All these released models are based on Llama. If it weren't for Llama, LLM Chatbots would all be closed source. Thanks to Meta (and 4chan for leaking their models).
the main benefit of chat gpt is context size as it is 2x bigger than this model while gpt 4 is like 4 times bigger so this bot will be able to make only very very short conversations.
With that open source code, there is absolutely no way that the development pause agreed by all CEOs working on AI will have any effect to slow down "cybercriminals".
True, it's one thing to contain and control some of the big online corporations when it comes to A.I. but it's another to control open source or locally run A.I. which more or less makes it impossible to control or contain by the ones that want to do that.
This is what many have been trying to tell the smooth-brains out there that think the letter has any merit whatsoever. The cat is out of the bag...and thankfully so.
@@jeffwads I don't think anyone thought the letter would actually convince anyone to stop, but it's more about communicating clearly their concerns. In that respect I think it's worthwhile to understand what their concerns are, even if what they're suggesting at the moment is unrealistic.
Thank you for sharing this tool. I'll emphasize that it is being compared to ChatGPT 3.5 in the comparison tool. Normally this wouldn't matter much, but it is telling that the comparison is being automatically judged by GPT 4.
Yes, great point. Its an interesting use case for GPT-4 as well. There is no objective way to score the output of a language model (its very subjective). So using GPT-4 is one way of doing it. Although another LLM (say Bard) would rate the same responses differently.
I am curious: on this scale, how does GPT-3 rank? Note that ChatGPT is GPT-3.5; if this model is equally as good as GPT-3, then we effectively have an open-source replacement for that model.
Wow I just asked another youtube channel why couldn't we have powerful Open-source Langugage AI like ChatGPT just like we have Stable Diffusion for text-to-image AI. And today I found your channel. So awesome. Thank you very much.
I tried GPT-3 and also Vicuna-13B and gave them both the same test. I gave them a task to complete. I said, "Compile a list of words using only 5 consonants and 2 vowels" (which I provided) I also asked them to make sure that 1 letter of the group of 7(which I also provided) must appear in every word. What do you suppose happened? Well, in short, they both failed miserably! The lists they provided were full of words using letters not among the 7 allowed letters. When asked to examine their list for errors they replied that every word contained only the letters specified. Wrong again. When challenged on a specific incorrect word, I received an apology for the mistake and a new list that was now "correct?) It too was full of similar errors. Words were also included that didn't contain the 1 letter that was supposed to be in every word generated. I repeated this exercise multiple times rephrasing my instructions to try to help it grasp what I was requesting of it, but it was no use. I'm assuming this is due to the newness of this technology, that it will improve over time, but I really thought this was a relatively easy task for any AI. I'll keep watching, but I'm not too impressed at the moment.
I am a newbee. As I understand, Lllama Alpaca, Vicuna are all pre-trained model. Is there a way to add to this pre-trained data? For eg: ask the model to read a bunch of documents I provide and then use that new knowledge to give me updated responses. Is this possible and how?
Yes, you can do that. The process is called fine-tuning. Llama is the original model that was trained by Meta on 1T tokens. Then Alpaca was fine-tuned (basically adding to the learning) on a new dataset. Same is the case with Vicuna. You can do that with your own dataset as well. I have another video on the topic which goes into the training part. Check it out (ua-cam.com/video/dF2eu-C87Pk/v-deo.html). I am making a video on fine-tuning so keep an eye out for that.
Alternatively, there's a Python package called Llama Index. You can feed it a bunch of documents and it can give the relevant parts to a LLM like ChatGPT and have it answer questions based on what's in the text.
@@engineerprompt First of all, thank you so much for responding. That video on GPT4All was very helpful. I eagerly look forward to your video on fine-tuning.
@@Smytjf11 Ever since you have mentioned that, I have been just reading about LLama Index as eagerly as I read my comics :) So basically what it does is I can take a pretrained model like Llama, then use some of my PDF's to add to that pretrained model, and query it just like I would with ChatGPT. Correct? Do you have any pointers/links/videos of a tut describing how to read PDF's and give the relevant parts to a pretrained LLM? I have created something similar using DaVinci but it uses the Open API and the queries are expensive. I want to do this locally on Llama (or similar to GPT-4) without incurring any query cost. I have a good GPU and 32 GB of RAM. Any pointers will be helpful, and thanks for reading my long comment.
I find them both censored and pretty much delirious. I wonder if learning after another AI model leads to accumulation of errors similar with Xerox copying after another copy.
I've tried running it on my machine locally but it only gives me simple short responses, if I ask it to do something It just keeps asking why and avoids doing it, eventually it just responds the same answer
@@engineerprompt Vicuna didn't seem to know either. :D The first answer was in English, but I asked for a translation and it has only gotten better since then. :)
Alpaca took me out with the final punchline! "-Well, I guess that's it for now. I can't top that nasty joke! -You already did! Amazing (sarcasticly) -I'm glad you enjoyed it! (sarcasticly)".
@Prompt Engineering, Can I use my own private dataset to have text classification, information retrieval without sharing my private dataset with Vicuna? How does this useful for private datasets? thanks
Yes, you can that do. You can check this video out to get some ideas (ua-cam.com/video/wrD-fZvT6UI/v-deo.html). You probably want to use a more powerful model than what I have used in that video but its possible.
@ Prompt Engineering, Thanks for the reply. I already tried it, the answers are not even close to what I expected. The token size was also another issue. Do you have something in your mind that would not store my private data in it, also giving flexibility of larger tokens, and higher accuracy? I watched your Vicuna and GPT4All videos. I am not sure, if they would be convenient ones? Thanks
Looking forward to the day when we can just install a program natively on our machine that can give us access to all the information we ever want, imagine the learning possibilities!
@@engineerprompt OpenAI ToS also restricts it's commercial use. Basically, Vicuna is a research purposes demonstration what -science have done- LLaMA based chatbots are capable of
Many commenters on this video about AI language models are impressed with the advancements in the AI space and the capabilities of the new models. They discuss the competition between different AI models, possible applications, and how these models compare to ChatGPT. Some users are curious about customizing models, using them for commercial purposes, or integrating them with other technologies. A few commenters point out the limitations and biases of these AI models, while others request tutorials or additional information.
Thanks for great content!! Would it not be considered biased for GPT-4 to judge its older release in comparison to another model? I assume GPT-4 /3.5 might have similar standard/values on certain subjective issues like grading a blog post article.
Just watch the video and tested it out myself and I've got to say, this is the closes I've seen to ChatGPT from an open source one. It will be interesting to see if this is doable locally on our computers and better yet, if it can be done so it can access data online for up-to-date information.
Yes, its possible. You will take an existing model and fine-tune it on your own dataset. I am making a detailed video on it but if you want to learn the process, check out this video (has details at the end of the video on the training process): ua-cam.com/video/dF2eu-C87Pk/v-deo.html
Which text generation model I can use for commercial use and which works nearly close as chatgpt may be 80-90%. Also it should be able to use with langchain. Which one do you recommend?
Unfortunately, at the moment, most of the useful ones are based on Llama which doesn't allow commercial usage. Check it this: github.com/stochasticai/xTuring
Vicuna-13B is an open-source chatbot that claims to achieve 90%* quality of well-known AI models like ChatGPT and Google Bard. It is trained on user-shared conversations from ShareGPT and fine-tuned on LLaMA, which allows it to have an enhanced dataset and a scalable infrastructure. The main objective of Vicuna-13B is to provide an alternative, open-source chatbot with comparable performance to proprietary models. The preliminary evaluation of Vicuna-13B, judged by GPT-4, demonstrates that it achieves more than 90%* quality compared to ChatGPT and Google Bard, while outperforming other models like LLaMA and Stanford Alpaca in over 90%* of cases. Vicuna-13B's performance is impressive and shows that it could be a potential competitor to other open-source models in the AI landscape. By providing an easy-to-use, scalable infrastructure and high-quality responses, Vicuna-13B aims to be an accessible and reliable chatbot for a wide range of users and applications.
Under the "Vicuna: An Open-Source Chatbot with Impressive GPT-4 Performance" demonstration, I notice that there are parameters specified, including "Temperature" and "Max Output Tokens." I am curious about the potential impact of these parameters on the chatbot's functionality and overall performance.
Good luck with that, in that many humans can’t seem to avoid inflicting their political leanings into everything they write where it could be done. For my needs/desires, I’d prefer to avoid political things and use AI for creating code, which seems most probable to not have any kind of political leanings by its nature. But see my first sentence…
@@strictnonconformist7369 If they display political bias, it means they're damaged goods. Political bias is about the ugliest deformity it could possibly come with. Unless the person is a drone that can't be without constant agreement and reaffirmation.
Why everyone is keep fooling themselves and viewers by constantly repeating “open source models”? The underlying LLAMA model by Meta is not open source, so the derivatives cannot be open source either. All the derivatives also used data generated by proprietary ChatGPT. If you take proprietary model and add proprietary data you cannot get open source as a result. You can open your sources but that ain’t gonna make the product FOSS
at 15:49 code it gave to you still had the same error as the original. so it didnt actually fix the code, you fixed it yourself. why did you made it seem like it was fixed by Vicuna?
Good to see a fellow programmer. You are right but vicuna was able to fix it. Didn't record the generated code. If you see the demo, I did say I had to play around with it for it to get it fixed but didn't record the actual code generation part.
I would have liked it if you showed what prompt you gave it when the code didn't work. Did you ask it to correct it? Not even that is necessary with GPT-4. Tell ChatGPT 'When I clicked on a button, it displayed an alert that said "Joke N does not exist.' That's it. It will infer that this behavior is not what was expected, that you want it to modify the code, and give you the corrected code.
Why don't you create a video tutorial and teach us how to install this AI on our PCs, and also teach us how to use it along with LangChain and Pinecone to have a strong AI on our own computers. Lastly, it would be great if you could teach us how to train it in other languages so that it's more similar to ChatGPT. It would be nice to have a graphical interface like the one used by ChatGPT as well. I know it's a lot to ask, but it would be a really interesting project.
Looks interessting. Thank you. I like GPT4ALL and "communicate" in german. Some test are looking quite good, but some other tests not so. So now I'm looking for information how to train, and how the "intelligenz" is working. Some problems to identify person name and locations.
After conducting a few tests, I feel that it falls short of the seemingly intelligent reasoning that GPT-4 displays. Perhaps improving the last 10% would bring it to the uncanny valley level.
This is not entirely correct way to evaluate language model as according to the source code this bot is roleplaying this setting "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions." which by my test is not exactly that good If you want more creative or philosophical question your should tell it that it is expert coder or philosopher or romantic partner and then put all those options to chose.
It was trained on 1T tokens (not sure if I mis-spoke in the video). The number of parameters for the largest model is 65B. Check out Table 2 in the paper: arxiv.org/pdf/2302.13971.pdf
Not sure what was going through their head when they named that model "Dolly". Why name it something that's a homophone of a much more well-known AI product?
This is what I mean when I say ai devepment is unstoppable there will always be sweaty programmers that can do this work all day on disability checks. Nothing is stopping 4chan or Reddit from creating a monster. I am sweaty to but I have a 9to 5 but if given the chance I’d be contributing more. Ai is the calculator of our century. But worst we have everyone the building blocks and now everyone has uniquely built ones
I wonder if AI programs are strong enough to use other AI programs to create another AI, this will surely have an effect on its quality but I wonder if it would exponentially worse or better
Soy peruano y me hace gracia los nombre porque en peru hay muchos auquenidos que se les llama así. Llama, alpaca, "vicuña" y falta uno que es "GUANACO" ojalá creen una inteligencia artificial por guanaco para que estén completos .
This bot is just as biased as ChatGPT. It'll happily generate a joke about men when prompted to, but ask it to make a joke about women and it'll lecture you about not wanting to offend groups of people. Pass.
i ask it to write me a story look Sure! Here's a short story about a cat written in the style of Franz Kafka: The cat sat on the windowsill, staring out at the empty street below. It was a grey, overcast day, and the cat's fur was matted with rain drops. The cat was lost in thought, wondering about the meaning of it all. Why was it here, in this small, cramped apartment? Why was it forced to live a life of endless boredom and solitude? The cat knew that there was more to life than this, but it didn't know how to escape. As the cat sat there, lost in thought, it suddenly heard a noise coming from the hallway. It tensed, ready to flee at the slightest sign of danger. But as the noise grew louder, the cat realized that it was just the landlord, come to collect the rent. The landlord was a tall, thin man with a stern expression. He never smiled, and he never spoke. He simply handed the cat a piece of paper and walked away, leaving the cat alone once again. The cat took the paper and looked at it, but it couldn't make sense of it. It was just a bunch of meaningless symbols, devoid of any real meaning. The cat let out a sigh and settled back onto the windowsill, feeling more lost and alone than ever. As the days passed, the cat continued to ponder the meaning of it all. It wondered about the purpose of its existence, and it longed for something more. But no matter where it looked, it couldn't find any answers. The cat was trapped in a world of endless confusion and despair, and it didn't know how to escape.
A problem with the response, seems always skip the first character, anyone faces the same ? m okay. ause I wanted to make it sound like it was about me, and I didn't want to sound like a robot. ou guys are such a cute couple.
Doesn’t anyone find it fascinating that the most human like AI are bad at math? Like even though its just a language model, it isn’t that great at those tasks which regular computers are pretty good at.
@Divergent Integral I know but I would have thought math would have been the first thing it learned to do well since it deals with probability and other formulas so much to understand human language. Its just interesting and maybe says a lot more about the nature of intelligence.
It's a *language* model. From what I understand, and in the simplest of terms, it's calculating the probability of what word comes next. So, if you ask it what 3 + 2 is, it's not going to add 3+2 but rather statistically figure out how often the word '5' come after the words '3 + 2'. So, you could say that it's not really thinking or reasoning, but for all we know that may be how our brains work too. Just maybe, most of the time, we don't actually reason either. We just talk. I can buy that. Could explain most comment sections ;) And yeah, Wolfram Alpha has solved this. The plugins are there and GPT4 can be told to get all the hard facts and calculation results you want. Funny though... one day it's some person going on about how he's hacked the APIs to get chat-gpt to pass queries to Wolfram, and a week or so later it's Stephen himself announcing the plugins. I spent my whole life living with the Moore's Law exponential curve, you'd think I'd be used to it. But, it still feels like this AI is accelerating freaky-fast.
@@4Fixerdave I can definitely buy our brains being prediction machines. Often we can finish other peoples words. Maybe it also plays into how our brains mirror each other at times. Maybe this is just part of the puzzle for intelligence. Maybe we have developed conversation and mirroring but something else is missing or silently emerging. I think what makes Ai feel different from Moores Law is that we are really at the beginning of this technology and we are experiencing it again. We are in an exciting time when we are going to see weird crazy shit. From randomized Ai with a wide range of weird personalities to full on work automation. As far as I know the Information Age officially ended on November 30,2022.We are officially in the age of Ai.
Prompt: someone gave me a riddle. try to find an answer: what does the crucifixion of jesus christ have to do with chocolate rabbits and fish? GPT-4: The connection between the crucifixion of Jesus Christ, chocolate rabbits, and fish lies within the celebration of Easter. Easter is a Christian holiday that commemorates the resurrection of Jesus Christ, which occurred three days after his crucifixion. Chocolate rabbits and fish, however, are not directly related to the crucifixion itself but are rather symbols associated with the celebration of Easter. Chocolate rabbits, or Easter bunnies, are a popular symbol of Easter in many cultures, representing new life and fertility. They are often made of chocolate and given as gifts during the holiday. Fish, on the other hand, are not as directly related to Easter as chocolate rabbits, but they still have a connection to Christianity. In the early days of Christianity, the fish symbol was used as a secret sign among Christians. The Greek word for fish, "ichthys," forms an acronym for "Iesous Christos Theou Yios Soter," which translates to "Jesus Christ, Son of God, Savior." Additionally, fish is often eaten on Good Friday, the day that commemorates Jesus' crucifixion, in place of meat as a form of fasting and penance. In summary, the connection between the crucifixion of Jesus Christ, chocolate rabbits, and fish lies in the celebration of Easter and the various symbols associated with Christianity and the holiday. Vicuna-13b: I'm sorry, but I cannot find a logical or meaningful connection between the crucifixion of Jesus Christ, chocolate rabbits, and fish. These are seemingly unrelated topics, and any connection between them would require a significant stretch of imagination or a specific context that I'm not aware of. Conclusion: I tried guiding it into getting it and I asked it "What about easter?" and it just explained to me what easter is. I'm curious about when these models actually get some kind of "smart" and work at similar levels as GPT but open-source and uncensored.
Try the Open Assistant (ua-cam.com/video/VFPrwxPBBVU/v-deo.html), here is the reponse: The connection between Easter, a Christian holiday celebrating the resurrection of Jesus Christ, and chocolate rabbits and fish is not immediately apparent. There is no biblical reference to either of these items in relation to Easter or any other religious event. However, there are several possible explanations for this seemingly arbitrary association: 1. Easter Bunny: One theory is that the idea of an Easter bunny comes from pagan beliefs in Europe, where it was thought that hares were sacred animals that could reproduce without sex. This may explain why eggs and rabbits became associated with springtime fertility festivals. In Christian tradition, the rabbit came to symbolize new life because of its rapid reproduction rate. Over time, the concept evolved into the modern-day Easter bunny who brings candy and treats, including chocolate bunnies, to children on Easter morning. 2. Chocolate Fish: Another explanation lies in the Catholic tradition of eating fish on Fridays as well as during Lent, which precedes Easter Sunday. Chocolatiers saw a market opportunity and began producing chocolate fish to satisfy customers' sweet tooth while still observing their Friday fasting traditions. Thus, the popularity of chocolate fish grew, particularly among Catholics, and they eventually became part of many people's Easter traditions. 3. Springtime: Finally, both chocolate bunnies and fish can be seen simply as springtime symbols of renewed growth and abundance after the long winter months. As chocolate making techniques improved over the years, confectioners began creating elaborate designs and moldings out of chocolate, resulting in cute, eye-catching products like chocolate bunnies and fish that proved irresistible to consumers. Today, these two items continue to be best sellers around Easter. While none of these theories directly links chocolate rabbits and fish to the crucifixion of Jesus Christ, all three illustrate how different cultural influences, culinary traditions, and consumer trends come together to create unique customs and rituals associated with major holidays like Easter. So, whether you prefer chocolate rabbits or fish (or both!), just remember that behind every delicious treat is a rich history that makes them even more special. 😊
⚠️ This transition sound every now and then is really annoying. Can you not use it ? Besides, it is 200% louder than your voice. You don't need this sound. Thanks. 🙏🙏👍
There is no way this is even close to chatGPT. ChatGPT was trained on basically the entire internet, which is what? Exobytes of data? The engineering challenges alone are something no one is even close to replicating, let alone modeling and fine tubing. The resting is flawed if a volunteer open source project is getting 90% of what ChatGPT is getting.
You are probably right but I think there is another perspective that is important. The goal doesn't have to be able to get similar performance like chatGPT on all the tasks, rather you can fine tune these models with much smaller datasets for a very specific task. These task specific models will be able to beat chatGPT on those task. Look up BloombergGPT. Second important thing to consider is that this shows the feasibility of fine-tuning your own models with much lower costs. I have seen some very interesting use cases by talking to so folks in the industry.
TLDR: Vicuna is nowhere near as good as chatgpt. Vicuna 70k questions is quite tiny in comparison to chatgpt billions of words and phrases. Specifically a dataset of over 45 terabytes of text data, which includes a wide range of sources such as books, websites, and other online content
You to learn more about how these models are trained, check out this video: ua-cam.com/video/dF2eu-C87Pk/v-deo.html
Commercial use?
Like a knife, a weapon available to everyone is not a weapon anymore it's just a mere tool... #hail_to_the_open_source
@@Shabasky1 No.
@@patrickangel4880 Agree!
As I noticed Vicuna is heavily cut by various censorship and affected by woke speech
All these released models are based on Llama. If it weren't for Llama, LLM Chatbots would all be closed source. Thanks to Meta (and 4chan for leaking their models).
Agree 👍
I hope opensource takes over this space honestly. Cant fucking stand open ai and them imposing their philosophy on the rest of us.
Thanks for the lore. I just looked it up.
now the barrier for entry is hardware.
@@xsuploader More than likely the government stepped in so they could benefit from the data and police what people are asking.
Getting GPT-4 to rate and compare the output of different models is a very interesting approach
I agree, this opens up a lot of use cases.
GPT-4 Can rate is own output (reflextion) and it make that more accurate
It's good to have competition for openai, so that they release extension features very fast. They are game changing.
Agree with you. Its always good not to have monopoly :)
the main benefit of chat gpt is context size as it is 2x bigger than this model while gpt 4 is like 4 times bigger so this bot will be able to make only very very short conversations.
Humans created AI, now AI creates AI, and soon AI will create humans.
@@engineerprompt it's never good, unless the monopology is open source.
@@kou6244 what do you mean by AI will create humans?
This model is so much better than I expected, I expected it to fall at multiple languages but seems to do both Japanese and Russian fairly well
We could ask gpt4 to translate those conversations into a whole bunch of different languages and train on them as well.
With that open source code, there is absolutely no way that the development pause agreed by all CEOs working on AI will have any effect to slow down "cybercriminals".
True, it's one thing to contain and control some of the big online corporations when it comes to A.I. but it's another to control open source or locally run A.I. which more or less makes it impossible to control or contain by the ones that want to do that.
Anyone who thought that could ever work were just deceiving themselves honestly.
This is what many have been trying to tell the smooth-brains out there that think the letter has any merit whatsoever. The cat is out of the bag...and thankfully so.
@@jeffwads I don't think anyone thought the letter would actually convince anyone to stop, but it's more about communicating clearly their concerns. In that respect I think it's worthwhile to understand what their concerns are, even if what they're suggesting at the moment is unrealistic.
The point is to stop significantly better models being trained. All the open source stuff is never going to be better than the state of the art.
Is it possible to make that locally installed Vicuna talk to a MySQL relational database?
Right now the weights are not released but once they are, I think it will be!
Thank you for sharing this tool. I'll emphasize that it is being compared to ChatGPT 3.5 in the comparison tool. Normally this wouldn't matter much, but it is telling that the comparison is being automatically judged by GPT 4.
Yes, great point. Its an interesting use case for GPT-4 as well. There is no objective way to score the output of a language model (its very subjective). So using GPT-4 is one way of doing it. Although another LLM (say Bard) would rate the same responses differently.
@@engineerprompt Could always ensemble. Ranked choice voting across multiple models
@@Smytjf11 Yes, that could be a good approach.
I am curious: on this scale, how does GPT-3 rank? Note that ChatGPT is GPT-3.5; if this model is equally as good as GPT-3, then we effectively have an open-source replacement for that model.
I asked Vicuna: are you free
It is taking an infinite amount of time to answer
Wow I just asked another youtube channel why couldn't we have powerful Open-source Langugage AI like ChatGPT just like we have Stable Diffusion for text-to-image AI. And today I found your channel. So awesome. Thank you very much.
Glad you found it useful. Consider subscribing :)
@Prompt Engineering Sure mate 👍🏼
I tried GPT-3 and also Vicuna-13B and gave them both the same test. I gave them a task to complete. I said, "Compile a list of words using only 5 consonants and 2 vowels" (which I provided) I also asked them to make sure that 1 letter of the group of 7(which I also provided) must appear in every word. What do you suppose happened? Well, in short, they both failed miserably! The lists they provided were full of words using letters not among the 7 allowed letters. When asked to examine their list for errors they replied that every word contained only the letters specified. Wrong again. When challenged on a specific incorrect word, I received an apology for the mistake and a new list that was now "correct?) It too was full of similar errors. Words were also included that didn't contain the 1 letter that was supposed to be in every word generated. I repeated this exercise multiple times rephrasing my instructions to try to help it grasp what I was requesting of it, but it was no use. I'm assuming this is due to the newness of this technology, that it will improve over time, but I really thought this was a relatively easy task for any AI. I'll keep watching, but I'm not too impressed at the moment.
I am a newbee. As I understand, Lllama Alpaca, Vicuna are all pre-trained model. Is there a way to add to this pre-trained data? For eg: ask the model to read a bunch of documents I provide and then use that new knowledge to give me updated responses.
Is this possible and how?
Yes, you can do that. The process is called fine-tuning. Llama is the original model that was trained by Meta on 1T tokens. Then Alpaca was fine-tuned (basically adding to the learning) on a new dataset. Same is the case with Vicuna. You can do that with your own dataset as well. I have another video on the topic which goes into the training part. Check it out (ua-cam.com/video/dF2eu-C87Pk/v-deo.html). I am making a video on fine-tuning so keep an eye out for that.
Alternatively, there's a Python package called Llama Index. You can feed it a bunch of documents and it can give the relevant parts to a LLM like ChatGPT and have it answer questions based on what's in the text.
@@engineerprompt First of all, thank you so much for responding. That video on GPT4All was very helpful. I eagerly look forward to your video on fine-tuning.
@@Smytjf11 Ever since you have mentioned that, I have been just reading about LLama Index as eagerly as I read my comics :)
So basically what it does is I can take a pretrained model like Llama, then use some of my PDF's to add to that pretrained model, and query it just like I would with ChatGPT. Correct?
Do you have any pointers/links/videos of a tut describing how to read PDF's and give the relevant parts to a pretrained LLM?
I have created something similar using DaVinci but it uses the Open API and the queries are expensive. I want to do this locally on Llama (or similar to GPT-4) without incurring any query cost. I have a good GPU and 32 GB of RAM.
Any pointers will be helpful, and thanks for reading my long comment.
So will models just keep being fine tuned with more data set releases etc?
Amazing... You always provide great value and updated news in the AI space. Really impressed. You have great knowledge about AI.
Thank you for your kind words. I am happy you find it useful :)
Vicuna is the Great Hero! I really dislike OpenAI as they now totally become CloseAI!
Well done .. a good thoughtful walk-thru of the Vicuna 13B model and it's capabilities.
bro, you deserved much more followers and likes, thanks for your amazing videos
Thanks for the kind words :)
I find them both censored and pretty much delirious.
I wonder if learning after another AI model leads to accumulation of errors similar with Xerox copying after another copy.
I've tried running it on my machine locally but it only gives me simple short responses, if I ask it to do something It just keeps asking why and avoids doing it, eventually it just responds the same answer
I haven't tried it locally yet but will try it once I get a chance.
Amazing! Vicuna copes with the Polish language, although she makes minor mistakes. But progress can be seen even in one conversation.
Nice, didn't know if supported languages other than English
@@engineerprompt Vicuna didn't seem to know either. :D The first answer was in English, but I asked for a translation and it has only gotten better since then. :)
Alpaca took me out with the final punchline!
"-Well, I guess that's it for now. I can't top that nasty joke!
-You already did! Amazing (sarcasticly)
-I'm glad you enjoyed it! (sarcasticly)".
@Prompt Engineering,
Can I use my own private dataset to have text classification, information retrieval without sharing my private dataset with Vicuna? How does this useful for private datasets? thanks
Yes, you can that do. You can check this video out to get some ideas (ua-cam.com/video/wrD-fZvT6UI/v-deo.html). You probably want to use a more powerful model than what I have used in that video but its possible.
@ Prompt Engineering, Thanks for the reply. I already tried it, the answers are not even close to what I expected. The token size was also another issue. Do you have something in your mind that would not store my private data in it, also giving flexibility of larger tokens, and higher accuracy? I watched your Vicuna and GPT4All videos. I am not sure, if they would be convenient ones? Thanks
Looking forward to the day when we can just install a program natively on our machine that can give us access to all the information we ever want, imagine the learning possibilities!
With your knowledge, is this capable of reflection like gpt4? How much time do you think they need to develop this for accepting image input?
When sharing a website, it'd be great if you can also show/display the address-bar. Thanks for the great video.
thanks for the feedback. Will try my best.
It can be commercial?
Unfortunately no, its based on Llama so you will need to ask permission from Meta.
@@engineerprompt OpenAI ToS also restricts it's commercial use. Basically, Vicuna is a research purposes demonstration what -science have done- LLaMA based chatbots are capable of
Many commenters on this video about AI language models are impressed with the advancements in the AI space and the capabilities of the new models. They discuss the competition between different AI models, possible applications, and how these models compare to ChatGPT. Some users are curious about customizing models, using them for commercial purposes, or integrating them with other technologies. A few commenters point out the limitations and biases of these AI models, while others request tutorials or additional information.
Thats a great summary!
@@markfethney7086 It seems a chatgpt generated one made wth cut and paste of all the comments flow
@@sambas9257 are you a bot too? XDDD Brave new world!
@@sambas9257 absolutely, but did it use the youtube api or just a copy paste jobby....? the latter I expect.
Thanks for great content!! Would it not be considered biased for GPT-4 to judge its older release in comparison to another model? I assume GPT-4 /3.5 might have similar standard/values on certain subjective issues like grading a blog post article.
That's a good point. I also if that's the case. Vicuna is also trained on the chatGPT data so it could be a confounding factor in GPT-4 evaluation.
Just watch the video and tested it out myself and I've got to say, this is the closes I've seen to ChatGPT from an open source one.
It will be interesting to see if this is doable locally on our computers and better yet, if it can be done so it can access data online for up-to-date information.
Where can i download llama 13b model
check this out: huggingface.co/decapoda-research/llama-13b-hf/tree/main
It's so incredibly good. I played a two hour round of DnD with the AI as the game master.
Nice, did you use the web UI or local installation?
WebUI. The local version is to huge for my System even with a 3070.
I wonder if I can train specialized domain related data
Yes that's possible. Google for BloombergGPT
can you train it to be better to code debugging?
Yes, the weights are now available so you can fine tune it on your own task.
@@engineerprompt love your response time. .. but how do you make it specialized to the weight of a programming language
Is it possible to make a custom Language model and can you do a tutorial.
Yes, its possible. You will take an existing model and fine-tune it on your own dataset. I am making a detailed video on it but if you want to learn the process, check out this video (has details at the end of the video on the training process): ua-cam.com/video/dF2eu-C87Pk/v-deo.html
Which text generation model I can use for commercial use and which works nearly close as chatgpt may be 80-90%. Also it should be able to use with langchain. Which one do you recommend?
Unfortunately, at the moment, most of the useful ones are based on Llama which doesn't allow commercial usage. Check it this: github.com/stochasticai/xTuring
There is a way to install this model?
Yes, they have released their weights to be public. I will be making a video on it soon.
@@engineerprompt Awesome I will looking forward for it!
Vicuna-13B is an open-source chatbot that claims to achieve 90%* quality of well-known AI models like ChatGPT and Google Bard. It is trained on user-shared conversations from ShareGPT and fine-tuned on LLaMA, which allows it to have an enhanced dataset and a scalable infrastructure.
The main objective of Vicuna-13B is to provide an alternative, open-source chatbot with comparable performance to proprietary models. The preliminary evaluation of Vicuna-13B, judged by GPT-4, demonstrates that it achieves more than 90%* quality compared to ChatGPT and Google Bard, while outperforming other models like LLaMA and Stanford Alpaca in over 90%* of cases.
Vicuna-13B's performance is impressive and shows that it could be a potential competitor to other open-source models in the AI landscape. By providing an easy-to-use, scalable infrastructure and high-quality responses, Vicuna-13B aims to be an accessible and reliable chatbot for a wide range of users and applications.
Was this generated by ai?
@@bowack194 I think it is 😂😂😂
In 3d there is this concept called uncanny valley. Reading through that passage its sterile of any emotion.
@@user-zi6rz4op5l fr bro we gotta make a word for uncanny valley text...
@@bowack194 yes
"Sorry, that joke doesn't exist." was actually a nice joke
Agree :)
Awesome thanks for the introduction, love finding new ai alternatives
Thanks for posting, thorough and timely. I’d love to see some posts on models with longer context.
More to come!
Under the "Vicuna: An Open-Source Chatbot with Impressive GPT-4 Performance" demonstration, I notice that there are parameters specified, including "Temperature" and "Max Output Tokens." I am curious about the potential impact of these parameters on the chatbot's functionality and overall performance.
They have a great impact on the output. I am trying to experiment with them and document their impact. Will share in a future video.
hi here .i can't properly install on Windows
how can i do To make it run
Check this out: ua-cam.com/video/8BVMcuIGiAA/v-deo.html
I’m looking for an HONEST accurate ai with no political leanings. No child guard rails. I don’t need it to protect my mind. I’ll be in charge of that.
Good luck with that, in that many humans can’t seem to avoid inflicting their political leanings into everything they write where it could be done.
For my needs/desires, I’d prefer to avoid political things and use AI for creating code, which seems most probable to not have any kind of political leanings by its nature.
But see my first sentence…
Well said Captain!
@@strictnonconformist7369
If they display political bias, it means they're damaged goods.
Political bias is about the ugliest deformity it could possibly come with.
Unless the person is a drone that can't be without constant agreement and reaffirmation.
What use case are you applying ai to, that requires it to be un ethical?
@@Jaimequin Words can't be unethical. Only actions can be. Words are just words.
can it write and debug code ?
Check the video for the answer :)
Can you make video about shellgpt with API LLaMa please 🙏
Didn't know about shellgpt, thanks for sharing this. Will definitely look into it.
Why everyone is keep fooling themselves and viewers by constantly repeating “open source models”? The underlying LLAMA model by Meta is not open source, so the derivatives cannot be open source either. All the derivatives also used data generated by proprietary ChatGPT. If you take proprietary model and add proprietary data you cannot get open source as a result. You can open your sources but that ain’t gonna make the product FOSS
at 15:49 code it gave to you still had the same error as the original. so it didnt actually fix the code, you fixed it yourself. why did you made it seem like it was fixed by Vicuna?
Good to see a fellow programmer. You are right but vicuna was able to fix it. Didn't record the generated code. If you see the demo, I did say I had to play around with it for it to get it fixed but didn't record the actual code generation part.
Looks like Vicuna also comes with the same restrictions as ChatGPT 4, unfortunately
It can be re-trained with the refusal messages removed from the dataset.
Many jailbreaks for ChatGPT also work pretty well with Vicuna 😉
Why not use the 65B LLaMA?
Limitation: The hardware you will need to run that!
@@engineerprompt Did you see the recent Ram shrink with the 30B?
@@Thedeepseanomad No, can you please share the link?
I would have liked it if you showed what prompt you gave it when the code didn't work. Did you ask it to correct it? Not even that is necessary with GPT-4. Tell ChatGPT 'When I clicked on a button, it displayed an alert that said "Joke N does not exist.' That's it. It will infer that this behavior is not what was expected, that you want it to modify the code, and give you the corrected code.
Why don't you create a video tutorial and teach us how to install this AI on our PCs, and also teach us how to use it along with LangChain and Pinecone to have a strong AI on our own computers. Lastly, it would be great if you could teach us how to train it in other languages so that it's more similar to ChatGPT. It would be nice to have a graphical interface like the one used by ChatGPT as well. I know it's a lot to ask, but it would be a really interesting project.
Most of this is on my list, hopefully in a week or so :)
@@engineerprompt I already see all your videos, but you dont show how to install Vicuna or how to run in web ui, at least i dont find it,
@@sunkwolf There are tutorial/automated script videos already.
Looks interessting. Thank you. I like GPT4ALL and "communicate" in german. Some test are looking quite good, but some other tests not so. So now I'm looking for information how to train, and how the "intelligenz" is working. Some problems to identify person name and locations.
After conducting a few tests, I feel that it falls short of the seemingly intelligent reasoning that GPT-4 displays. Perhaps improving the last 10% would bring it to the uncanny valley level.
You should have a look at RWKV
Thanks for sharing, I wasn't aware of it. Transformer free is an interesting idea. Will have a closer look.
Imagine an open sourced AGI that there could be millions of differently aligned variants of...
that would be interesting.....
That'll probably be how it turns out. More than one AGI, all developed slightly differently.
This is INSANITY 😂😂😂😂, things are happening SO FAST, gonna get freaking whiplash
Yes, really hard to keep up with!
I tried Vicuna and it clearly trained on the on the skewed data used by charGPT
I agree with you!
But you can modify the skewed data yourself and re-train it.
Like a knife, a weapon available to everyone is not a weapon anymore it's just a mere tool... #hail_to_the_open_source
This is not entirely correct way to evaluate language model as according to the source code this bot is roleplaying this setting
"A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions."
which by my test is not exactly that good
If you want more creative or philosophical question your should tell it that it is expert coder or philosopher or romantic partner and then put all those options to chose.
Are you sure LLamA has 1Trillion Parameters ? source please
It was trained on 1T tokens (not sure if I mis-spoke in the video). The number of parameters for the largest model is 65B. Check out Table 2 in the paper: arxiv.org/pdf/2302.13971.pdf
Is it free like gpt ?
Not sure what was going through their head when they named that model "Dolly". Why name it something that's a homophone of a much more well-known AI product?
Sent you a ko-fi enjoy, Great tutorials.
Thank you, appreciate the support, keeps me going :)
why do none of these have realtime text to speech?
This is what I mean when I say ai devepment is unstoppable there will always be sweaty programmers that can do this work all day on disability checks. Nothing is stopping 4chan or Reddit from creating a monster. I am sweaty to but I have a 9to 5 but if given the chance I’d be contributing more. Ai is the calculator of our century. But worst we have everyone the building blocks and now everyone has uniquely built ones
Resoning?
with me-almost other AI, they basic use chatgpt power for it
the hell happened to the model and weights?
They "will" release them in future
Do u know Alethea AI's CharacterGPT is the world's first multimodal AI system for Text-to-Character Generation. We would love for you to check out .
Will check it out!
hei mate, how about google BARD?
Nah, Bard is way behind in the race.
Awesome!!! Thanks for keeping us up to date with the latest developments. Well deserved the growth of your channel!!!!
Thank you, I am glad people are finding it useful :)
What about non-english question?
Not sure, it depends on the training data. You can try it.
This surely breaks the ChatGPT/OpenAI terms of use :(
I wonder if AI programs are strong enough to use other AI programs to create another AI, this will surely have an effect on its quality but I wonder if it would exponentially worse or better
Well there is a whole field called AutoML. Its possible now
Are you a prompt engineer or just use reddit to post content on yt??
A lot more than that :-)
Soy peruano y me hace gracia los nombre porque en peru hay muchos auquenidos que se les llama así.
Llama, alpaca, "vicuña" y falta uno que es "GUANACO" ojalá creen una inteligencia artificial por guanaco para que estén completos .
google again at last place in the group of ai research and technologies, vicuna is gonna be huge considering is open source for real
Very good thank you for info
Glad you found it useful.
Power.
At 41 seconds Reasoning not Resoning
This bot is just as biased as ChatGPT. It'll happily generate a joke about men when prompted to, but ask it to make a joke about women and it'll lecture you about not wanting to offend groups of people. Pass.
i tryed its good. ty
i ask it to write me a story look
Sure! Here's a short story about a cat written in the style of Franz Kafka:
The cat sat on the windowsill, staring out at the empty street below. It was a grey, overcast day, and the cat's fur was matted with rain drops. The cat was lost in thought, wondering about the meaning of it all.
Why was it here, in this small, cramped apartment? Why was it forced to live a life of endless boredom and solitude? The cat knew that there was more to life than this, but it didn't know how to escape.
As the cat sat there, lost in thought, it suddenly heard a noise coming from the hallway. It tensed, ready to flee at the slightest sign of danger. But as the noise grew louder, the cat realized that it was just the landlord, come to collect the rent.
The landlord was a tall, thin man with a stern expression. He never smiled, and he never spoke. He simply handed the cat a piece of paper and walked away, leaving the cat alone once again.
The cat took the paper and looked at it, but it couldn't make sense of it. It was just a bunch of meaningless symbols, devoid of any real meaning. The cat let out a sigh and settled back onto the windowsill, feeling more lost and alone than ever.
As the days passed, the cat continued to ponder the meaning of it all. It wondered about the purpose of its existence, and it longed for something more. But no matter where it looked, it couldn't find any answers.
The cat was trapped in a world of endless confusion and despair, and it didn't know how to escape.
A problem with the response, seems always skip the first character, anyone faces the same ?
m okay.
ause I wanted to make it sound like it was about me, and I didn't want to sound like a robot.
ou guys are such a cute couple.
Doesn’t anyone find it fascinating that the most human like AI are bad at math? Like even though its just a language model, it isn’t that great at those tasks which regular computers are pretty good at.
@Divergent Integral I know but I would have thought math would have been the first thing it learned to do well since it deals with probability and other formulas so much to understand human language. Its just interesting and maybe says a lot more about the nature of intelligence.
It's a *language* model. From what I understand, and in the simplest of terms, it's calculating the probability of what word comes next. So, if you ask it what 3 + 2 is, it's not going to add 3+2 but rather statistically figure out how often the word '5' come after the words '3 + 2'. So, you could say that it's not really thinking or reasoning, but for all we know that may be how our brains work too. Just maybe, most of the time, we don't actually reason either. We just talk. I can buy that. Could explain most comment sections ;)
And yeah, Wolfram Alpha has solved this. The plugins are there and GPT4 can be told to get all the hard facts and calculation results you want. Funny though... one day it's some person going on about how he's hacked the APIs to get chat-gpt to pass queries to Wolfram, and a week or so later it's Stephen himself announcing the plugins. I spent my whole life living with the Moore's Law exponential curve, you'd think I'd be used to it. But, it still feels like this AI is accelerating freaky-fast.
@@4Fixerdave I can definitely buy our brains being prediction machines. Often we can finish other peoples words.
Maybe it also plays into how our brains mirror each other at times. Maybe this is just part of the puzzle for intelligence. Maybe we have developed conversation and mirroring but something else is missing or silently emerging.
I think what makes Ai feel different from Moores Law is that we are really at the beginning of this technology and we are experiencing it again. We are in an exciting time when we are going to see weird crazy shit. From randomized Ai with a wide range of weird personalities to full on work automation.
As far as I know the Information Age officially ended on November 30,2022.We are officially in the age of Ai.
@Divergent Integral obsolutlty !!!
You literally answer your own question.
- Why are most human-like AIs bad at math?
- Aren't most humans bad at math?
RLAIF: Reinforcement learning with ai feedback. We've just obsoleted ourselves
Very interesting idea!
Prompt:
someone gave me a riddle. try to find an answer:
what does the crucifixion of jesus christ have to do with chocolate rabbits and fish?
GPT-4:
The connection between the crucifixion of Jesus Christ, chocolate rabbits, and fish lies within the celebration of Easter. Easter is a Christian holiday that commemorates the resurrection of Jesus Christ, which occurred three days after his crucifixion.
Chocolate rabbits and fish, however, are not directly related to the crucifixion itself but are rather symbols associated with the celebration of Easter. Chocolate rabbits, or Easter bunnies, are a popular symbol of Easter in many cultures, representing new life and fertility. They are often made of chocolate and given as gifts during the holiday.
Fish, on the other hand, are not as directly related to Easter as chocolate rabbits, but they still have a connection to Christianity. In the early days of Christianity, the fish symbol was used as a secret sign among Christians. The Greek word for fish, "ichthys," forms an acronym for "Iesous Christos Theou Yios Soter," which translates to "Jesus Christ, Son of God, Savior." Additionally, fish is often eaten on Good Friday, the day that commemorates Jesus' crucifixion, in place of meat as a form of fasting and penance.
In summary, the connection between the crucifixion of Jesus Christ, chocolate rabbits, and fish lies in the celebration of Easter and the various symbols associated with Christianity and the holiday.
Vicuna-13b:
I'm sorry, but I cannot find a logical or meaningful connection between the crucifixion of Jesus Christ, chocolate rabbits, and fish. These are seemingly unrelated topics, and any connection between them would require a significant stretch of imagination or a specific context that I'm not aware of.
Conclusion:
I tried guiding it into getting it and I asked it "What about easter?" and it just explained to me what easter is. I'm curious about when these models actually get some kind of "smart" and work at similar levels as GPT but open-source and uncensored.
Try the Open Assistant (ua-cam.com/video/VFPrwxPBBVU/v-deo.html), here is the reponse:
The connection between Easter, a Christian holiday celebrating the resurrection of Jesus Christ, and chocolate rabbits and fish is not immediately apparent. There is no biblical reference to either of these items in relation to Easter or any other religious event. However, there are several possible explanations for this seemingly arbitrary association:
1. Easter Bunny: One theory is that the idea of an Easter bunny comes from pagan beliefs in Europe, where it was thought that hares were sacred animals that could reproduce without sex. This may explain why eggs and rabbits became associated with springtime fertility festivals. In Christian tradition, the rabbit came to symbolize new life because of its rapid reproduction rate. Over time, the concept evolved into the modern-day Easter bunny who brings candy and treats, including chocolate bunnies, to children on Easter morning.
2. Chocolate Fish: Another explanation lies in the Catholic tradition of eating fish on Fridays as well as during Lent, which precedes Easter Sunday. Chocolatiers saw a market opportunity and began producing chocolate fish to satisfy customers' sweet tooth while still observing their Friday fasting traditions. Thus, the popularity of chocolate fish grew, particularly among Catholics, and they eventually became part of many people's Easter traditions.
3. Springtime: Finally, both chocolate bunnies and fish can be seen simply as springtime symbols of renewed growth and abundance after the long winter months. As chocolate making techniques improved over the years, confectioners began creating elaborate designs and moldings out of chocolate, resulting in cute, eye-catching products like chocolate bunnies and fish that proved irresistible to consumers. Today, these two items continue to be best sellers around Easter.
While none of these theories directly links chocolate rabbits and fish to the crucifixion of Jesus Christ, all three illustrate how different cultural influences, culinary traditions, and consumer trends come together to create unique customs and rituals associated with major holidays like Easter. So, whether you prefer chocolate rabbits or fish (or both!), just remember that behind every delicious treat is a rich history that makes them even more special. 😊
It is not smart as any gpt version. Hoping it'll be in future
Too bad Vicuna isn't actually available yet. (Demos are. Weights are not.)
Weights are now available!
@@engineerprompt only diffs. And need more than 32 gb RAM to run...
Waiting 7b quantized version...
@@enlightenment5d There is already a 4 bit version on huggingface
@@enlightenment5d that will be out in less than a week :)
@@engineerprompt I saw that. Pretty awesome!
Ask it how to spell reasoning in the title slide
I hope they didn’t use that ai to come up with that name.
Haha, may be. All of these are based on animals. Soon they will run out of animal names
Oh wow
Tried it. It's crap.
⚠️ This transition sound every now and then is really annoying. Can you not use it ? Besides, it is 200% louder than your voice. You don't need this sound. Thanks. 🙏🙏👍
Thanks for the feedback, will definitely keep it in mind.
Tried it locally. Its not like the online-demo version. It's crap and to some extend deeply offensive and rude.
it was designed by open ai company so no wonder
There is no way this is even close to chatGPT. ChatGPT was trained on basically the entire internet, which is what? Exobytes of data? The engineering challenges alone are something no one is even close to replicating, let alone modeling and fine tubing. The resting is flawed if a volunteer open source project is getting 90% of what ChatGPT is getting.
You are probably right but I think there is another perspective that is important. The goal doesn't have to be able to get similar performance like chatGPT on all the tasks, rather you can fine tune these models with much smaller datasets for a very specific task. These task specific models will be able to beat chatGPT on those task. Look up BloombergGPT. Second important thing to consider is that this shows the feasibility of fine-tuning your own models with much lower costs. I have seen some very interesting use cases by talking to so folks in the industry.
It's not even close.
Hello
Hi :)
No, in short it's not.
:)
TLDR: Vicuna is nowhere near as good as chatgpt.
Vicuna 70k questions is quite tiny in comparison to chatgpt billions of words and phrases. Specifically a dataset of over 45 terabytes of text data, which includes a wide range of sources such as books, websites, and other online content