I thought the same thing that you realized in this video. However, its clear that we need to be using embeddings, and not fine tunes. Then we can use the D3 model with our data, to get the same response flows we get from GPT3.
would love a semantic search video. Im looking to make a gpt assistant that uses a bunch of video transcripts to provide assistance and recommendations
Really great video. I probably wasted $200-$300 on trying to fine-tune models to sound like a specific person. They all ended up hallucinating and made no sense. It's really annoying that so many people are claiming that it's possible and so few are actually showing examples where it's done in a good way. The only thing that they are good at is labeling or filtering out data from a big text or something. But I don't see why you would fine-tune a model for that. It's easier to just have it in the prompt of the regular ChatGPT or GPT-4. Anyway, love your videos, keep them coming!
Your are the best! So glad I found this channel. Could you please do a video on Sematic Search? I realize that this is what I have been trying to achieve and I though fine-tuning will help achieve that.
Hi Liam! Great vid, as always. I think a video about 'teaching' more data to GPT-3, whether by semantic search or Langchain would be really informative. You mentioned that Langchain was better with numerical data; would that mean that semantic search is better with text?
@@LiamOttley if you just want to fetch text data, use semantic search. however if you want something like the max heartbeat in a table of patients, or if you want the average of male athletes from your database, than use the SQL query method.
Hi Liam, great video! I am a little confused however. Are you saying that we should not attempt something like you ran through in your "How to Fine Tune GPT3 | Beginner's Guide to Building Businesses w/ GPT-3" video (ie fine tuning on data about player statistics then attempting to ask queries about it)? I gather that in this case (asking questions about a players statistics) it would be a better approach to instead train on a number of prompts in English with completions that are some sort of structured query (eg a line of SQL), use that query on our own API serving the data and go back to the user with a hand crafted response with the queried data inserted? Does that mean it's impossible/not recommended to build a model (via fine tuning only) to answer questions that would require an "understanding" of the data? For example "Give me the average number of shots attempted by all players in team X" without explicitly giving examples of that type of query and creating a hand crafted response?
7:08 is it possible to work with the code? for example you have code that has 4000 lines and you want to work with this code. So I can't really paste this into the chat or send it to GPT. Can I use semantic search to send this code to GPT and send API requests to GPT about this code? Well.... I can't really find any good tutorial or explanation of the use cases for this.
If we do semantic search does that take the context from the data and add it to the Api request ? And If so using a vector database will still have the limit of 4000 tokens which in this case will be the Original Query + the context from the search results ?
Really enjoy your videos! We have idea and fund and look for a software engineer with GPT model experiences (prompt engineering and fine tuning etc.) It's still new skills and a lot of software engineers don't have the expertise. Where can we find such talents?
Hey Liam. This video content is crystal clear and interesting. Please do a video on Sematic Search by taking up a use case. It would also be better if you could implement the fine tuning of model for the resume use case that you demonstrated. Looking forward for this 😄🤘
The part I am confused is that when fine tuning it completes my prompts perfectly, but when I try to have a conversation it doesnt remember the context of previous messages. So i got the idea to feed it back in as an input the entire conversation up to that point, however it gets all crazy after 3-4 messages because the training dataset does not contain any prompts that have conversation history in it, so how is this done in order to have conversations with it like chatgpt?? Please someone help me
So its all about fine tuning. How you do it that's important to achieve best results. However getting the data and labelling it requires a good amount of time too. Super informative video. I also misunderstood that using GPT-3 API is like chatgpt responsiveness. 😂
Nice video however I’m still confused by reading in different places conflicting information. I have a couple use cases Im working to see if possible. One of them would be to fine tune GPT with internal product manuals, knowledge base and engineering documentation and be able to ask GPT for an information that is inside these documents and for it to respond. Some places says it’s possible and in your video sounds like it’s not possible. Another use case I have is to fine tune with multiple questions and answers we have in some forms and the end result of a multiple line quote of services. The idea is to see if GPT can help do these estimations and how far off it will be from a person.
You're referencing 'Text-Embeddings' I believe. You'd need to create text embeddings of your internal docs -> Store those in a Vector DB like Pinecone or pgVector -> Then generate text embeddings off your prompt. You use semantic search and cosine similarity
@@joshuashepherd7189 Is openai's ada 2 the best for generating embedding or can you use other things, I think I heard you can use fasttext or a tensorflow universal sentence encoder
@@nattyzaddy6555 Good question! Honestly, It just depends on the project. If you're doing something BIG and need to store like production data for a service, Pinecone is probably better since its geared toward that kind of use. However, If you're working a proof of concept project, or something you want to use just for yourself or friends, pgVector is going to be your best, free option. But you may need to do more manual config, than you would with Pinecone
@@nattyzaddy6555 You can definitely use other embedding tools for sure. You just need to make sure the embed dimensions are compatible with whatever vector database you're using. Like 1536 dimensions is the current Ada config. More dimensions = more quality, but it doesn't always directly mean that
6:01 I totally don't understand what are you talking about here. I see that you are trying your best to share your knowledge, but for me it's too fast and too chaotic.
A i automation agency is to be offered a real money from this business so that l had attracted in this business and l will be expecting a lot of money from this business
Leave your questions below! 😎
📚 My Free Skool Community: bit.ly/3uRIRB3
🤝 Work With Me: www.morningside.ai/
📈 My AI Agency Accelerator: bit.ly/3wxLubP
I thought the same thing that you realized in this video. However, its clear that we need to be using embeddings, and not fine tunes. Then we can use the D3 model with our data, to get the same response flows we get from GPT3.
would love a semantic search video. Im looking to make a gpt assistant that uses a bunch of video transcripts to provide assistance and recommendations
Really great video. I probably wasted $200-$300 on trying to fine-tune models to sound like a specific person. They all ended up hallucinating and made no sense. It's really annoying that so many people are claiming that it's possible and so few are actually showing examples where it's done in a good way. The only thing that they are good at is labeling or filtering out data from a big text or something. But I don't see why you would fine-tune a model for that. It's easier to just have it in the prompt of the regular ChatGPT or GPT-4. Anyway, love your videos, keep them coming!
Your are the best! So glad I found this channel. Could you please do a video on Sematic Search? I realize that this is what I have been trying to achieve and I though fine-tuning will help achieve that.
A very good explanation. There is so much misunderstanding about what is contained in a model.
Hi Liam! Great vid, as always. I think a video about 'teaching' more data to GPT-3, whether by semantic search or Langchain would be really informative. You mentioned that Langchain was better with numerical data; would that mean that semantic search is better with text?
Semantic search works on searching for the meaning of text, rather than the words themselves. By nature it works on text rather than numbers :)
@@LiamOttley if you just want to fetch text data, use semantic search.
however if you want something like the max heartbeat in a table of patients, or if you want the average of male athletes from your database, than use the SQL query method.
Awesome vdo … definitely would want to see vdo on semantic search.
Would be great if you can do a video on semantic search too!
Sure!
Hi Liam, great video!
I am a little confused however. Are you saying that we should not attempt something like you ran through in your "How to Fine Tune GPT3 | Beginner's Guide to Building Businesses w/ GPT-3" video (ie fine tuning on data about player statistics then attempting to ask queries about it)?
I gather that in this case (asking questions about a players statistics) it would be a better approach to instead train on a number of prompts in English with completions that are some sort of structured query (eg a line of SQL), use that query on our own API serving the data and go back to the user with a hand crafted response with the queried data inserted?
Does that mean it's impossible/not recommended to build a model (via fine tuning only) to answer questions that would require an "understanding" of the data? For example "Give me the average number of shots attempted by all players in team X" without explicitly giving examples of that type of query and creating a hand crafted response?
would love to see a video on semantic search please Liam
Hi Liam! Awesome video! thank you! please do video on Semantic Search, please!
7:08 is it possible to work with the code? for example you have code that has 4000 lines and you want to work with this code. So I can't really paste this into the chat or send it to GPT. Can I use semantic search to send this code to GPT and send API requests to GPT about this code? Well.... I can't really find any good tutorial or explanation of the use cases for this.
It’s a great video as usual. Would like to learn more about semantic search
If we do semantic search does that take the context from the data and add it to the Api request ?
And If so using a vector database will still have the limit of 4000 tokens which in this case will be the Original Query + the context from the search results ?
Another awesome video bro.
Really enjoy your videos! We have idea and fund and look for a software engineer with GPT model experiences (prompt engineering and fine tuning etc.) It's still new skills and a lot of software engineers don't have the expertise. Where can we find such talents?
Hey Liam. This video content is crystal clear and interesting. Please do a video on Sematic Search by taking up a use case. It would also be better if you could implement the fine tuning of model for the resume use case that you demonstrated. Looking forward for this 😄🤘
do you put a filter on your mic audio? it sounds bad
Thank you captain. I've been searching all over for an explainer like this. Thanks man.
This is really really helpful sir. Thanks for the amazing video
The part I am confused is that when fine tuning it completes my prompts perfectly, but when I try to have a conversation it doesnt remember the context of previous messages. So i got the idea to feed it back in as an input the entire conversation up to that point, however it gets all crazy after 3-4 messages because the training dataset does not contain any prompts that have conversation history in it, so how is this done in order to have conversations with it like chatgpt?? Please someone help me
Awesome video as always mate!
So its all about fine tuning. How you do it that's important to achieve best results. However getting the data and labelling it requires a good amount of time too.
Super informative video. I also misunderstood that using GPT-3 API is like chatgpt responsiveness. 😂
Thanks Liam!! Great video! :)
Nice video however I’m still confused by reading in different places conflicting information. I have a couple use cases Im working to see if possible. One of them would be to fine tune GPT with internal product manuals, knowledge base and engineering documentation and be able to ask GPT for an information that is inside these documents and for it to respond. Some places says it’s possible and in your video sounds like it’s not possible. Another use case I have is to fine tune with multiple questions and answers we have in some forms and the end result of a multiple line quote of services. The idea is to see if GPT can help do these estimations and how far off it will be from a person.
You're referencing 'Text-Embeddings' I believe. You'd need to create text embeddings of your internal docs -> Store those in a Vector DB like Pinecone or pgVector -> Then generate text embeddings off your prompt. You use semantic search and cosine similarity
@@joshuashepherd7189 Is openai's ada 2 the best for generating embedding or can you use other things, I think I heard you can use fasttext or a tensorflow universal sentence encoder
@@joshuashepherd7189 Also do you prefer pgvector or pinecone? Thanks dude
@@nattyzaddy6555 Good question! Honestly, It just depends on the project. If you're doing something BIG and need to store like production data for a service, Pinecone is probably better since its geared toward that kind of use. However, If you're working a proof of concept project, or something you want to use just for yourself or friends, pgVector is going to be your best, free option. But you may need to do more manual config, than you would with Pinecone
@@nattyzaddy6555 You can definitely use other embedding tools for sure. You just need to make sure the embed dimensions are compatible with whatever vector database you're using. Like 1536 dimensions is the current Ada config. More dimensions = more quality, but it doesn't always directly mean that
It would be good if you created a presentation on the value you and your team can add to a AI enabled project.
$500 for a 45 minute call, that’s twice as much as a lawyer….
6:01 I totally don't understand what are you talking about here. I see that you are trying your best to share your knowledge, but for me it's too fast and too chaotic.
Tunnig BERT pls also
here before 100k !!!!
Virtual assistant jobs gonna go on the rise?
When l would be expected your first income from this business
A i automation agency is to be offered a real money from this business so that l had attracted in this business and l will be expecting a lot of money from this business
Bro love your videos but never get that haircut again