output, and pass my prompt into something like Dall-E. Or if I asked "write an essay on Albert Einstein" it would hook onto the word "write" and understand I want text output and pass my prompt to BLOOM.
I've tried to dive into bloom when I tried to understand it, you just made clear what I've seen but didn't understand...Thank'x for taking time to explain to mortal npc like me 😅
*and for that reason I just subscribed.... And Realy want that book..... I'll find a way to pay it 😂
Just wanna say to you dude..... tu hi meri jaan hai... seriously learned a lot from u....and still learning
Learn Python for free:
czcams.com/video/5lLkLi9SAlA/video.html
This is truly inspiring. I'm trying to use it for getting technical information out of repair inquiries for my small business. Unfortunately I have to translate it first, as Dutch is not (yet) supported. This makes it less accurate.
Just adding an unpaid and unprompted thumbs-up for the book , highly recommended.